7373tpPath.indd 1
5/19/10 3:29:55 PM
May 13, 2010
10:6
WSPC/SPIB913
b913fm
This page intentionally left blank
FA
World Scientific
7373tpPath.indd 2
5/19/10 3:29:56 PM
Published by World Scientific Publishing Co. Pte. Ltd. 5 Toh Tuck Link, Singapore 596224 USA office: 27 Warren Street, Suite 401402, Hackensack, NJ 07601 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Library of Congress CataloginginPublication Data Alternative investments and strategies / edited by Rüdiger Kiesel, Matthias Scherer & Rudi Zagst. p. cm. ISBN13: 9789814280105 ISBN10: 9814280100 1. InvestmentsMoral and ethical aspects. 2. Portfolio managementMoral and ethical aspects. I. Kiesel, Rüdiger, 1962– II. Scherer, Matthias. III. Zagst, Rudi, 1961– HG4515.13.A498 2010 332.6dc22 2010013167
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
Copyright © 2010 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
Typeset by Stallion Press Email:
[email protected] Printed in Singapore.
Juliet  Alternative Investments.pmd
1
8/2/2010, 6:18 PM
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
PREFACE
Asset allocation investigates the optimal division of a portfolio among different asset classes. Standard theory involves the optimal mix of risky stocks, bonds, and cash together with various subdivisions of these asset classes. Underlying this is the insight that diversification allows for achieving a balance between risk and return: by using different types of investment, losses may be limited and returns are made less volatile without losing too much potential gain. These insights are made precise using the benchmark theory of mathematical finance, the BlackScholesMerton theory, based on Brownian motion as the driving noise process for risky asset prices. Here, the distributions of financial returns of the risky assets in a portfolio are multivariate normal, thus relating to the standard meanvariance portfolio theory of Markowitz with its riskreturn paradigm as above. Recent years have seen many empirical studies shedding doubt on the BlackScholesMerton model, and motivating various alternative modeling approaches, which were able to reproduce the stylized facts of asset returns (such as heavy tails and volatility clustering) much better. Also, various new asset classes and specific financial tools for achieving better diversification have been created and entered the investment universe. This book combines academic research and practical expertise on these new (often called alternative) assets and trading strategies in a unique way. We include the practitioners’ viewpoint on new asset classes as well as academic research on modeling approaches, for new asset classes. In particular, alternative asset classes such as power forward contracts, forward freight agreements, and investment in photovoltaic facilities are discussed in detail, both on a standalone basis and with a view to their effects on diversification in combination with classical asset. We also analyse creditrelated portfolio instruments and their effect in achieving an optimal asset allocation. In this context, we highlight aspects of financial structures which may sometimes be neglected, such as default risk of issuer in case of certificates or the role that model v
May 13, 2010
vi
10:6
WSPC/SPIB913
b913fm
FA
Preface
risk plays within asset allocation problems. This leads naturally to the use of robust asset allocation strategies. Extending the classical meanvariance portfolio setting, we include dynamic portfolio strategies and illustrate different portfolio protection strategies. In particular, we compare the benefits of such strategies and investigate conditions under which Constant Proportion Portfolio Insurance (CPPI) may be prefered to OptionBased Portfolio Insurance (OBPI) and vice versa. We also contribute to the understanding of gap risk by analyzing this risk for CPPI and Constant Proportion Debt Obligations (CPDO) in a sophisticated modeling framework. Such analyses are supplemented and extended by an investigation of the optimality of hedging approaches such as varianceoptimal hedging and semistatic variants of classical hedging strategies. Many of the articles can serve as guides for the implementation of various models. In addition, we also present stateoftheart models and explain modern tools from financial mathematics, such as MarkovSwitching models, timechanged Lévy models, variants of lognormal approximations, and copula structures. This books combines a unique mix of authors. Also many of our students improved the outcome of the project with critical and insightful comments. Particular thanks goes to Georg Grüll, Peter Hieber, Julia Kraus, Matthias Lutz, JanFrederik Mai, Kathrin Maul, Kevin Metka, Daniela Neykova, Johannes Rauch, Andreas Rupp, Daniela Selch, and Christofer Vogt. R. Kiesel, M. Scherer, and R. Zagst
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
CONTENTS
Preface
v
Part I. Alternative Investments Chapter 1. Socially Responsible Investments
3
Sven Hroß, Christofer Vogt and Rudi Zagst 1.1 1.2 1.3
1.4
1.5
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . Recent Research on SRI . . . . . . . . . . . . . . . . . . . How Sustainable is Sustainability? . . . . . . . . . . . . . . 1.3.1 Description of the Dataset . . . . . . . . . . . . . . 1.3.2 Introduction to Markov Transition Matrices . . . . . 1.3.3 Results of Markov Transition Matrices . . . . . . . SRI in Portfolio Context . . . . . . . . . . . . . . . . . . . 1.4.1 Description of the Dataset and Statistical Properties 1.4.2 MarkovSwitching Model . . . . . . . . . . . . . . 1.4.3 Fitting the Model Parameters . . . . . . . . . . . . 1.4.4 Simulation of Returns . . . . . . . . . . . . . . . . 1.4.5 Portfolio Optimization Models . . . . . . . . . . . 1.4.6 Definition of Investor Types . . . . . . . . . . . . . 1.4.7 Optimal Portfolios . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
Chapter 2. Listed Private Equity in a Portfolio Context
4 5 6 6 6 7 8 8 11 11 13 13 15 15 18 21
Philipp Aigner, Georg Beyschlag, Tim Friederich, Markus Kalepky and Rudi Zagst 2.1 2.2
Introduction . . . . . . . . . . . . . Defining Private Equity Categories . 2.2.1 Financing Stages . . . . . . 2.2.2 Divestment Strategies . . . vii
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
22 23 23 24
May 13, 2010
10:6
WSPC/SPIB913
b913fm
viii
Contents
2.2.3 2.2.4
2.3
2.4
2.5
FA
Type of Financing . . . . . . . . . . . . . . . . . . . . . . Classification of Private Equity Fund Investments . . . . . 2.2.4.1 Venture capital funds . . . . . . . . . . . . . . . . 2.2.4.2 Buyout funds . . . . . . . . . . . . . . . . . . . . 2.2.4.3 Leveraged buyouts (LBO) . . . . . . . . . . . . . Investment Possibilities — One Asset, Many Classes . . . . . . . . 2.3.1 Direct Investments . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Private Equity Funds . . . . . . . . . . . . . . . . . . . . . 2.3.2.1 Key players . . . . . . . . . . . . . . . . . . . . 2.3.3 Cash Flow Structure of a Private Equity Fund . . . . . . . . 2.3.4 FundofFunds . . . . . . . . . . . . . . . . . . . . . . . . 2.3.4.1 Structure of a private equity fundoffunds . . . . 2.3.4.2 Advantages . . . . . . . . . . . . . . . . . . . . . 2.3.4.3 Disadvantages . . . . . . . . . . . . . . . . . . . 2.3.5 Publicly Traded Private Equity . . . . . . . . . . . . . . . 2.3.6 Secondary Transactions . . . . . . . . . . . . . . . . . . . 2.3.6.1 Types of secondary transactions . . . . . . . . . . 2.3.6.2 Buyer’s motivation . . . . . . . . . . . . . . . . . Private Equity as Alternative Asset Class in an Investment Portfolio . . . . . . . . . . . . . . . . . . . . . . 2.4.1 Characteristics of LPE Return Series . . . . . . . . . . . . 2.4.2 Modeling Return Series with MarkovSwitching Processes . 2.4.2.1 Markov–Switching models . . . . . . . . . . . . 2.4.2.2 Fitting the parameters . . . . . . . . . . . . . . . 2.4.2.3 Simulation of return paths . . . . . . . . . . . . . 2.4.3 Listed Private Equity in Asset Allocation . . . . . . . . . . 2.4.3.1 Performance measurement . . . . . . . . . . . . . 2.4.3.2 Portfolio optimization frameworks . . . . . . . . 2.4.3.3 Definition of investor types . . . . . . . . . . . . 2.4.3.4 Optimization of portfolios . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
25 26 26 27 27 28 28 29 29 31 32 32 32 33 33 34 34 35
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
35 36 37 37 39 40 40 40 42 43 44 47
Chapter 3. Alternative Real Assets in a Portfolio Context
51
Wolfgang Mader, Sven Treu and Sebastian Willutzky 3.1 3.2 3.3
Introduction . . . . . . . . . . . . . . . . . Overview on Alternative Real Assets . . . . Modeling Photovoltaic Investments . . . . 3.3.1 General Approach . . . . . . . . . 3.3.2 Definition of the Investment Project
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
52 52 53 53 54
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
ix
Contents
3.3.3
3.4
3.5
Modeling of Risk Factors . . . . . . . . . . . . . . . . . 3.3.3.1 Economic factors . . . . . . . . . . . . . . . . 3.3.3.2 Noneconomic factors . . . . . . . . . . . . . . 3.3.3.3 Historical analysis of monthly global irradiance 3.3.3.4 Monte Carlo analysis of yearly global irradiance Photovoltaic Investments in a Portfolio Context . . . . . . . . . . 3.4.1 Setting the Portfolio Context . . . . . . . . . . . . . . . . 3.4.2 Including Photovoltaic Investments in a Portfolio . . . . . 3.4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
Chapter 4. The Freight Market and Its Derivatives
56 56 57 58 61 63 63 64 66 68 71
Rüdiger Kiesel and Patrick Scherer 4.1
4.2
4.3
4.4
4.5 4.6 4.7
Introduction: the Freight Market . . . . . . . 4.1.1 Vessels . . . . . . . . . . . . . . . . 4.1.2 Cargo . . . . . . . . . . . . . . . . . 4.1.3 Routes . . . . . . . . . . . . . . . . Freight Rates: What Drives the Market? . . . 4.2.1 Demand for Shipping Capacity . . . 4.2.2 Supply of Shipping Capacity . . . . 4.2.3 Costs . . . . . . . . . . . . . . . . . Freight Derivatives: Hedging or Speculating? 4.3.1 Forward Freight Agreement . . . . . 4.3.2 Freight Futures . . . . . . . . . . . . Explanatory Variables . . . . . . . . . . . . 4.4.1 Explanatory Power . . . . . . . . . . 4.4.2 Granger Causality . . . . . . . . . . 4.4.3 Selection Algorithm “Top Five” . . . 4.4.4 Cointegration . . . . . . . . . . . . . Predicting Freight Spot and Futures Rates . . The Backtesting Algorithm . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
Chapter 5. On Forward Price Modeling in Power Markets
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
72 72 72 73 74 75 76 77 77 77 78 79 80 82 83 84 86 88 90 93
Fred Espen Benth 5.1 5.2 5.3
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 HJM Approach to Power Forward Pricing . . . . . . . . . . . . . . . . . 95 Power Forwards and Approximation by Geometric Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
x
Contents
5.3.1
5.4 5.5
A Geometric Brownian Motion Dynamics by Volatility Averaging . . . . . . . . . . . . . . . . . 5.3.2 A Geometric Brownian Motion Dynamics by Moment Matching . . . . . . . . . . . . . . . . . 5.3.3 The Covariance Structure Between Power Forwards . 5.3.4 The Distribution of a Power Forward . . . . . . . . . 5.3.5 Numerical Analysis of the Power Forward Distribution Pricing of Options on Power Forwards . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . 101 . . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
Chapter 6. Pricing Certificates Under Issuer Risk
103 106 108 110 114 119 123
Barbara Götz, Rudi Zagst and Marcos Escobar 6.1 6.2 6.3
6.4
Introduction . . . . . . . . . . . . . . . . . The Model . . . . . . . . . . . . . . . . . Pricing of Certificates Under Issuer Risk . . 6.3.1 Building Blocks . . . . . . . . . . 6.3.2 Index Certificates . . . . . . . . . 6.3.3 Participation Guarantee Certificates 6.3.4 Bonus Guarantee Certificates . . . 6.3.5 Discount Certificates . . . . . . . . 6.3.6 Bonus Certificates . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
Chapter 7. Asset Allocation with Credit Instruments
124 125 126 126 130 132 134 135 136 139 147
Barbara Menzinger, Anna Schlösser and Rudi Zagst 7.1 7.2 7.3 7.4
7.5
7.6
Introduction . . . . . . . . . . . . . . . . . . . . . Simulation Framework . . . . . . . . . . . . . . . Framework for Total Return Calculation . . . . . . Optimization Framework . . . . . . . . . . . . . . 7.4.1 MeanVariance Optimization . . . . . . . 7.4.2 CVaR Optimization . . . . . . . . . . . . Model Calibration and Simulation Results . . . . . 7.5.1 MeanVariance Approach . . . . . . . . . 7.5.2 Conditional Value at Risk . . . . . . . . . 7.5.3 Comparison of Selected Optimal Portfolios Summary and Conclusion . . . . . . . . . . . . .
Chapter 8. Cross Asset Portfolio Derivatives Stephan Höcht, Matthias Scherer and Philip Seegerer
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
148 150 153 156 156 157 157 162 164 167 170 175
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
xi
Contents
8.1 8.2 8.3
8.4
8.5 8.6
Introduction to Cross Asset Portfolio Derivatives 8.1.1 Definitions and Examples . . . . . . . . Collateralized Obligations . . . . . . . . . . . . A Comparison of CFO with CTSO . . . . . . . . 8.3.1 Structural Features of CFO . . . . . . . 8.3.2 Structural Features of CTSO . . . . . . . 8.3.3 The Different Risks . . . . . . . . . . . 8.3.4 Correlation of Tail Events in CTSO . . . Pricing Cross Asset Portfolio Derivatives . . . . 8.4.1 Pricing Trigger Swaps . . . . . . . . . . 8.4.2 Pricing nthtoTrigger Baskets . . . . . . 8.4.3 Pricing CTSO . . . . . . . . . . . . . . 8.4.4 Modeling Approaches . . . . . . . . . . 8.4.4.1 The structural approach . . . . 8.4.4.2 The copula approach . . . . . . 8.4.5 An Example for an nthto Trigger Basket 8.4.5.1 A pricing exercise of Example 3 (structural approach) . . . . . 8.4.5.2 A pricing exercise of Example 3 (copula approach) . . . . . . . 8.4.5.3 Resulting model spreads . . . . Outlook . . . . . . . . . . . . . . . . . . . . . . Conclusion . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
175 176 179 179 179 181 181 181 182 182 183 184 185 185 186 188
. . . . . . . . . . . . . 188 . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
189 190 194 195
Part II. Alternative Strategies Chapter 9. Dynamic Portfolio Insurance Without Options
201
Dominik Dersch 9.1 9.2
9.3 9.4
Introduction . . . . . . . . . . . Simple Strategies . . . . . . . . 9.2.1 BuyandHold . . . . . 9.2.2 StopLoss . . . . . . . 9.2.3 The Bond Floor Strategy 9.2.4 Plain Vanilla CPPI . . . Historical Simulation I . . . . . Advanced Features . . . . . . . 9.4.1 Transaction Costs . . . 9.4.2 Transaction Filter . . . 9.4.3 Lockin Levels . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
202 203 203 203 204 205 206 209 210 210 211
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
xii
9.5
9.6 9.7
Contents
9.4.4 Leverage and Constrain of Exposure . . . . . 9.4.5 Rebalancing Strategies for the Risky Portfolio 9.4.6 CPPI and Beyond . . . . . . . . . . . . . . . Historical Simulation II . . . . . . . . . . . . . . . . 9.5.1 Transaction Costs and Transaction Filter . . . 9.5.2 Lockin Levels . . . . . . . . . . . . . . . . . 9.5.3 The Use of Leverage . . . . . . . . . . . . . . 9.5.4 CPPI on a MultiAsset Risky Portfolio . . . . Implement a Dynamic Protection Strategy with ETF . Closing Remarks . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
Chapter 10. How Good are Portfolio Insurance Strategies?
212 213 213 214 214 216 220 222 223 224 227
Sven Balder and Antje Mahayni 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Optimal Portfolio Selection with Finite Horizons . . . . . . . . 10.2.1 Problem (A) . . . . . . . . . . . . . . . . . . . . . . . 10.2.2 Problem (B) . . . . . . . . . . . . . . . . . . . . . . . 10.2.3 Problem (C) . . . . . . . . . . . . . . . . . . . . . . . 10.2.4 Comparison of Optimal Solutions . . . . . . . . . . . . 10.3 Utility Loss Caused by Guarantees . . . . . . . . . . . . . . . . 10.3.1 Justification of Guarantees and Empirical Observations 10.3.2 Utility Loss . . . . . . . . . . . . . . . . . . . . . . . . 10.4 Utility Loss Caused by Trading Restrictions and Transaction Costs . . . . . . . . . . . . . . . . . . . . . . 10.4.1 DiscreteTime CPPI . . . . . . . . . . . . . . . . . . . 10.4.2 DiscreteTime OptionBased Strategy . . . . . . . . . . 10.4.3 Comments on Utility Loss and Shortfall Probability . . 10.5 Utility Loss Caused by Guarantees and Borrowing Constraints . . . . . . . . . . . . . . . . . . . . 10.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
228 230 233 234 235 238 242 242 242
. . . .
. . . .
. . . .
. . . .
. . . .
246 246 249 250
. . . . . 252 . . . . . 254
Chapter 11. Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
259
Elisabeth Joossens and Wim Schoutens 11.1 Introduction . . . . . . . . . . . . . . 11.2 Credit Risk and Credit Default Swaps 11.2.1 Credit Risk . . . . . . . . . . 11.2.2 Credit Default Swaps (CDS) . 11.3 Portfolio Insurances . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
260 261 261 265 267
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
xiii
Contents
11.4 Modeling of CPPI Dynamics Using Multivariate JumpDriven Processes . . . . . . . . . . . . . . . . . . . . . . . . 11.4.1 Multivariate Variance Gamma Modeling . . . . . . . . . . 11.4.2 Swaptions on Credit Indices . . . . . . . . . . . . . . . . . 11.4.2.1 Black’s model . . . . . . . . . . . . . . . . . . . 11.4.2.2 The variance gamma model . . . . . . . . . . . . 11.4.3 Spread Modeling by Correlated VG Processes . . . . . . . 11.4.3.1 The pricing of CPPIs . . . . . . . . . . . . . . . . 11.4.3.2 Gap risk . . . . . . . . . . . . . . . . . . . . . . 11.5 Recent Developments for CPPI . . . . . . . . . . . . . . . . . . . 11.5.1 Portfolio Insurance: The Extreme Value Approach to the CPPI Method . . . . . . . . . . . . . . . . . . . . . 11.5.2 VaR Approach for Credit CPPI . . . . . . . . . . . . . . . 11.5.3 CPPI with Cushion Insurance . . . . . . . . . . . . . . . . 11.6 A New Financial Instrument: Constant Proportion Debt Obligations 11.6.1 The Structure . . . . . . . . . . . . . . . . . . . . . . . . . 11.6.2 CPDOs in the Spotlight . . . . . . . . . . . . . . . . . . . 11.6.3 Rating CPDOs Under VG Dynamics . . . . . . . . . . . . 11.7 Comparison Between CPPI and CPDO . . . . . . . . . . . . . . . 11.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
270 270 273 273 274 275 275 279 281
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
282 283 284 285 285 289 289 291 292
Chapter 12. On the Benefits of Robust Asset Allocation for CPPI Strategies
295
Katrin Schöttle and Ralf Werner 12.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2 The Financial Market . . . . . . . . . . . . . . . . . . . . . . . . . 12.2.1 The Basic Financial Market . . . . . . . . . . . . . . . . . 12.2.2 The Riskless Asset . . . . . . . . . . . . . . . . . . . . . . 12.2.3 The Risky Asset . . . . . . . . . . . . . . . . . . . . . . . 12.2.4 Classical Mean–Variance Analysis . . . . . . . . . . . . . 12.2.5 The Trading Strategy . . . . . . . . . . . . . . . . . . . . . 12.3 The Standard CPPI Strategy . . . . . . . . . . . . . . . . . . . . . 12.3.1 The Simple Case . . . . . . . . . . . . . . . . . . . . . . . 12.3.2 The General Case . . . . . . . . . . . . . . . . . . . . . . 12.3.3 Shortfall Probability of CPPI Strategies . . . . . . . . . . . 12.3.4 Improving CPPI Strategies . . . . . . . . . . . . . . . . . . 12.3.5 CPPI Strategies Under Estimation Risk . . . . . . . . . . . 12.4 Robust Mean–Variance Optimization and Improved CPPI Strategies 12.4.1 Robust Mean–Variance Analysis . . . . . . . . . . . . . . . 12.4.2 Uncertainty Sets Via Expert Opinions or Related Estimators
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
296 296 297 298 298 300 302 302 303 305 308 310 313 316 317 317
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
xiv
Contents
12.4.3 Uncertainty Sets Via Confidence Sets . . . 12.4.4 Usage and Implications for CPPI Strategies 12.4.5 CPPIs with Robust Asset Allocations . . . 12.5 Conclusion . . . . . . . . . . . . . . . . . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
Chapter 13. Robust Asset Allocation Under Model Risk
319 321 323 324 327
Pauline Barrieu and Sandrine Tobelem 13.1 Background . . . . . . . . . . . . . . . . . . . . . . . 13.2 A Robust Approach to Model Risk . . . . . . . . . . . 13.2.1 The Absolute Ambiguity Robust Adjustment . 13.2.2 Relative Ambiguity Robust Adjustment . . . . 13.2.3 ARA Parametrization . . . . . . . . . . . . . 13.3 Some Definitions Relative to the AmbiguityAdjusted Asset Allocation . . . . . . . . . . . . . . . . . . . . 13.4 Empirical Tests . . . . . . . . . . . . . . . . . . . . . 13.4.1 Portfolios Tested . . . . . . . . . . . . . . . . 13.4.2 Performance Measures . . . . . . . . . . . . . 13.4.3 Results . . . . . . . . . . . . . . . . . . . . . 13.4.3.1 Performances of the different models 13.4.3.2 SEU portfolio . . . . . . . . . . . . 13.4.3.3 Ambiguity robust portfolios . . . . . 13.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
328 329 330 333 334
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
335 336 337 339 340 341 342 342 343
Chapter 14. SemiStatic Hedging Strategies for Exotic Options
345
Hansjörg Albrecher and Philipp Mayer 14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.2 Hedging PathIndependent Options . . . . . . . . . . . . . . 14.2.1 Plain Vanilla Options with Arbitrary Strikes are Liquid 14.2.2 Finitely Many Liquid Strikes . . . . . . . . . . . . . 14.3 Hedging Barrier and Other Weakly Path Dependent Options . 14.3.1 ModelDependent Strategies: Perfect Replication . . . 14.3.2 ModelDependent Strategies: Approximations . . . . 14.3.3 ModelIndependent Strategies: Robust Strategies . . . 14.4 Hedging Strongly PathDependent Options . . . . . . . . . . 14.4.1 Lookback Options . . . . . . . . . . . . . . . . . . . 14.4.2 Asian Options . . . . . . . . . . . . . . . . . . . . . 14.5 Case Study: ModelDependent Hedging of Discretely Sampled Options . . . . . . . . . . . . . . . . . . . . . . . . 14.6 Conclusion and Future Research . . . . . . . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
346 347 348 349 350 351 357 359 361 362 364
. . . . . . 367 . . . . . . 370
May 13, 2010
10:6
WSPC/SPIB913
b913fm
FA
xv
Contents
Chapter 15. DiscreteTime VarianceOptimal Hedging in Affine Stochastic Volatility Models
375
Jan Kallsen, Richard Vierthauer, Johannes MuhleKarbe and Natalia Shenkman 15.1 15.2 15.3 15.4
Introduction . . . . . . . . . . . . . . . . DiscreteTime VarianceOptimal Hedging The Laplace Transform Approach . . . . Application to Affine Stochastic Volatility Models . . . . . . . . . . . . . 15.5 Numerical Illustration . . . . . . . . . . Index
. . . . . . . . . . . . . . . . . 376 . . . . . . . . . . . . . . . . . 377 . . . . . . . . . . . . . . . . . 378 . . . . . . . . . . . . . . . . . 380 . . . . . . . . . . . . . . . . . 388 395
This page intentionally left blank
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
FA
Part I
Alternative Investments
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
This page intentionally left blank
FA
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
FA
1
SOCIALLY RESPONSIBLE INVESTMENTS
SVEN HROß∗ , CHRISTOFER VOGT† and RUDI ZAGST‡ HVBStiftungsinstitut für Finanzmathematik, Technische Universität München, Boltzmannstr. 3, 85747 München, Germany ∗
[email protected] †
[email protected] ‡
[email protected] Within the last two decades, the market of socially responsible investing (SRI) has seen unprecedented growth and has become more and more important, not only because of the current financial crisis. This chapter gives a survey of the asset class SRI in general, i.e., market development and investment possibilities. Moreover, the question “How sustainable is sustainability?” is addressed by analyzing SAM Group sustainability rankings of the years 2001–2007. Furthermore, the ability of SRI to contribute to diversification within a portfolio is scrutinized. The analysis is based on simulated returns generated by an autoregressive MarkovSwitching model and accounts for different levels of investors’ risk aversion. Optimal portfolios consisting of stocks, bonds, and the respective SRI index show that risk–averse investors mix SRI to an established portfolio consisting of bonds and stocks to reduce the risk and increase the performance. Additionally, the asset class SRI is found to be a substitute for the asset class stocks.
3
May 12, 2010
17:46
WSPC/SPIB913
4
b913ch01
FA
Hroß et al.
1.1. INTRODUCTION There are different ways to describe socially responsible investing (SRI). Reference 1 defines SRI as the integration of environmental, social, and corporate governance (ESG) considerations into investment management processes and ownership practices hoping that these factors can have an impact on financial performance. Responsible investment can be practiced across all asset classes. Several reasons can be stated, why the field of SRI has gained great public interest as well as rising economic importance in recent years. Simultaneously to the ongoing climate change debate, public scrutiny and political attention have put pressure on businesses to consider both social and environmental issues in their activities. Accompanied by these developments, the SRI market grew strongly during the last decade. SRI does no longer represent a negligible economical niche, but as stated in [2] it might play a crucial financial role in the future. The current size of the worldwide SRI market is according to [3] approximately 5 trillion. With 53% market share, the greatest part of the SRI market is based in Europe followed by the United States with 39%. The rest of the world represents only 8% of the SRI market. According to [4], the size of the SRI market in the United States was $639 billion in 1995 and then grew up to $2159 billion in 1999, which means an average annual growth rate of 36%. From 1999 to 2005, SRI investment volumes only slightly grew up to $2290 billion, but then growth accelerated again resulting in $2711 billion in 2007. The European SRI market experienced an average growth rate of 51% since 2002 from an absolute investment volume of 336 billion in 2002 up to 2665 billion in 2007. Reference 3 estimates that the share of SRI in the total European fund market is about 17.6% in 2008 and largely driven by institutional investors. There are several possibilities to invest into SRI. For example, the SAM Group (www.samgroup.com) offers a wide range of funds covering the total SRI market and also special funds, e.g., on Islamic sustainability. There are also sustainably managed fixedincome funds available. Another possibility is the direct investment into nonlisted companies or projects. In this context, projects like wind farms or solar parks can be mentioned as suitable investment possibilities. Moreover, certificates are available on the market which allow the investor to participate in the SRI market, e.g., index certificates on the European Renewable Energy Index (ERIX Index Certificate, Societe Generale, ISIN: DE000SG1ERX7). The structure of this chapter is as follows. Section 1.2 gives an overview on recent research on SRI. Section 1.3 answers the question “How sustainable is sustainability?” by using Markov transition matrices. Section 1.4 then analyzes SRI in a portfolio context by generating optimal portfolios for different investors using a MarkovSwitching model and different optimization frameworks. Finally, Sec. 1.5 concludes.
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
Socially Responsible Investments
FA
5
1.2. RECENT RESEARCH ON SRI During the last years, several empirical studies analyzed whether SRI produces or destroys shareholder wealth. Many early studies on the performance of SRI use regression models with one or two factors and try to measure Jensen’s alpha. Reference 5 compares 32 SRI funds to 320 nonSRI funds in the United States between 1981 and 1990 and finds no significant average alphas with respect to a valueweighted NYSE index. More advanced studies apply a matching approach to compare SRI and nonSRI funds with similar characteristics, e.g., fund universe and size. Within this approach, management and transaction costs can be included into the analysis, see, e.g., [6] or [7]. As a result, no significant performance differences between SRI and nonSRI could be observed. One problem is that important characteristics might not be taken into consideration. Reference 8 applies a four factor model according to [9] using as regression factors the excess market return, SMB (“SmallminusBig”: The difference between the return of a small and of a largecap portfolio), HML (“HighminusLow”: The return difference between a value and a growthportfolio, i.e., a portfolio containing firms that dispose of a high booktomarket ratio versus firms with a low value relating to this ratio), and MOM (“Momentum”: The return difference betweeen two portfolios, one consisting of last year’s best performers and the other of the worst performers) in order to analyze the performance of United States, German, and British SRI funds. The authors build two portfolios for each country, one containing all SRI funds, the other the conventional funds, and find under — as well as outperformance of SRI, but none of the differences are significant. Furthermore, SRI funds seem to have an investment bias toward growth stocks (low booktomarket value) and small caps (lower marketcapitalization). Reference 10 uses ecoefficiency rankings of Innovest to evaluate two equity portfolios that differ in ecoefficiency. The highranked portfolio shows significantly higher returns than its lowranked counterpart over the period 1995–2003. In contrast, [11] finds that SRI investors have to pay for their constrained investment style. Another approach is to look at SRI equity indices to avoid usual problems of mutual funds during a performance analysis, e.g., transaction costs of funds or effects of management skills. Reference 12 analyzes 29 SRI indices and applies different settings to test for differences in riskadjusted performance compared to a suitable benchmark. The study concludes that SRI screens do not lead to significant performance difference of SRI indices. Yet, no final answer to the question whether SRI produces or destroys shareholder wealth can be given. Independent of these findings, SRI market growth might simply come from the nonfinancial utility gained by SRI investors. To the authors’ best knowledge, there is yet no such study scrutinizing this effect. Therefore, the focus of Sec. 1.4 lies on the benefits of SRI in a portfolio context. For this, optimal portfolios of bonds, stocks, and SRI will be constructed for different investor types and in different optimization frameworks.
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
FA
6
Hroß et al.
1.3. HOW SUSTAINABLE IS SUSTAINABILITY? In this section, the endurance of sustainability is analyzed. This is especially important for an SRI investor, who does not want to have too many reallocations in his portfolio. Moreover, sustainability scores should be enduring by the pure definition of the word “sustainability”. For this aim, sustainability scores from SAM Group, one of the world’s most respected companies in the field of SRI assessment, are scrutinized. This study is implemented using Markov transition matrices.
1.3.1. Description of the Dataset The dataset used for the analysis contains the sustainability scores (hereinafter called total score) of 822 companies. The methodology for calculating the total score of a firm is given as follows. A company’s economic, ecologic, and social performance is analyzed, where each of the three dimensions is divided into several criteria. These criteria are weighted with an individual percentage of contribution to derive the final total score. There are general criteria for all industries and specific criteria for companies in a certain sector. The complete dataset consists of 4432 total scores for the different firms and years between 2001 and 2007. However, not every company receives a sustainability score by SAM every year, simply due to the fact that there are firms that are not willing to participate in the assessment process every year. To be more precise, only 185 companies were evaluated by SAM Group in every single of the seven assessment years. The companies in the dataset are a mixture of worldwide wellknown multinational companies, such as Adidas AG, Allianz SE, the CocaCola Company, and Sony Corporation, as well as rather regional established firms such as Eniro AB from Sweden or the Italian Beni Stabili SpA. It can be seen from Table 1.1 that the total scores over the whole time period range between a rather low rating of 4.97 and a very high score of 92.37, i.e., that the predefined range between 0 and 100 is actually utilized. Interestingly, the median and mean of the overall total scores are slightly above 50, and barely half of the companies received a sustainability score between 43 and 65.
1.3.2. Introduction to Markov Transition Matrices In this section, Markov transition matrices are used to analyze the evolution of the sustainability scores. A high degree of variation within the total scores would be Table 1.1 Minimum 4.97
1st quartile 43.60
Statistics on Total Score. Median 55.48
Mean 53.67
3rd quartile 65.19
Maximum 92.37
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
FA
7
Socially Responsible Investments
counterintuitive, due to the fact that sustainability is a longterm affair and thus should not be subject to largesized jumps, unless extraordinary events occur, e.g., an environmental disaster on an oil producer’s platform. For the following analysis, data of those companies are used for which the sustainability scores are available for two consecutive years. For the entire sixyear time period, this leads to a total dataset of 2125 observations. The calculation of the transition matrices is performed as follows: For every single year, companies are ranked by their sustainability score, whereby for every year the 25% best rated companies are assigned to the 1st quartile, the next 25% to the 2nd quartile, and so on. Based on this allocation, empirical transition probabilities from one of the four quartiles to any of the four quartiles after one year can be calculated.
1.3.3. Results of Markov Transition Matrices From the average oneyear transition probabilities in Table 1.2, it can be seen that the probability of staying in the current quartile is the highest and ranges from 47.53% for the 2nd quartile to 72.21% for the last quartile.Additionally, the probability decreases in the distance between two quartiles. Furthermore, the probability that a topranked firm will end up in the 4th quartile in the following year is only 0.37% and the probability of a “bad” company to be part of the first quartile in the following period is 1.23%. Moreover, Markov transition matrices for every single year 2001–2007 were scrutinized. The results for the single years are quite similar to the average observation in Table 1.2. Finally, a sixyear Markov transition matrix was computed. The results are shown in Table 1.3. Nearly half of the companies that were ranked in the first quartile in 2001 were still in the first quartile in 2007. The probability that a highly sustainable company will be part of the worst quartile at the end of the six years is 5.36% and the probability of the opposite case, i.e., a “bad” company ending as a sustainability leader after six years, is 7.02%. Table 1.2 Average OneYear Markov Transition Probabilities (Year 2001–2007). Next year quartile
Last year quartile
1 2 3 4
1 (%)
2 (%)
3 (%)
4 (%)
69.74 23.83 5.15 1.23
25.00 47.53 21.74 5.96
4.90 24.35 49.94 20.60
0.37 4.29 23.17 72.21
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
8
FA
Hroß et al. Table 1.3 Markov SixYear Transition Probabilities (Year 2001–2007). Next year quartile
Last year quartile
1 2 3 4
1 (%)
2 (%)
3 (%)
4 (%)
46.43 25.00 21.43 7.02
26.79 39.29 14.29 19.30
21.43 32.14 32.14 14.04
5.36 3.57 32.14 59.65
Altogether, the results provide evidence to the assumption that sustainability rankings do not have a high degree of shortterm variation.
1.4. SRI IN PORTFOLIO CONTEXT After having analyzed the sustainability of sustainability in the preceding section, this section will scrutinize how SRI can be evaluated with regard to the portfolio context. The main questions to be answered are whether investors shall add SRI investments to their portfolio, and if so, with which weighting. In the conducted portfolio case study, the SRI market is represented by the Advanced Sustainable Performance Index (ASPI). The ASPI is a European index consisting of 120 companies and is published by Vigeo Group, an extrafinancial supplier and rating agency in the field of sustainable development and social responsibility (for further information see [13]). In order to include dividend payments to the analysis, total return indices are used, i.e., dividends are reinvested. This approach has two main advantages. First, the index already represents a selected basket of the asset category SRI and the time series are readily available. Second, the predefined index is widespread and thus has the advantage that the companies’ specific risks are already eliminated by diversification. As a result, only the diversification effect of the asset class SRI itself is observed.
1.4.1. Description of the Dataset and Statistical Properties The portfolio analysis is based on daily logreturns of the asset classes bonds (represented by the JP Morgan Global Government Bond Index), stocks (represented by the Dow Jones Total Markets World Index), and SRI (represented, as described above, by the ASPI index) between 1 January 1992 and 30 September 2008. The main empirical statistics are shown in Table 1.4.
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
FA
9
Socially Responsible Investments Table 1.4
Empirical Statistics of Daily LogReturns.
Empirical statistics Mean Mean (annualized) Standard deviation Standard deviation (annualized) Skewness Excess kurtosis Autocorrelation: lag 1 Autocorrelation: lag 2 Autocorrelation: lag 3 Autocorrelation of squared returns (lag 1) 5% critical value for autocorrelation
Bonds
Stocks
SRI
0.00024 0.05917 0.00389 0.06151 −0.00619 1.44597 0.03266 0.00386 −0.01373 0.03859 0.03026
0.00026 0.06399 0.00803 0.12702 −0.29037 4.11292 0.16873 −0.02818 −0.02108 0.13695 0.03026
0.00038 0.09425 0.01216 0.19230 −0.13149 3.58599 0.00494 −0.02391 −0.06324 0.19301 0.03026
By comparing mean and standard deviation of bonds and stocks, it becomes evident that most of the risk–averse investors would invest the bulk of their wealth in bonds. This is due to the extremely high mean for bonds (5.92% per annum) combined with a low standard deviation. Additionally, bonds display the highest skewness and lowest excess kurtosis, which is generally preferred by risk–averse investors. As it is a debatable point whether past returns indicate the future in a sufficient way, experts’ forecasts about expected returns are often used to solve this shortcoming. By using the Black–Litterman approach to adjust the empirical returns, the empirical mean µemp itself as well as absolute and relative forecasts are taken into account (see, e.g., [14]). This approach can be interpreted as a linear combination of these two components at a given confidence level τ regarding the forecasts. The Black–Litterman expectations µBL can be expressed by (given that L is invertible) µBL = τ · L−1 q + (1 − τ) · µemp ,
(1.1)
where L represents the linear transformation Lµ of the asset classes expected return vector µ and for each forecast, whose actual value is specified in q. The assumptions about the forecasts are taken from [15]. This means that the annual return of bonds is expected to be 3.96% and the equity risk premium amounts to 3.5%. The additional assumption that the difference of the means of stocks and SRI does not change leads to 1 0 0 0.0396/250 L = −1 1 0 and q = 0.0350/250 . 0 −1 1 0.0303/250 For τ = 0, the Black–Litterman expectations are equal to the empirical means, while τ = 1 leads to expectations which are completely driven by the forecasts. Table 1.5 provides the Black–Litterman expectations for a confidence level of τ = 0.75.
May 12, 2010
17:46
WSPC/SPIB913
b913ch01
FA
10
Hroß et al. Table 1.5 Black–Litterman Expectations for Asset Class LogReturns. Black–Litterman expectations Daily return Annual return
Bonds
Stocks
SRI
0.00018 0.04449
0.00029 0.07195
0.00041 0.10221
Table 1.6 PValues of Jarque–Bera and Ljung–BoxQ Tests. Test
Null hypothesis
Jarque–Bera Ljung–BoxQ (Q1) Ljung–BoxQ (Q2) Ljung–BoxQ (Q3) Ljung–BoxQ (QS1)
Bonds
Stocks
Normal distribution K} )  Ft ]
− (1 − R)EQ [(1 + rI + 1{S2,T >K} )1{ς≤T }  Ft ]),
(6.19)
where ς is as in (6.5), rI the basic interest, the bonus payment rate described in the contract, and K the bonus barrier. In our framework, the risk–neutral price of the bonus guarantee certificate is given by BG(t, S1 , S2 ) = R(ZI,t + Cd,t (S2 , K)) D D + (1 − R)(ZI,t (S2 ) + Cd,t (S1 , S2 , K)).
(6.20)
To show the impact of debt, volatility, and correlation, we structure a bonus guarantee certificate with notional 88, bonus barrier K = 120, basic interest rI = 3.0%, and bonus payment rate = 3.4%.
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
135
Pricing Certificates Under Issuer Risk Bonus guarantee certificate 80 Defaultable Nondefaultable
75 70 65
Value
60 55 50 45 40 35 30
0
200
400
Figure 6.5
600 Debt level
800
1000
1200
Impact of debt.
For the bonus guarantee certificate, we observe the same typical feature of the graph, which shows prices for different debt levels, see Fig. 6.5, as for the other examples: the graph decreases sharply for lower debt levels and approaches R times the price of the analogous certificate of a nondefaultable issuer. In Fig. 6.6 the relationship between price and volatility is of nearly linear kind: the higher the volatility, the lower the price the investor has to pay. The impact of the correlation is less distinct as in the case of index certificates due to the fact that a major part of the value of the bonus guarantee depends on the zerocoupon bond.
6.3.5. Discount Certificates The risk protection of a discount certificate consists in a risk buffer: the investor buys the certificate at a discount on the actual value of the underlying. This risk limitation is again financed by a gain limit. The structure can be hedged by investing in the underlying and writing a call option, i.e., the value is specified by T
DC(t, S1 , S2 ) = S2,t − Ct (S2 , K) − (1 − R)e− t rs ds
× EQ S2,T − max[S2,T − K, 0] 1{ς≤T }  Ft ,
(6.21)
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
136
Götz et al. Bonus guarantee certificate
75
Value
70
65
60
1 0.5
55 0.7
0 0.6
0.5
0.4
−0.5 0.3
Volatility S1
0.2
−1
Correlation
Figure 6.6 Impact of volatility and correlation.
where ς is as in (6.5). The discount certificate can be valued by the following formula DC(t, S1 , S2 ) = R(S2,t − Ct (S2 , K)) D + (1 − R)(S2,t (S1 , S2 ) − CtD (S1 , S2 , K)).
(6.22)
We value a discount certificate with K = 120 in different scenarios. With respect to the debt level, the value of the discount certificate does not differ in its characteristics from the certificates analyzed before, see Fig. 6.7. In Fig. 6.8, we look into the dependence of the price on the issuer’s volatility and the correlation. Regarding the volatility, we see the known structure, i.e., the price falls when the volatility increases. The slopes are similar for all correlation scenarios. As the certificate consists of building blocks clearly dependent on the correlation, the impact of the correlation on the value of the certificate is similar to the index certificate example.
6.3.6. Bonus Certificates The investor in this certificate is protected from a decline of the underlying up to a certain point, the protection barrier B. Below this point, the investor fully participates in any fluctuations of the underlying. The same is true for the performance of the underlying beyond the bonus barrier K. The certificate can be fully hedged by an
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
137
Pricing Certificates Under Issuer Risk Discount certificate 75 Defaultable Nondefaultable
70 65
Value
60 55 50 45 40 35 30
0
200
400
600 Debt level
Figure 6.7
800
1000
1200
Impact of debt.
Discount certificate
75 70 65
Value
60 55 50 45 40 35 0.7
1 0.5 0.6
0.5
0 0.4
Volatility S1
0.3
0.2
−1
−0.5 Correlation
Figure 6.8 Impact of volatility and correlation.
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
138
Götz et al.
investment in the underlying and by buying a knockout put option. The respective barrier level is the protection barrier. When the issuer risk is incorporated in the pricing model, the price is specified by BC(t, S1 , S2 ) = S2,t + e−
T t
rs ds
− (1 − R)e−
EQ [max[K − S2,T , 0]1{τ2 >T }  Ft ]
T t
rs ds
EQ [(S2,T + max[K − S2,T , 0]1{τ2 >T } )1{ς≤T }  Ft ], (6.23)
where τ2 = inf(t ∈ (t0 , T ] : S2,t ≤ B),
(6.24)
where ς is as in (6.5). The following formula evaluates the payoff: BC(t, S1 , S2 ) = R(S2,t + Pk,t (S2 , K, B)) D D + (1 − R) S2,t (S1 , S2 ) + Pk,t (S1 , S2 , K, B) .
(6.25)
Finally, we show some exemplary computations of the bonus certificate. We assumed a protection barrier B = 80 and a bonus barrier K = 120 for our computations. Not surprisingly, the graph, i.e., Fig. 6.9, plotting the debt level of the issuer against the value of the bonus certificates shows the same features as in the examples above. Bonus certificate 110 Defaultable Nondefaultable 100
Value
90
80
70
60
50
40
0
200
400
Figure 6.9
600 Debt level
800
Impact of debt.
1000
1200
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
139
Pricing Certificates Under Issuer Risk Bonus certificate
100 90
Value
80 70 60 50 1
40 0.5 30 2.5
0 2
1.5
1
−0.5 0.5
0
−1
Volatility S1
Correlation
Figure 6.10 Impact of volatility and correlation.
In Fig. 6.10, we see a strong impact of correlation in high volatility scenarios and small influence in lower volatility scenarios. These characteristics can be explained by the fact that correlation between issuer and underlying has a considerable impact on the price of the building blocks of the bonus certificate.
6.4. CONCLUSION We have derived closedform expressions for index, discount, participation guarantee, and bonus certificates under issuer risk in a Black–Scholes model framework. Our scenario computations clearly depict that, depending on the issuer’s capital soundness, a pricing formula which neglects issuer risk considerably overprices the value of the singular certificate. Thus, for a retail investor, a simple comparison of the prices of certificates of different issuers is not appropriate in order to find the security with the best priceperformance ratio. For his investment decision, the investor has to take the rating and equity to debt ratio into consideration.
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
140
Götz et al.
A.1. PROOF OF PROPOSITION 6.1
D Proof. S2,t fulfils the following partial differential equation and boundary conditions
D D D ∂2 S2,t ∂2 S2,t ∂2 S2,t 1 1 2 + σ22 S22 + ρσ1 σ2 (S1 + D1,t )S2 σ1 (S1 + D1,t )2 2 2 2 2 ∂S1 ∂S2 ∂S1 ∂S2 D D D ∂S2,t ∂S2,t ∂S2,t D + (rt − d1,t )S1 + (rt − d2,t )S2 + = 0, − rt S2,t ∂S ∂S ∂t 1 2 D S2 (t, 0, S2 ) = 0, D S2 (T, S1 , S2 ) = S2,T . By introducing, the following transformations S2D (t, S1 , S2 ) = e−
T t
rs ds ατ+β1 y1 +β2 y2 D∗ e S2 (t, y1 , y2 ),
τ = T − t, S1,t + D1,t y1 = ln , D1,t
χ = σ1 σ2 τ,
T
y2 = ln (S2,t e t (rs −d2,s )ds ), σ2 z1 = y1 , σ1 1 σ1 − ρσ2 , β1 = 2 (1 − ρ2 )σ1 α=−
σ1 y2 , σ2 1 σ2 − ρσ1 β2 = , 2 (1 − ρ2 )σ2 z2 =
1 σ12 − 2σ1 σ2 ρ + σ22 , 4 2(1 − ρ2 )
the PDE can be reduced to D∗ 1 ∂2 S2D∗ ∂2 S2D∗ 1 ∂2 S2D∗ ∂S 2 − − − ρ = 0, 2 ∂z21 2 ∂z22 ∂z1 ∂z2 ∂χ S2 (χ, 0, z2 ) = 0, σ σ S D∗ (0, z , z ) = (e−β1 σ12 z1 +(1−β2 ) σ21 z2 ). 1 2 2 D∗
(A.1)
This problem can be rewritten in terms of its transition probability density p(χ, z1 , z2 , z1 , z2 ). p(χ, z1 , z2 , z1 , z2 ) is a fundamental solution of the partial differential equation above and satisfies the backward Kolmogorov equation, see [17], page 368f, with initial condition p(0, z1 , z2 , z1 , z2 ) = δ(z1 − z1 )δ(z2 − z2 ),
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
141
Pricing Certificates Under Issuer Risk
see [15], page 493. We apply the method of images, see [18], page 476ff, to restrict the fundamental solution to the area the problem is defined for. This method is appropriate when the region to which the solution should be bounded is highly symmetric: a solution for the free space is first derived (pF (χ, z1 , z2 , z1 , z2 )) and then restricted to the defined region via symmetry, i.e., the principle of reflection. In this case, the image point (z1 , z2 ) of the source point (z1 , z2 ) is given by (z1 , z2 ) = (−z1 , −2ρz1 + z2 ). Thus, the solution is found by p(χ, z1 , z2 , z1 , z2 ) = pF (χ, z1 − z1 , z2 − z2 ) − pF (χ, z1 + z1 , z2 + 2ρz1 − z2 ). In this case the solution for the free space is 2
pF (χ, z1 , z2 , z1 , z2 ) =
2
(z +z −2ρz1 z2 ) 1 − 1 2 e 2(1−ρ2 )χ . 2π 1 − ρ2 χ
The solution of the original Dirichlet problem can be found by computing, see [17], page 364ff, S2D∗ (χ, z1 , z2 ) =
∞
∞
−∞
0
∞
−
p (χ, z1 − F
−∞
∞
0
With
z1 , z2
−
−β1 z2 )e
σ1 z +(1−β2 ) σ2 1
pF (χ, z1 + z1 , z2 + 2ρz1 − z2 )e
ϕ= =e
∞ −∞
0
∞
−β1
σ2 z σ1 2
dz1 dz2
σ1 z +(1−β2 ) σ2 1
σ2 z σ1 2
dz1 dz2 .
pF (χ, z1 , z2 )ek1 z1 +k2 z2 dz1 dz2
χ(k12 +2ρk1 k2 +k22 ) 2
√ lim N2 ( χ(k1 + ρk2 ), z2 , ρ),
z2 →∞
we get Proposition 6.1.
A.2. PROOF OF PROPOSITION 6.2
Proof. The proof is similar to the one for the investment in the underlying. By the means of the Feyman Kac Theorem, see [19], page 143ff, the following PDE can be
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
142
Götz et al.
derived: 2 D 1 2 2 ∂2 CD 1 2 ∂2 C D 2∂ C (S + D ) + S + ρσ σ (S + D )S σ σ 1 1,t 1 2 1 1,t 2 1 2 2 2 2 ∂S1 ∂S2 ∂S12 ∂S22 ∂CD ∂CD ∂CD + (rt − D1,t )S1 + (rt − d2,t )S2 + − rt C D = 0, ∂S ∂S ∂t 1 2 D C (t, 0, S2 ) = 0, D C (T, S1 , S2 ) = max(S2 − K, 0). By applying the transformations, which have been introduced before the PDE can be reduced to D∗ ∗ ∗ ∗ 1 ∂2 CD ∂2 C D 1 ∂2 C D ∂C − − ρ = 0, − ∂χ 2 ∂z21 2 ∂z22 ∂z1 ∂z2 ∗ CD (χ, 0, z2 ) = 0, σ σ σ σ D∗ −β1 σ1 z1 +(1−β2 ) σ2 z2 −β1 σ1 z1 −β2 σ2 z2 2 1 2 1 −e , 0). C (0, z1 , z2 ) = max(e Using the method of images on the probability density function as before the solution is given by ∗
CD (χ, z1 , z2 )
∞
=
0
∞
0
−
∞
0
− 0
∞
p (χ, z1 −
z1 , z2
p (χ, z1 +
z1 , z2
F
0 ∞
0
−
pF (χ, z1 − z1 , z2 − z2 )e
∞
F
0 ∞
∞
0
with
0
=e
σ1 z +(1−β2 ) σ2 1
−
−β1 z2 )e
+
2ρz1
−
σ1 z −β2 σ2 1
−β1 z2 )e
pF (χ, z1 + z1 , z2 + 2ρz1 − z2 )e
∞
ϕ=
−β1
0
∞
−β1
σ2 z σ1 2
σ2 z σ1 2
dz1 dz2
dz1 dz2
σ1 z +(1−β2 ) σ2 1
σ1 σ2 z1 −β2
pF (χ, z1 , z2 )ek1 z1 +k2 z2 dz1 dz2
χ(k12 +2ρk1 k2 +k22 ) 2
√ √ N2 ( χ(k1 + ρk2 ), χ(k2 + ρk1 )).
σ2 σ1 z2
σ2 z σ1 2
dz1 dz2
dz1 dz2
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
FA
143
Pricing Certificates Under Issuer Risk
A.3. PROOF OF PROPOSITION 6.3
Proof. The PDE and the boundary conditions can be indicated by 2 D ∂2 CdD 1 2 2 ∂2 CdD 1 2 2 ∂ Cd (S + D ) + S + ρσ σ (S + D )S σ σ 1 1,t 1 2 1 1,t 2 1 2 2 2 2 2 2 ∂S1 ∂S2 ∂S1 ∂S2 ∂CD ∂CD ∂CD + (rt − d1,t )S1 d + (rt − d2,t )S2 d + d − rt CdD = 0, ∂S1 ∂S2 ∂t CdD (t, 0, S2 ) = 0, CdD (T, S1 , S2 > K) = 1. Applying the introduced transformations and using the method of images on the transformed probability density function, one gets ∗
CdD (χ, z1 , z2 ) =
1 K
∞ ∞ 0
− 0
0
∞ 0
∞
−β1
pF (χ, z1 − z1 , z2 − z2 )e
p (χ, z1 + F
z1 , z2
+
2ρz1
−
σ1 σ2 z1 −β2
−β1 z2 )e
σ2 σ1 z2
dz1 dz2
σ1 z −β2 σ2 1
σ2 z σ1 2
dz1 dz2
A.4. PROOF OF PROPOSITION 6.4
Proof. The barrier option can be denoted by T
PkD (t, S1 , S2 ) = EQ e− t rs ds max[K − S2,t , 0]1{τ2 >T } 1{τ1 >T }  Ft . The PDE and boundary conditions are given by ∂2 BPkD ∂2 PkD 1 2 2 ∂2 PkD 1 σ12 (S1 + D1,t )2 + S + ρσ σ (S + D )S σ 1 2 1 1,t 2 2 2 2 2 ∂S22 ∂S1 ∂S2 ∂S12 D D D ∂P ∂P ∂P + (rt − d1,t )S1 k + (rt − d2,t )S2 k + k − rt PkD = 0, ∂S1 ∂S2 ∂t PkD (t, 0, S2 ) = 0, PkD (t, S1 , B) = 0, D Pk (T, S1 , S2 ) = max(K − S2 , 0).
.
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
144
FA
Götz et al.
Introducing the following transformations PkD (t, S1 , S2 ) = Ke−
T t
rs ds α∗ τ+β1 x1 +β2 x2∗
e
∗
PkD (t, x1 , x2∗ ),
1 1 1 1 2 2 α∗ = − σ12 β1 − σ22 − (r − d2 ) β2 + σ12 β1 + σ22 β2 + ρσ1 σ2 β1 β2 , 2 2 2 2 x1 = ln
S1 + D1,t D1,t
τ = T − t, x∗ − b x1 −ρ 2 , σ2 1 − ρ2 σ1 1 σ σ − ρ 12 σ22 − (r − d2 ) 2 1 2 β1 = , (1 − ρ2 )σ1 σ2 z∗1 =
1
x2∗ = ln
S2 , K
b = ln
B , K
z∗2 =
x2∗ − b , σ2
β2 =
1 σ2 − ρσ1 r − d2 − 2 , 2 2 (1 − ρ )σ2 σ2 (1 − ρ2 )
the PDE can be reduced to ∗ ∗ ∗ ∂PkD 1 ∂2 PkD 1 ∂2 PkD − + − = 0, 2 ∂z2 2 ∂z22 ∂τ 1 D∗ ∗ Pk (τ, z1 , 0) = 0,
1 − ρ2 ∗ D∗ ∗ Pk τ, z1 , − z1 = 0, ρ √ D∗ ∗ ∗ −β1 (σ1 1−ρ2 z∗1 +σ1 ρz∗2 )−β2 (z∗2 σ2 +b) (0, z , z ) = max(e P 1 2 k √ 2∗ ∗ ∗ − e−β1 (σ1 1−ρ z1 +σ1 ρz2 )+(1−β2 )(z2 σ2 +b) , 0). This PDE can be solved by integrating the payoff over the probability density function, see [16] √ ∗ ∗ 2 ∗ ∗ PkD (τ, z∗1 , z∗2 ) = h(τ, z∗1 , z∗2 , y1∗ , y2∗ ) max(e−β1 (σ1 1−ρ y1 +σ1 ρy2 )−β2 (y2 σ2 +b)
− e−β1 (σ1
√
1−ρ2 y1∗ +σ1 ρy2∗ )+(1−β2 )(y2∗ σ2 +b)
, 0)dy2∗ dy1∗ ,
where ∞ q q 2 − (q2 +qt2 ) nπθt nπθ t e 2τ sin sin I nπϕ , ϕτ n=1 ϕ ϕ τ 1 − ρ2 , ϕ ∈ [0, π], tan ϕ = − ρ
h(τ, z∗1 , z∗2 , y1∗ , y2∗ ) =
May 12, 2010
17:47
WSPC/SPIB913
b913ch06
Pricing Certificates Under Issuer Risk
tan θ =
y2∗ , y1∗
FA
145
θ ∈ [0, ϕ],
z∗2 , θt ∈ [0, ϕ], z∗1 q = y∗1 2 + y∗2 2 ,
tan θt =
qt =
z∗1 2 + z∗2 2 .
While the region is given as follow:
1 − ρ2 ∗ (−∞, 0) × −∞, − y1 ∪ (0, +∞) ρ
1 − ρ2 ∗ y1 , +∞ ∪ (−∞, 0) (−∞, 0) × − ρ
= 1 − ρ2 ∗ y1 (0, +∞) × 0, − ρ
2 1 − ρ (−∞, 0) × − y1∗ , 0 ρ
if {ρ > 0} ∩ {y1∗ > 0} if {ρ > 0} ∩ {y1∗ < 0} if {ρ < 0} ∩ {y1∗ > 0} if {ρ < 0} ∩ y1∗ < 0 ,
where 0 denotes the values in t. The region is obtained from the relationships arctan(y2∗ /y1∗ ) ∈ [0, arctan(− 1 − ρ2 /ρ)] and arctan(− 1 − ρ2 /ρ) ∈ [0, π].
References [1] Canabarro, E and D Duffie (2003). Measuring and marking counterparty risk. In: Asset/Liability Management for Financial Institutions, L Tilman (ed.). Institutional Investor Books. [2] Pykhtin, M and S Zhu (2006). Measuring counterparty credit risk for trading products under Basel II, Working Paper. [3] Pykhtin, M and S Zhu (2007). A guide to modeling counterparty credit risk, GARP Risk Review. [4] Merton, RC (1974). On the pricing of corporate debt: The risk structure of interest rates. Journal of Finance, 29, 449–470. [5] Johnson, H and R Stulz (1987). The pricing of options with default risk. Journal of Finance, 42, 267–280. [6] Hull, JC and AD White (1995). The impact on default risk on the prices of options and other derivative securities. Journal of Banking and Finance, 19, 299–322. [7] Jarrow, R and S Turnbull (1995). Pricing derivatives on financial securities subject to credit risk. Journal of Finance, 50, 53–85.
May 12, 2010
146
17:47
WSPC/SPIB913
b913ch06
FA
Götz et al.
[8] Cherubini, U and E Luciano (2002). Copula vulnerability. Risk, 15, 83–86. [9] Cherubini, U and E Luciano (2004). Pricing swap credit risk with copulas, Working Paper. [10] Klein, P (1996). Pricing Black–Scholes options with correlated credit risk. Journal of Banking and Finance, 20, 1211–1229. [11] Klein, P and M Inglis (2001). Pricing vulnerable European options when the option payoff can increase the risk of financial distress. Journal of Banking and Finance, 25, 993–1012. [12] Stamicar, R and C Finger (2005). Incorporating equity derivatives into the CreditGrades model, RiskMetrics Group. [13] Sepp,A (2006). Extended CreditGrades model with stochastic volatility and jumps. Wilmott Magazine, 54, 50–62. [14] Hull, JC (2002). Options, Futures, and Other Derivatives, 5th edn. Prentice Hall. [15] Lipton, A (2001). Mathematical Methods for Foreign Exchange: A Financial Engineer’s Approach. Singapore: World Scientific Publishing Co. Pte. Ltd. [16] He, H et al. (1998). Double lookbacks. Mathematical Finance, 8, 201–228. [17] Karatzas, I and SE Shreve (1991). Brownian Motion and Stochastic Calculus, 2nd edn. Berlin: SpringerVerlag. [18] Zauderer, E (2006). Partial Differential Equations of Applied Mathematics. John Wiley & Sons, Inc. [19] Øksendal, B (2003). Stochastic Differential Equations, 6th edn. Berlin: SpringerVerlag.
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
FA
7
ASSET ALLOCATION WITH CREDIT INSTRUMENTS
BARBARA MENZINGER∗,‡ , ANNA SCHLÖSSER∗,§ and RUDI ZAGST†,¶ ∗
risklab GmbH, Seidlstr. 2424a, 80335 Munich, Germany † HVBStiftungsinstitut für Finanzmathematik, Technische Universität München, Parkring II, 85748 Garching, Germany ‡
[email protected] §
[email protected] ¶
[email protected] This chapter presents a consistent, scenariobased asset allocation framework for analyzing traditional financial instruments and credit instruments in a portfolio context. Our framework accounts for the distinct return characteristics of credit instruments by incorporating potential defaults into the total return calculation. We generate correlated default times with a Normal Inverse Gaussian onefactor copula. To determine optimal portfolios, we use a meanvariance and a conditional value at risk optimization. Performing a case study for the U.S. market, we find that the meanvariance optimization overestimates the benefits of lowrated credit instruments. Though, optimal portfolios always contain a considerable proportion of credit instruments.
147
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
148
FA
Menzinger et al.
7.1. INTRODUCTION The credit market experienced an enormous growth over the last years. As part of the credit market, the credit derivatives market was the worldwide fastestgrowing derivatives market (see [1]). The market size expanded from USD 180 billion in 1996 via USD 5,021 billion in 2004 to USD 20,207 billion in 2006 and to estimated USD 33,120 billion in 2008 (see [2]). Due to the financial crisis, the growth was rapidly stopped in 2008 and turned into a decline of the market size. Still, the actual notional amount outstanding of the credit default swaps (CDS), which are the most important instruments in this market, exceeded the estimation of the British Bankers’ Association by far with a notional amount outstanding of USD 41,868 billion (see [3]). After a standardization of credit derivative contracts and the introduction of CDS indices, a revolution in terms of liquidity has taken place, which is only one reason why credit instruments are very attractive to investors. In addition, credit instruments such as corporate bonds, credit derivatives, and securitizations often have an appealing risk–return profile allowing to enhance the portfolio return. Furthermore, due to the correlation structure of their returns to those of traditional asset classes such as stocks and government bonds, they offer high potential for diversification. Finally, they allow to manage the credit risk exposure. Even knowing the potential benefits of different credit instruments, investors still have to know how to combine them optimally with traditional asset classes, i.e., they have to decide on the optimal proportion of creditrelated products, especially for their individual level of risk–aversion. A reasonable asset allocation including credit instruments needs to account for the distinct return characteristics of these instruments. We exemplarily analyze the return properties of daily logreturns of U.S. Lehman aggregates over the period from 7 November, 2002 to 29 September, 2006. The descriptive statistics of the returns are summarized in Table 7.1, where it becomes obvious that the risk–return profile of credit instruments cannot be sufficiently described by mean and variance alone.
Table 7.1
U.S. aggregate Aaa Aa A Baa Corporates
Descriptive Statistics of U.S. Lehman Aggregates. No. of issuers
Mean (Ann.)
Volatility (Ann.)
Skewness
Excess kurtosis
3560 759 1404 1215 2721
3.50% 4.23% 5.30% 6.99% 5.65%
3.27% 4.18% 4.55% 4.72% 4.55%
−0.1310 −0.2944 −0.1787 −0.0508 −0.1979
1.7812 1.5802 1.4967 1.3991 1.2648
May 12, 2010
17:47
WSPC/SPIB913
Asset Allocation with Credit Instruments
b913ch07
FA
149
We observe negatively skewed return distributions and positive excess kurtosis indicating nonnormal distribution of return data. This is confirmed by Jarque–Bera tests for normality. The nonnormality might be attributed to defaults within the aggregates. The logical consequence is that the shape of the return distributions should be considered in a realistic asset allocation framework. In this chapter, we propose a consistent, scenariobased asset allocation framework that is composed of a simulation, a total return calculation, and an optimization. With the simulation framework, we produce consistent capitalmarket scenarios for interest rates, credit spreads, and returns of equity indices, on the one hand, and correlated default times, on the other hand. For the scenario generation, we use the fourfactor model according to [4] and the model suggested by [5]. To generate nonnormal return distributions of credit instruments, we simulate correlated default times for issuers with different ratings. In this context, the onefactor copula approach for modeling correlated default times between reference entities has become very popular. We use this approach for a NormalInverseGaussian (NIG) onefactor copula as suggested by [6] since it is able to produce more realistic properties for default times than the widespread Gaussian version, e.g., a higher probability of joint defaults of different companies. This property can be predominantly observed during a crisis. Then the probability of joint defaults increases and the value of a portfolio of credit instruments can decrease tremendously, as recently seen during the subprime crisis. Therefore, it is particularly important that events having a large impact on the portfolio value are modeled realistically. After having determined the return distribution of the government bonds, the equity index, and the credit instruments, we are able to calculate optimal asset allocations according to different optimization criteria. We do not only apply traditional meanvariance portfolio optimization according to [7] but also conditional value at risk (CVaR) optimization. The latter is more appropriate to capture the distinct distributional properties of credit instruments as it takes into account the left tail of a return distribution. We perform a case study, applying the asset allocation framework to the U.S. market. So we can show that investors benefit from adding creditrelated products to their portfolio, i.e., with the same level of risk they can generate a higher expected return. This chapter is organized as follows. The second section describes the simulation framework for the risk factors and the correlated default times. In the third section, we give the general conditions for the total return calculations. In particular, we explain the pricing of funded CDS and funded CDS indices. The applied optimization framework is presented in the fourth section. In the fifth section, we provide the model parameters and present the simulation and optimization results. We close in the last section with a summary of the main results.
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
150
FA
Menzinger et al.
7.2. SIMULATION FRAMEWORK The simulation framework consists of a model to simulate risk factors on the one hand and correlated default times on the other hand. First, we introduce a model to simulate risk factors. We simulate the short rate, the growth rate of the Gross Domestic Product (GDP), and the shortrate spread with the extended model of Schmid and Zagst (see [4]). The short inflation and the returns of an equity index are simulated with the integrated market model suggested by [5]. The latter can be embedded into the extended model of Schmid and Zagst. In the following, we assume that markets are frictionless and perfectly competitive, that trading takes place continuously, that there are neither taxes nor transaction costs or informational asymmetries, and that investors act as price takers. We fix a terminal time horizon T ∗ . Uncertainty in the financial market is modeled by a complete probability space (, G, P) and all random variables and stochastic processes introduced below are defined on this probability space. We assume that (, G, P) is equipped with three filtrations H, F, and G, i.e., three increasing families of subσfields of G. The default time τ of an obligor is an arbitrary nonnegative random variable on (, G, P). For the sake of convenience, we assume that P[τ = 0] = 0 and P[τ > t] > 0 for every t ∈ (0, T ∗ ]. For a given default time τ, we introduce the associated default indicator or hazard function H(t) = 1{τ≤t} , t ∈ (0, T ∗ ]. Let H = (Ht )0≤t≤T ∗ be the filtration generated by the process H. In addition, we define the filtration F = (Ft )0≤t≤T ∗ as the filtration generated by the multidimensional standard Brownian motion Wt = (Wr,t , Wω,t , Wu,t , Ws,t ) and G = (Gt )0≤t≤T ∗ as the enlarged filtration G = H ∨ F, i.e., for every t we set Gt = Ht ∨Ft . All filtrations are assumed to satisfy the usual conditions of completeness and rightcontinuity. For the sake of simplicity, we furthermore assume that F0 is trivial. It should be emphasized that τ is not necessarily a stopping time with respect to the filtration F but of course with respect to the filtration G. If we assumed that τ was a stopping time with respect to F, then it would be necessarily a predictable stopping time. We assume that there exists a measure Q ∼ P such that all discounted price processes of the financial instruments are martingales relative to (Q, G).As numéraires, t we choose the moneymarket account Bt = e 0 rl dl with rt denoting the nondefaultable short rate. We will assume throughout the chapter that for any t ∈ (0, T ∗ ] the σfields FT ∗ and Ht are conditionally independent (under Q) given Ft . Following [8], p. 167 and p. 242, this is equivalent to the assumption that for any t ∈ (0, T ∗ ] and any Qintegrable FT ∗ measurable random variable X, we have EQ [X  Gt ] = EQ [X  Ft ]. We start by defining all processes under the realworld measure P and will then work under the equivalent martingale measure Q.
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
Asset Allocation with Credit Instruments
FA
151
The dynamics of the nondefaultable nominal short rate r is described by a twofactor Hull–White model and is given by drt = (θr,t + br ωt − ar rt )dt + σr dWr,t ,
(7.1)
dωt = (θω − aω ωt )dt + σω dWω,t ,
(7.2)
where ar , br , σr , aω , and σω are positive constants, θω is a nonnegative constant, θr,t is a continuous, deterministic Nelson–Siegel function that fits the market term structure at simulation start date. (For details see, for example, [9, 10].) dWr,t and dWω,t are independent standard Brownian motions. ω represents the GDP growth rate so that the interestrate levels directly depend on general economic conditions. The dynamics of the shortrate spread are described by dst = [θs + bsu ut − bsω ωt − as st ]dt + σs dWs,t ,
(7.3)
dut = [θu − au ut ]dt + σu dWu,t ,
(7.4)
where bsu , bsω , as , σs , au , and σu are positive constants, θs , θu are nonnegative constants and dWu,t , dWs,t are independent standard Brownian motions. u is the socalled uncertainty index and can be interpreted as an aggregation of all available information about the quality of the firm, i.e., it represents the firmspecific risk. The higher its value, the lower the firm’s quality. The GDP behaves reversely. If it grows at a higher rate, spreads usually tighten as the probability of default of a firm becomes smaller. We need the dynamics of the short inflation i for the modeling of the equityreturn. It is given by dit = (θi − ai it )dt + σi dWi,t ,
(7.5)
where ai and σi are positive constants, θi is a nonnegative constant and dWi,t is a standard Brownian motion. Then the continuous return of an equity index RE,t is described by dRE,t = [αE + bEω ωt − bEi it + bER rt ]dt + σE dWE,t ,
(7.6)
where αE ∈ R, bEω , bEi , bER , and σE are positive constants and dWE,t is a standard Brownian motion. Note that this process reflects the socalled “stock returninflation puzzle” stating that inflation negatively influences stock returns (see for example [11]– [15]). So far, we considered the dynamics of the SDEs in Eqs. (7.1)–(7.4) under the realworld measure P. Though, as we are interested in zerocoupon bond prices as well as interest rates and credit spreads for different terms, we need to know the parameters of the SDEs under the risk–neutral equivalent martingale measure Q. Replacing ar , aω , as , au by aˆ r , aˆ ω , aˆ s , aˆ u and using independent standard Brownian ˆ ω,t , d W ˆ s,t , d W ˆ u,t instead of dWr,t , dWω,t , dWs,t , dWu,t , leads to the ˆ r,t , d W motions d W
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
152
FA
Menzinger et al.
processes under this measure. The relationship between the parameters is given by aˆ k = ak + λk σk2 , where k = r, ω, s, u. We obtain λr , λω , λs , λu when changing measure from P to Q by applying Girsanov’s Theorem (see, for example, [4]). The processes introduced earlier imply analytical formulas for the zero rates at time t to maturity T of nondefaultable and defaultable bonds, R(t, T ) and Rd (t, T ), as well as for the credit spreads, S(t, T ). So, we can determine the complete term structure. The explicit formulas can be found in the Appendix. To simulate correlated default times, we use Li’s approach, who presents an efficient algorithm in [16]. We use it with a onefactor copula as applied for example by [17]. The onefactor copula is a simple, but powerful way to quickly define a correlation structure between several variables. Therefore, it has become very popular and the standard approach in practice to simulate correlated default times. The underlying idea is the following: In reality, more firms default during a recession than during a booming period. This implies that each firm is subject to the same set of macroeconomic environment and that there exists a dependence among the firms. The Gaussian onefactor copula is often applied due to an easy implementation and the appealing properties of a standard normal distribution, such as the stability under convolution. We rather use the NIG copula which is able to overcome some modeling deficiencies of the Gaussian copula, e.g., the lack of tail dependence. The NIG distribution is a mixture of normal and inverse Gaussian distributions. It is a four parameter distribution with very interesting properties. It can produce fat tails and skewness, it is stable under convolution (under certain conditions), and the density function, the distribution function and the inverse distribution function can be computed sufficiently fast (see [6] and [18]). A definition of the NIG distribution and its most important properties can be found in [6]. Before presenting the copula model, we explain the idea of the Large Homogeneous Portfolio (LHP) approach, introduced by [19]. It assumes a constant default correlation structure over the reference credit portfolio, with the same default probabilities and the same recovery rate in case of default, and it models default using a onefactor Gaussian copula. Reference [6] modified the LHP model by replacing the Gaussian distribution with the NIG distribution. Their onefactor copula model is briefly introduced in the following. Consider a homogeneous portfolio of m credit instruments. The standardized asset return of the ith issuer in the portfolio, Ai , is assumed to be of the form Ai = aM + 1 − a2 Xi (7.7) with independent random variables
βγ 2 γ 3 M ∼ N IG α, β, − 2 , 2 α α
,
(7.8)
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
FA
153
Asset Allocation with Credit Instruments
and
√ √ √ √ 1 − a2 1 − a2 1 − a2 βγ 2 1 − a2 γ 3 , (7.9) Xi ∼ N IG , α, β, − a a a α2 a α2 where γ = α2 − β2 . Then, the asset returns Ai also follow an NIG distribution with the parameters βγ 2 γ 3 α β , , − 2 , 2 =: N IG A Ai ∼ N IG (7.10) a a aα aα
and Ai has zero mean and unit variance. The factor M represents the systematic common market factor and Xi represents firmspecific factors. Equation (7.7) defines a correlation structure between the random variables Ai . Then, the correlation between the asset returns of two issuers is given by a2 , in the case of a homogeneous portfolio. Conditional on M the asset returns of different issuers are independent. Let us assume that default at time t occurs when the asset return of obligor i falls below the threshold C(t), i.e., Ai ≤ C(t). Using this copula model, the variable Ai is then mapped to default time τi of the ith issuer with a percentiletopercentile transformation as described for example by [17] or [20]: p(t) = P[τi ≤ t] = P[Ai ≤ C(t)] = N IG A (C(t)).
(7.11)
p(t) = P[τi ≤ t] is the realworld distribution of default times estimated with a migration matrix following [20]. According to Eq. (7.11), we conclude that C(t) = N IG −1 A (p(t)) and thus −1 P[τi ≤ t] = P[Ai ≤ N IG −1 A (p(t))] = P[p (N IG A (Ai )) ≤ t].
Now, we can simulate Ai according to Eq. (7.7) and determine the default times via τi = p−1 (N IG(Ai )).
7.3. FRAMEWORK FOR TOTAL RETURN CALCULATION Based on the simulations according to the previous section, we price the following financial instruments: Government and corporate coupon bonds, funded CDS and funded CDS indices. The pricing for all instruments is done under the following conditions. We price the instruments at every simulated time step tk , k ∈ {0, . . . , p}, with t0 < t1 < · · · < tp = T sim , where T sim denotes the end date of the simulation. We assume that the cash flows occurring during the lifetime of the instrument are reinvested in the respective instrument. For simplicity reasons, we assume a constant recovery rate REC of 40% for all credit instruments in the case of a default rather than using a stochastic recovery model. The constant recovery rate of 40% is in accordance with the recovery rates for the iTraxx Europe (see www.indexco.com). The cash flows
May 12, 2010
154
17:47
WSPC/SPIB913
b913ch07
FA
Menzinger et al.
at default time τ are invested in a riskfree cash account and they are compounded at every simulated time step with the default riskfree government rate. We always assume an initial investment of one unit. The pricing and total return calculation of the bonds and the equity index is straightforward. The pricing of the funded CDS and the funded CDS index, however, is not trivial and is therefore derived in the following. A CDS can be issued in a funded version. Then, the investor (protection seller) buys a floating rate note (FRN), which pays a coupon of the threemonth LIBOR plus the fixed CDS spread on a quarterly basis. If no credit event occurs, the coupon is paid until maturity. In this case, the investor receives the notional amount of the FRN. If, however, a credit event of the reference entity takes place, the investor receives the recovery value and the contract terminates. The functionality of a funded CDS index is similar to funded CDS. It can be considered as a portfolio of funded CDS. If a credit event of a reference entity takes place, the investor receives the recovery value, but the notional amount of the FRN is reduced by the weight of the reference entity. If we assume, for example, the first default in the reference portfolio containing 125 equally weighted names, like the broadest and most actively traded investmentgrade indices CDX.NA.IG and the iTraxx Europe for North America and Europe, respectively, the notional amout is reduced by 1/125 from 100% to 99.2%. The future coupon payments are based on the new notional amount while the coupon rate remains unchanged. We consider CDS and the CDS index from an investor’s perspective. Hence, we view these instruments as an investment rather than as a means of hedging. Furthermore, we use the funded version of CDS and the CDS index for two reasons. First, we can exclude counterparty risk from our considerations. Second, there are investors who are for regulatory restrictions or due to internal investment policies not allowed to enter into unfunded creditderivative contracts. Before explaining the total return components of these instruments, we briefly describe the relationship between the relevant time steps. At every simulated time step, tk ∈ {0, . . . , T sim }, we price the funded CDS/CDS index. Furthermore, we assume that t1c < · · · < tnc = T denote the spread payment dates, where t1c denotes the next spread payment date following the current pricing day, tk , and T denotes the maturity of the CDS/CDS index. We denote the previous spread payment date or the settlement date, if the first spread payment has not yet been made, by t0c . Spread payments are made in c to tic . arrear — at time tic for the payment period from ti−1 A funded CDS/CDS index has the following total return components: Present value of the pure, default riskfree FRN at tk paying LIBOR (PVFRN ), coupon payments on the notional amount at t0c , comprising LIBOR and the fixed spread S CDS , compounded to the following simulated date tk (PVS ), recovery payments at default time τ, compounded to the following simulated date tk (PVREC ), and the present value of the CDS/CDS index (PVCDS ). PVCDS can be calculated for every pricing day tk . Its value can be either
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
FA
Asset Allocation with Credit Instruments
155
positive, or negative, or zero depending on possible changes in the default intensity λk at tk and on changes in the interestrate curve. At inception of a funded CDS/CDS index, the fixed spread is determined so that PVCDS is equal to zero, i.e., the fixed spread is a fair price for the default intensity of the reference entity. During the term of the contract, however, the default intensity of the reference entity can change, while the fixed spread does not, resulting in a change in the value of PVCDS . To determine the present value of the funded CDS/CDS index at tk , we need to know the expected loss at tk up to every spread payment day, tic , i ∈ {1, . . . , n}, the premium leg and the protection leg. With the constant default intensity model, we calculate the expected loss at tk up to tic according to EL(tk , tic , λk ) = 1−exp(−λk (tic − tk )). This model assumes a constant default intensity, λk , for a given time tk . (This implies that the partial recovery of market value in the extended model of Schmid and Zagst is stochastic over time.) The premium leg is the present value of all expected spread payments. It is calculated according to Prem Leg(tk , T, S CDS , λk ) =
n
tic S CDS (1 − EL(tk , tic , λk ))P(tk , tic ),
i=1 c where tic = tic −ti−1 , S CDS is the fixed annual spread of the CDS/CDS index, P(tk , tic ) is the discount factor and 1 − EL(tk , tic , λk ) denotes the probability of no default up to time tic . The protection leg is the present value of all expected protection payments made by the protection seller. As we assume a constant recovery rate REC, we can calculate the protection leg by T Prot Leg(tk , T, λk ) = (1 − REC) P(tk , l) dEL(tk , l, λk ) tk
≈ (1 − REC)
n
c (EL(tk , tic , λk ) − EL(tk , ti−1 , λk ))P(tk , tic ).
i=1
At issuance of the funded CDS, the fixed annual spread S CDS is determined so that the value of the premium leg equals the value of the protection leg: c (1 − REC) ni=1 (EL(tk , tic , λk ) − EL(tk , ti−1 , λk ))P(tk , tic ) n S CDS = . (7.12) c c c i=1 ti (1 − EL(tk , ti , λk ))P(tk , ti ) The equality of premium leg and protection leg as shown in Eq. (7.12) must also hold for every day, when S CDS is substituted by the quoted par yield spread. For the sake of simplicity, we use the simulated zero spread as the quoted par yield spread. The absolute difference between the spreads, however, is very low. We analyzed historical credit spreads from 31 December, 1991 to 29 September, 2006. The average difference
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
156
FA
Menzinger et al.
in this period was at most 1.5 bps for rating classes AA, A2 and BBB and maturities of up to five years. The worst case on a single day was 4.9 bps. After this substitution, we can solve the equation for λk and determine the constant (implied) default intensity at tk . Let N(tk , τ) denote the notional amount of a funded CDS/CDS index at time step tk . Then, the present value of the CDS/CDS index is given by PVCDS (tk , T, S CDS , λk ) = (Prem Leg(tk , T, S CDS , λk ) − Prot Leg(tk , T, λk ))N(tk , τ).
(7.13)
The calculation of the present values of the other total return components is straightforward. Then, we determine the total return of the investment, RCDS , for every tk > 0 by RCDS (tk−1 , tk ) =
PVCDS (tk , T, S CDS , λk ) + PVFRN (tk ) + PVREC (tk ) + PVS (tk ) − 1. PVCDS (tk−1 , T, S CDS , λk−1 ) + PVFRN (tk−1 ) + PVREC (tk−1 )
7.4. OPTIMIZATION FRAMEWORK For all optimization criteria, we make the following assumptions. There are neither transaction costs nor taxes; all securities can be divided arbitrarily; the portfolios remain unchanged over time; there are n given assets to invest in with returns Ri , i = 1, . . . , n; the expected return of asset i is given by µi := EP [Ri ] and µ := (µ1 , . . . , µn )T ; xi is the portfolio weight of asset i with ni=1 xi = 1, and the portfolio is denoted by x := (x1 , . . . , xn )T .
7.4.1. MeanVariance Optimization The meanvariance optimization can be viewed as traditional portfolio optimization and is based on the model of [7]. The basic assumption is that investors select their portfolios taking into account only the first two moments of the asset returns — mean and variance — and the correlation between the assets. Let the covariance matrix be denoted by C = (cij )i,j=1,...,n , where cij = Cov[Ri , Rj ], and let the variance be denoted by σi2 := cii > 0. Then, the meanvariance optimization is given by the following optimization problem min x
xT Cx
(7.14)
s.t. µT x ≥ µ, ¯ 1T x = 1, x ≥ 0, where 1 = (1, . . . , 1)T . If we solve the optimization problem in Eq. (7.14) for every possible µ, ¯ we obtain the set of all efficient portfolios. Obviously, the optimization problem in Eq. (7.14) only takes into account mean and variance of a portfolio. This is appropriate, for example, if returns are normally distributed. However, the returns of credit instruments are not normally distributed,
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
FA
157
Asset Allocation with Credit Instruments
as seen earlier. Therefore, this criterion is only of limited use for the optimization of portfolios including credit instruments.
7.4.2. CVaR Optimization The conditional value at risk (CVaR) represents the expected value of all losses that exceed a certain value at risk (VaR). Formally, we can define the CVaR as follows. Let 1 − α be the confidence level for the VaR with α ∈ (0, 1). Then, the CVaR of a portfolio’s return R(x) is given by CVaR(x, α) = −EP [R(x)  R(x) ≤ CR (α)] = −EP [R(x)  R(x) ≤ −VaR(x, α)],
(7.15)
where CR (α) denotes the αquantile of the portfolio’s return distribution R(x). From the formulas above, it becomes evident that the CVaR provides information on the negative tail of a return distribution since it is not only focussed on the αquantile but also takes into account the shape of its tail. This is of great importance if instruments may default and so produce fat tails. The corresponding CVaR optimization is given by min x
CVaR(x, α)
(7.16)
s.t. µT x ≥ µ, ¯ 1T x = 1, x ≥ 0.
7.5. MODEL CALIBRATION AND SIMULATION RESULTS We fit our model to market data as of 30 September, 2006 (simulation start date). To calibrate the simulation model, we use parameters estimated by [21] and [22]. The parameters for the short rate and the GDP growth rates are given by ar = 0.37867, aˆ r = 0.24782, br = 0.13315, σr = 0.01496, aω = 1.18532, aˆ ω = 0.26847, θω = 0.01583, σω = 0.00601. The credit spread parameters are displayed in Table 7.2. We adjust the estimated parameters aˆ s , aˆ u , and bsu to better meet historical data in terms of average spreads and spread ranges. For the inflation process and the process for the equityindex returns, we use the following parameters (see [5] and [22]): ai = 0.64073, aˆ i = 0.50319, θi = 1.04790, σi = 0.01447. αE = −3.07791, bER = 3.94000, bEi = 9.32436, bEω = 5.10643, σE = 0.16000. For the equity index, we adjust the level of the returns and the standard deviation to better reflect the updated historical data. We want the U.S. CDS index to represent the current composition of the CDX.NA.IG in terms of proportion of different rating classes. However, we assume the U.S. CDS index (short CDX) only to be composed of the rating classes AA, A, and BBB with the weights 12.28%, 39.47%, and 48.25%. We added the AAA rating class
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
158
FA
Menzinger et al. Table 7.2 Parameter Estimation for Processes s and u for Rating Classes AA, A2, BBB1.
as aˆ s θs σs bsω au aˆ u θu σu bsu
AA
A2
BBB1
2.96099 2.96099 0.00275 0.00389 0.07373 0.11269 0.02180 0.12173 0.00459 0.98404
2.80727 2.80727 0.00251 0.00309 0.07634 0.11057 0.03210 0.18566 0.00476 0.92214
2.41739 2.41739 0.00238 0.00287 0.09369 0.11048 0.04075 0.18980 0.00489 1.26089
to rating class AA and BB to BBB, as there are only parameters for the rating classes AA, A, and BBB available. Furthermore, we normalized the weights to sum up to one. For the simulation of correlated default times, we use the average oneyear migration matrix for the United States, from 1981 to 2005, provided by [23]. The parameters for the onefactor copula as of simulation start date are estimated in accordance with the model introduced by [6]. We use the following parameters a = 0.37029, α = 0.70138, β = 0. So, the correlation a of a single reference entity to the common market factor is equivalent to a correlation between two reference entities of a2 = 0.13711. Having simulated 5,000 scenarios on a quarterly basis with government interest rates, credit spreads and correlated default times, we can now price various financial instruments. All bonds, CDS and the CDX, are assumed to have an initial term of approximately five years. Table 7.3 reports the return characteristics for investment horizons one year and three years. Here, we make some general observations. We see higher expected returns and higher standard deviations or levels of CVaR with decreasing credit quality of the bonds. Besides, we observe that — as soon as defaults have occurred — the distributions of credit instruments become nonnormal. This can be shown by some statistics: The skewness of the credit instruments is negative, the excess kurtosis is significantly higher than 0 and minimum values always indicate defaults. The nonnormality can also be verified easily by Jarque–Bera tests. For the one year investment horizon, the AA and Arated CDS and the CDX have considerably less volatility and a lower CVaR than government and corporate bonds. At the same time, CDS and the CDX have a similar expected return to government bonds. Though, comparing the oneyear and the threeyear investment horizon, we make an interesting observation. For the threeyear horizon, the CDS have a similar or higher risk than the corporate bonds and a
Bond AA
Bond A
Bond BBB
CDS index
CDS AA
CDS A
CDS BBB
5.55% 5.50% 3.16% 2.55% −4.35% 17.01% 0.0900 0.0276
10.43% 8.25% 19.30% 32.14% −43.08% 88.10% 0.5568 0.4571
6.03% 6.01% 3.35% 3.24% −59.17% 17.18% −1.3820 28.2614
6.19% 6.18% 3.57% 5.34% −58.22% 18.12% −3.3200 59.8199
6.37% 6.51% 4.73% 17.71% −58.17% 19.63% −6.7629 90.1579
5.49% 5.52% 0.74% −2.60% −13.27% 7.59% −5.1385 109.6775
5.75% 5.77% 1.13% −2.64% −57.44% 8.46% −34.6157 1927.6047
5.92% 5.95% 1.70% −0.39% −57.72% 8.54% −30.6476 1135.5933
6.10% 6.27% 3.46% 13.24% −57.28% 10.00% −16.6226 298.5225
16.57% 16.58% 2.67% −9.41% 7.32% 25.13% −0.0188 −0.0459
33.44% 24.29% 52.74% 54.45% −67.75% 439.46% 1.3790 3.8584
18.04% 18.16% 3.86% −1.53% −55.18% 27.26% −9.2575 172.4228
18.55% 18.86% 5.34% 14.83% −55.41% 28.12% −9.8637 131.3360
19.08% 19.99% 8.77% 53.12% −55.20% 30.60% −7.1960 56.2670
16.43% 16.46% 3.33% −5.77% −30.08% 27.25% −0.9070 10.8976
17.20% 17.25% 4.18% 0.65% −54.21% 29.06% −7.3939 123.0516
17.73% 17.93% 5.43% 15.09% −54.58% 30.19% −8.4513 107.1262
18.41% 19.10% 8.06% 51.90% −54.77% 32.05% −7.1697 59.9592
1 Year Mean Median σ CVaR(1%) Min Max Skewness Excess kurtosis
WSPC/SPIB913
Equity index
17:47
Govt. bond
May 12, 2010
Key Statistics of Total Returns of Financial Instruments.
Asset Allocation with Credit Instruments
Table 7.3
3 Years b913ch07 FA
Mean Median σ CVaR(1%) Min Max Skewness Excess kurtosis
159
May 12, 2010
160
17:47
WSPC/SPIB913
b913ch07
FA
Menzinger et al.
lower return. This phenomenon can be easily explained by the nature of the instruments. Fixedincome instruments are exposed to market risk which is the higher, the longer the time to maturity. CDS and the CDX are very similar to an FRN and only have a very small market risk. They pay the shortterm interest rate, which is lower than the mediumterm interest rates if the curve is normal. In addition, the outcome of interest rates is more volatile over a longer investment horizon. In Table 7.4, we show the linear correlations between the returns of the financial instruments for the investment horizons one year and three years. We make the following main observations: Governmentbond and corporatebond returns have a high positive correlation. This is reasonable as the main driver of the bond return is the government rate, which is the same for all bonds. Differences in bond returns come from different credit spreads and different default times. Bond and equityindex returns are negatively correlated, which is also known from historical time series. There is a small positive or negative correlation between returns of corporate bonds and funded CDS/CDX. This may seem counterintuitive at first sight. A closer look at the nature of these instruments proves otherwise. We need to differentiate between two opposing effects, which can be attributed to single return components of the funded CDS/CDX. On the one hand, the present value of the CDS PVCDS (see Eq. (7.13)) strongly behaves like the price of the corporate bond. Zero rates being equal, PVCDS and the price of the corporate bond decrease if spreads increase resulting in a high positive correlation. On the other hand, bonds are fixed income instruments with prices that strongly depend on the level and the structure of zero rates. If zero rates are falling, the bond returns increase as bond prices increase. For the FRN linked to the CDS/CDX, it is the other way round. Since its return strongly depends on the floating LIBOR, the return of the FRN tends to decrease with falling interest rates resulting in a high negative correlation between bond returns and returns of funded CDS/CDX. All in all, the two opposing effects cancel each other out to a certain degree resulting in a small positive or negative correlation between corporate bonds and funded CDS/CDX. The returns of the equity index and CDS or the CDX are positively correlated. This is reasonable since both equityindex returns and the threemonth government rate are to a large proportion driven by the short rate, which was empirically confirmed by [5]. To conclude, we identify appealing risk–return profiles and low correlations between bonds, the equity index and CDS/CDX. Therefore, investors should benefit from holding a portfolio consisting of traditional financial instruments and credit instruments. After having analyzed the return characteristics, we can turn to portfolio optimization. For this purpose, we consider a oneyear and a threeyear investment horizon and the following three investment universes: Initial investment universe consisting of government bonds and an equity index; extension by corporate bonds; extension by CDS and the CDX.
Bond AA
Bond A
Bond BBB
CDS index
CDS AA
CDS A
CDS BBB
1.0000 −0.1184 0.9508 0.8824 0.6828 −0.2338 −0.1569 −0.1013 −0.0510
−0.1184 1.0000 −0.1182 −0.1110 −0.0835 0.1124 0.0751 0.0639 0.0104
0.9508 −0.1182 1.0000 0.8391 0.6511 −0.2227 −0.0781 −0.0986 −0.0465
0.8824 −0.1110 0.8391 1.0000 0.6002 −0.1934 −0.1385 −0.0417 −0.0482
0.6828 −0.0835 0.6511 0.6002 1.0000 −0.0104 −0.1061 0.0265 0.0560
−0.2338 0.1124 −0.2227 −0.1934 −0.0104 1.0000 0.2652 0.2802 0.1830
−0.1569 0.0751 −0.0781 −0.1385 −0.1061 0.2652 1.0000 0.1107 0.0632
−0.1013 0.0639 −0.0986 −0.0417 0.0265 0.2802 0.1107 1.0000 0.0395
−0.0510 0.0104 −0.0465 −0.0482 0.0560 0.1830 0.0632 0.0395 1.0000
1.0000 −0.1629 0.7001 0.5086 0.3271 −0.4689 −0.3812 −0.2919 −0.1814
−0.1629 1.0000 −0.1068 −0.1014 −0.0593 0.2651 0.2201 0.1764 0.1045
0.7001 −0.1068 1.0000 0.3543 0.2511 −0.2249 −0.2387 −0.1934 −0.0579
0.5086 −0.1014 0.3543 1.0000 0.2027 −0.1772 −0.1930 −0.0712 −0.0726
0.3271 −0.0593 0.2511 0.2027 1.0000 −0.0816 −0.1103 −0.0901 −0.0209
−0.4689 0.2651 −0.2249 −0.1772 −0.0816 1.0000 0.6809 0.5481 0.4063
−0.3812 0.2201 −0.2387 −0.1930 −0.1103 0.6809 1.0000 0.4113 0.2732
−0.2919 0.1764 −0.1934 −0.0712 −0.0901 0.5481 0.4113 1.0000 0.2070
−0.1814 0.1045 −0.0579 −0.0726 −0.0209 0.4063 0.2732 0.2070 1.0000
WSPC/SPIB913
Equity index
17:47
Govt. bond
May 12, 2010
1 Year Govt. bond Equity index Bond AA Bond A Bond BBB CDS index CDS AA CDS A CDS BBB
Correlation of Total Returns of Financial Instruments.
Asset Allocation with Credit Instruments
Table 7.4
3 Years b913ch07 FA
Govt. bond Equity index Bond AA Bond A Bond BBB CDS index CDS AA CDS A CDS BBB
161
May 12, 2010
17:47
WSPC/SPIB913
162
b913ch07
FA
Menzinger et al.
7.5.1. MeanVariance Approach With a meanvariance optimization according to the optimization problem in Eq. (7.14) for the relevant investment universes as described above, we obtain efficient frontiers and the corresponding asset allocations. The results are displayed in Fig. 7.1 for the oneyear investment horizon and in Fig. 7.2 for an investment horizon of three years. The upper part of the figures show the efficient frontiers, the lower parts of the figures contain the optimal asset allocations along the efficient frontiers. For the sake of readability, the results of the asset’s weights in the portfolio are partly aggregated, i.e., the element “Bonds AA/A/BBB” represents the corporate bonds with a rating of AA, A, or BBB and is the sum of all corporate bond weights in a certain optimal asset allocation, “CDS AA/A/BBB” represents all singlename CDS rated AA, A, or BBB in a certain allocation. The figures exhibit some similar structures. At first, we examine the efficient frontiers. Allowing for corporate bonds in the portfolio optimization leads to an upward shift of the efficient frontiers compared to the initial investment universe. For the same level of risk, a higher expected return can be generated. Allowing additionally for CDS and the CDX leads to a shift of the efficient frontier to the left, i.e., an investor can reduce the portfolio risk. The proportion of the potential risk reduction or return enhancement depends on the investment horizon. In Table 7.5, we show the potential improvement of the investor’s portfolio risk and/or return position exemplarily for a oneyear horizon using the minimumvariance portfolio of the initial investment universe as a reference point. Table 7.5 shows that the minimum variance portfolio based on the initial investment universe has an expected return of 5.76% and a standard deviation of 3.04%. This portfolio is compared to equivalent portfolios on the most extended investment universe, i.e., allowing not only for government bonds and an equity index, but also for all relevant credit instruments. The appropriate portfolios can be identified by either holding the expected return constant and looking for the standard deviation of the equivalent portfolio on the efficient frontier of the most extended investment universe, or holding the standard deviation constant and looking for the expected return on the efficient frontier. In the first case, the portfolio’s standard deviation is only 0.70%, in the second case the expected return is 6.72%. The results from the Table 7.5 allow the conclusion that a meanvariance investor can reduce his portfolio’s risk and/or enhance the return by adding credit instruments to the portfolio of government bonds and an equity index. Similar improvements of the risk and/or return position can be observed for the threeyear investment horizon. Now, we focus on the meanvariance optimal asset allocations, i.e., the lower parts of Figs. 7.1 and 7.2. In the initial investment universe, optimal allocations only consist of government bonds and an equity index. In the largest investment universe,
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
FA
163
Asset Allocation with Credit Instruments 11
Expected return (%)
10 9 8 7
6 5 4 0
5
10
15
20
25
Standard deviation (%) Equity Index and Government Bonds Equity Index, Government Bonds, Corporate Bonds
18.17
16.99
15.83
14.67
13.52
11.26
12.38
9.09
10.16
8.05
7.03
6.04
5.08
4.17
3.30
2.48
1.75
1.14
0.76
100 90 80 70 60 50 40 30 20 10 0
0.61
Portfolio weights (%)
Equity Index, Government Bonds, Corporate Bondsand CDS/CDX
Standard deviation (%) Equity Index
Government Bonds
CDS AA/A/BBB
Bonds AA/A/BBB
CDX
Figure 7.1 Results of meanvariance optimization, oneyear investment horizon.
government bonds are substituted to a large extent with corporate bonds, CDS, and the CDX. As already seen from Table 7.3, AA and Arated CDS and the CDX have a similar expected return compared to government bonds but a considerably lower standard deviation for a oneyear horizon. Therefore, particularly for lower levels of standard deviation, optimal portfolios contain a considerable proportion of CDS and the CDX. For levels of standard deviation higher than 7.64% (oneyear horizon) and 10.38% (threeyear horizon), an optimal portfolio only consists of corporate bonds and an equity index. For lower levels of standard deviation, corporate bonds are partially replaced by CDS and the CDX. There is only a rather small proportion of government bonds for a low level of risk, due to the risk–return profile of government bonds and CDS.
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
164
FA
Menzinger et al. 35 33 Expected return (%)
31 29 27 25 23 21 19 17 15 0
10
20
30
40
50
60
Standard deviation (%) Equity Index and Government Bonds Equity Index, Government Bonds, Corporate Bonds Equity Index, Government Bonds, Corporate Bonds and CDS/CDX
49.74
46.62
43.50
40.40
37.31
34.23
31.17
28.13
25.13
22.17
19.26
16.38
13.55
8.21
10.81
5.86
3.86
2.49
1.84
1.50
Portfolio weights (%)
100 90 80 70 60 50 40 30 20 10 0
Standard deviation (%) Equity Index
Government Bonds
CDS AA/A/BBB
Bonds AA/A/BBB
CDX
Figure 7.2 Results of meanvariance optimization, threeyear investment horizon.
7.5.2. Conditional Value at Risk In the CVaR optimization, we assume α to be 1%, i.e., we consider the mean of the worst 1% of the portfolio return as risk measure. The resulting efficient frontiers and the optimal asset allocations are displayed in Fig. 7.3 for the oneyear investment horizon and in Fig. 7.4 for the threeyear investment horizon. When comparing the results of the CVaR optimization with those of the meanvariance optimization, the first impression is very similar. In the following, we describe the main results of
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
FA
165
Asset Allocation with Credit Instruments Table 7.5 Comparison of Selected MeanVariance Optimal Portfolios Based on Two Different Investment Universes, OneYear Investment Horizon. Investment universe Initial Most extended Most extended
Portfolio selection criterion
µ (%)
σ (%)
Minimum variance portfolio Equal portfolio return Equal portfolio standard deviation
5.76 5.76 6.72
3.04 0.70 3.04
the CVaR optimization at first. Then, we describe and explain some differences of meanvariance and CVaR optimization. We begin with the efficient frontiers. Adding credit instruments to a portfolio only consisting of the initial investment universe leads to an upward shift and a shift to the left of the efficient frontier. This indicates that there is an enormous potential for return enhancement or risk reduction. The potential depends on the investment horizon. It is exemplarily illustrated for the minimumCVaR portfolio as a reference point, in Table 7.6. Table 7.6 reveals that for a oneyear investment horizon there is a high potential to either reduce the risk for a given level of return (from a CVaR of 2.10% to a CVaR of only −3.12%), or to enhance the portfolio return for a given level of risk (from 5.79% to 6.78%). Also for a threeyear investment horizon there is a high potential to reduce risk and/or enhance return. Next, we analyze the optimal asset allocations for an investor using the CVaRcriterion. In an optimal portfolio, government bonds are partially substituted with corporate bonds, CDS and the CDX compared to the initial investment universe. As explained earlier, AA and Arated CDS and the CDX have a similar expected return to government bonds, but a lower CVaR for a oneyear horizon. Particularly for low levels of CVaR, corporate bonds are partially substituted with CDS and the CDX. For levels of CVaR higher than 18.29% (oneyear horizon) and 14.47% (threeyear horizon), an optimal portfolio only consists of corporate bonds and an equity index. If, however, lower levels of CVaR are of interest, optimal allocations are composed of a considerable proportion of credit derivatives. Having a closer look at the resulting optimal allocations of meanvariance and CVaR optimizations, we can identify some differences. For a oneyear investment horizon, optimal allocations contain CDS even for high levels of expected returns. While meanvariance optimal portfolios are composed of CDS up to an expected portfolio return of 7.91%, CVaR optimal portfolios hold CDS up to a return level of 8.94%. In contrast, CVaR optimal portfolios contain significantly less BBBrated, but significantly more AArated corporate bonds. For a threeyear investment horizon, CVaR optimal portfolios hold CDS up to an expected portfolio return of 25.30%, while meanvariance optimal portfolios only consist of CDS up to returns of 21.53%.
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
166
FA
Menzinger et al. 11
Expected return (%)
10 9 8 7
6 5 4 −10
−5
0
5
10
15
20
25
30
35
CVaR (%) Equity Index and Government Bonds Equity Index, Government Bonds, Corporate Bonds
31.36
28.97
26.64
24.32
22.03
19.77
17.62
15.58
13.55
9.62
11.57
7.74
5.92
4.27
2.77
1.27
−0.20
−1.60
−2.78
100 90 80 70 60 50 40 30 20 10 0
−3.30
Portfolio weights (%)
Equity Index, Government Bonds, Corporate Bonds and CDS/CDX
CVaR (%) Equity Index
Government Bonds
CDS AA/A/BBB
Bonds AA/A/BBB
CDX
Figure 7.3 Results of CVaR optimization, oneyear investment horizon.
Furthermore, the former are composed of a significant larger proportion of government bonds than the latter. Again, we observe a considerably lower proportion of BBBrated corporate bonds for the CVaR optimal portfolios. BBBrated bonds provide a rather high expected return. Though, the standard deviation is not able to adequately capture the tail events. From these observations, we can conclude that the meanvariance optimization underestimates the benefits of credit derivatives, particularly for longer investment horizons, by only considering the first two moments of the return distribution, and not its tail. Moreover, it overestimates the benefits of BBBrated corporate bonds by only looking at the expected return and volatility but mainly ignoring the tail events. Our analysis show that an investor, optimizing his portfolio
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
FA
167
Asset Allocation with Credit Instruments 35 33 31 Expected return (%)
29 27 25 23 21 19 17 15 −20
−10
0
10
20
30
40
50
60
CVaR (%) Equity Index and Government Bonds Equity Index, Government Bonds, Corporate Bonds
51.57
47.12
42.79
38.55
30.07
34.31
25.93
21.92
18.02
14.17
10.51
6.98
3.78
0.84
−1.94
−4.67
−7.18
−9.34
−11.02
100 90 80 70 60 50 40 30 20 10 0
−12.00
Portfolio weights (%)
Equity Index, Government Bonds, Corporate Bonds and CDS/CDX
CVaR (%) Equity Index
Government Bonds
CDSAA/A/BBB
Bonds AA/A/BBB
CDX
Figure 7.4 Results of CVaR optimization, threeyear investment horizon.
with either the meanvariance or the CVaR criterion, can add performance to his portfolio for a given level of risk, or he can reduce risk for a given target return level. This can be realized due to a low correlation between the different instruments and their attractive risk–return profile.
7.5.3. Comparison of Selected Optimal Portfolios We close our analysis of the simulation results with a comparison of selected optimal portfolios. We examine the optimal asset allocation for two representative investors which we denote by risk–averse and risk–affine. Note that the term risk–affine does
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
168
FA
Menzinger et al. Table 7.6 Comparison of Selected CVaROptimal Portfolios Based on Two Different Investment Universes, OneYear Investment Horizon. Investment universe Initial Most extended Most extended
Table 7.7 Investor.
Portfolio selection criterion
µ (%)
CVaR (%)
MinimumCVaR portfolio Equal portfolio return Equal portfolio standard deviation
5.79 5.79 6.78
2.10 −3.12 2.10
Benchmark Portfolios for a Risk–Averse and a Risk–Affine
Govt. bond Equity index
Risk–averse investor (%)
Risk–affine investor (%)
70.00 30.00
30.00 70.00
not mean that this investor is seeking risk. He is rather willing to bear more risk than the risk–averse investor in compensation for a higher risk premium. Their respective benchmark portfolio is defined on the initial investment universe and is exhibited in Table 7.7. The procedure for selecting the optimal portfolios to compare with the benchmark portfolios is the following. At first, we determine the risk and return characteristics of the benchmark portfolios, denoted as “riskB ” and “returnB ”, where “riskB ” refers to the relevant risk measure: If optimal portfolios are selected using meanvariance approach (MV), the relevant risk measure is standard deviation, in the case of CVaR optimization it is the CVaR. We use these risk values as reference points to find the optimal portfolios with identical risk values on the efficient frontiers of the most extended investment universe based on the results of meanvariance and the CVaRoptimization presented in the previous sections (see Figs. 7.1–7.4). The results are shown in Tables 7.8 and 7.9 for the oneyear and the threeyear investment horizon, respectively. In the upper parts of the Tables 7.8 and 7.9 the risk–return profiles of the benchmark portfolios and the optimal portfolios (denoted as “risk∗ ”, “return∗ ”) can be found. In addition, the corresponding optimal asset allocations are displayed in the lower part of these tables. Tables 7.8 and 7.9 reveal some interesting characteristics of meanvariance optimal and CVaR optimal portfolios for the risk–averse and risk–affine investor defined in Table 7.7. At first, we focus on some similarities. Meanvariance and CVaR optimization lead to similar expected portfolio returns. Extending the investment universe by credit instruments an investor always can enhance the portfolio return. A risk– averse investor with a oneyear investment horizon can increase his portfolio return
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
FA
169
Asset Allocation with Credit Instruments Table 7.8 Comparison of Benchmark Portfolios and Optimal Portfolios for Risk–Averse and Risk–Affine Investor, OneYear Investment Horizon.
RiskB ReturnB Risk∗ Return∗ Govt. bond Equity index Bond AA Bond A Bond BBB CDS index CDS
Risk–averse investor
Risk–affine investor
MV (%)
CVaR (%)
MV (%)
CVaR (%)
5.95 7.01 5.95 7.50
6.52 7.01 6.52 7.47
13.43 8.97 13.43 9.21
20.52 8.97 20.52 9.20
0.00 29.50 0.00 20.38 37.94 0.00 12.18
0.00 31.01 24.05 27.10 6.46 0.00 11.39
0.00 69.82 0.00 0.00 30.18 0.00 0.00
0.00 69.93 0.00 7.16 22.90 0.00 0.00
Table 7.9 Comparison of Benchmark Portfolios and Optimal Portfolios for Risk–Averse and Risk–Affine Investor, ThreeYear Investment Horizon. Riskaverse investor
Riskaffine investor
MV (%)
CVaR (%)
MV (%)
CVaR (%)
RiskB ReturnB Risk∗ Return∗
15.63 21.63 15.63 23.13
4.52 21.63 4.52 22.88
36.80 28.38 36.80 29.12
32.76 28.38 32.76 29.18
Govt. bond Equity index Bond AA Bond A Bond BBB CDS Index CDS
0.00 29.04 0.00 23.32 47.64 0.00 0.00
0.00 30.14 24.80 21.86 11.20 0.00 12.00
0.00 69.89 0.00 0.00 30.11 0.00 0.00
0.00 70.89 0.00 15.86 13.25 0.00 0.00
from 7.01% to 7.50% applying the meanvariance optimization and to 7.47% applying the CVaR optimization. A risk–affine investor with a threeyear investment horizon increases his portfolio return from 28.38% to 29.12% (MV) or to 29.18% (CVaR). The proportion of the equity index in an optimal portfolio is always rather close to the
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
170
FA
Menzinger et al.
proportion of the benchmark portfolio, i.e., it is approximately 30% for the risk–averse investor and 70% for the risk–affine investor. However, the equity proportion is always slightly higher for investors using the CVaR criterion. The two investors should never allocate funds in government bonds to build an optimal portfolio. The proportion of the government bonds in the benchmark portfolios is rather reallocated to the credit instruments in the optimized portfolios, particularly to corporate bonds and CDS. For the considered levels of risk, the CDS Index does not play a role. For the oneyear horizon both optimization approaches lead to a proportion of approximately 11–12% of CDS for the risk–averse investor. Independent of optimization criterion or investment horizon, the risk–affine investor should not invest in CDS. Having a closer look at the proportions of the credit instruments, we can figure out some interesting differences. A risk–averse investor applying the CVaR criterion and having threeyear investment horizon should still allocate approximately 12% to CDS, while this investor using the meanvariance criterion should not hold CDS in his portfolio. This indicates that meanvariance optimization underestimates the benefit of credit derivatives for longer investment horizons. For both investment horizons, meanvariance optimal portfolios contain a considerably higher proportion of BBBrated corporate bonds than the CVaRoptimal portfolios. This means that meanvariance optimization overestimates the benefits of BBBrated corporate bonds by only taking the expected return and volatility into account but mainly ignoring the tail events. These effects were already explained in the previous section. To sum up, comparing the resulting optimal asset allocations for a representative risk–averse and risk–affine investor, we see that, independent of investment horizon and optimization criterion, an investor always benefits from substituting government bonds by corporate bonds, CDS, and the CDX. The resulting optimal allocations, however, strongly depend on the investor type, the optimization criterion and the investment horizon.
7.6. SUMMARY AND CONCLUSION Constructing portfolios with credit instruments requires an appropriate asset allocation framework in order to account for the distinct return characteristics of these instruments, such as nonnormality of the return distribution due to potential defaults. We have presented a consistent, scenariobased asset allocation framework, which is able to determine optimal portfolios consisting of traditional instruments and credit instruments. The entire framework is composed of a simulation, a total return calculation, and an optimization framework. We have suggested a model to simulate consistent capitalmarket scenarios offering analytical formulas for the whole term structure of interest rates and credit spreads. Furthermore, we have introduced a
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
Asset Allocation with Credit Instruments
FA
171
model to simulate correlated default times with an NIG onefactor copula, which is an extension of the Gaussian onefactor copula. The NIG version is appealing since it allows to generate more realistic properties of default times such as a higher probability of joint defaults of different issuers. After the introduction of general conditions for the total return calculations, we have explained how to price CDS. For portfolio optimization we have applied two criteria — the meanvariance and the CVaR optimization. The former can be viewed as traditional portfolio optimization. Its main disadvantage is that it only takes the first two moments of a return distribution into account, which is obviously not appropriate when also allowing for credit instruments. The latter is able to overcome this drawback since it considers the left tail of a distribution. So, it particularly takes defaults into account. Finally, we have presented the model parameters and we have applied our asset allocation framework to the U.S. market for a oneyear and a threeyear investment horizon. We have found that credit instruments have an appealing risk–return profile and a correlation structure providing a considerable potential for diversification. Moreover, we have found that meanvariance optimization overestimates the benefits of lowrated bonds by only focussing on mean and variance, and mainly ignoring the tail events. Realistic modeling of return characteristics is very important when instruments with specific return characteristics are included in an asset allocation. Otherwise, optimization results are not more than an approximation. Our model provides realistic return distributions. Furthermore, it can be fitted to market data and it can be easily extended by other credit instruments. Therefore, it is appropriate for an asset allocation with credit instruments.
APPENDIX The zero rates R(t, T ) and Rd (t, T ) at time t to maturity T of nondefaultable and defaultable bonds as well as the credit spreads S(t, T ) are given by
R(t, T ) = − Rd (t, T ) = − S(t, T ) = −
T T T
1 [A(t, T ) − B(t, T )rt − D(t, T )ωt ], −t 1 [Ad (t, T ) − B(t, T )rt − Dd (t, T )ωt − Ed (t, T )ut − F d (t, T )st ], −t 1 [Ad (t, T ) − A(t, T ) − (Dd (t, T ) − D(t, T ))ωt −t
− Ed (t, T )ut − F d (t, T )st ],
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
172
FA
Menzinger et al.
with
A(t, T ) =
T
T
t
Ad (t, T ) =
1 2
1 2 2 2 2 (σ B(l, T ) + σω D(l, T ) ) − θr (l)B(l, T ) − θω D(l, T ) dl, 2 r [σr2 B(l, T )2 + σω2 Dd (l, T )2 + σu2 Ed (l, T )2 + σs2 F d (l, T )2 ]dl
t
T
−
[θr (l)B(l, T ) + θω Dd (l, T ) + θu Ed (l, T ) + θs F d (l, T )]dl,
t
B(t, T ) =
1 (1 − e−ˆar (T −t) ), aˆ r
br D(t, T ) = aˆ r
1 − e−ˆaω (T −t) e−ˆaω (T −t) − e−ˆar (T −t) + aˆ ω aˆ ω − aˆ r
bsω D (t, T ) = D(t, T ) − aˆ s d
bsu E (t, T ) = aˆ s d
F d (t, T ) =
,
1 − e−ˆaω (T −t) e−ˆaω (T −t) − e−ˆas (T −t) + aˆ ω aˆ ω − aˆ s
1 − e−ˆau (T −t) e−ˆau (T −t) − e−ˆas (T −t) + aˆ u aˆ u − aˆ s
,
,
1 (1 − e−ˆas (T −t) ). aˆ s
For an explicit derivation of the zero rates, we refer to [4].
References [1] FitchRatings (2005). Global credit derivatives survey: Risk dispersion accelerates. [2] British Bankers’ Association (2006). BBA credit derivatives report 2006 — executive summary. [3] Bank for International Settlements (2009). BIS quarterly review June 2009: International banking and financial market developments. [4] Antes, S., M Ilg, B Schmid and R Zagst (2008). Empirical evaluation of hybrid defaultable bond pricing models. Journal Applied Mathematical Finance, 15(3), 219–249. [5] Zagst, R., T Meyer and H Hagedorn (2007). Integrated modelling of stock and bond markets. International Journal of Finance, 19(1), 4252–4277. [6] Kalemanova,A., B Schmid and R Werner (2007). The Normal Inverse Gaussian distribution for sythetic CDO pricing. Journal of Derivatives, Spring. [7] Markowitz, H. (1952). Portfolio selection. Journal of Finance, 7, 77–91. [8] Bielecki, T and M Rutkowski (2004). Credit Risk: Modeling, Valuation and Hedging.
May 12, 2010
17:47
WSPC/SPIB913
Asset Allocation with Credit Instruments
b913ch07
FA
173
[9] Hull, J and A White (1994). Numerical procedures for implementing term structure models ii: Twofactor models. Journal of Derivatives. [10] Nelson, CR and A Siegel (1987). Parsimonious modelling of yield curves. The Journal of Business, 60(4), 473–489. [11] Arin, KP and A Mamun (2005). Is it inflation or inflation variability? A note on the stock returninflation puzzle. Finance Letters, 3(6), 19–23. [12] Sellin, P (2001). Monetary policy and the stock market. Theory and empirical evidence. Journal of Economic Surveys, 15(4). [13] Feldstein, M (1980). Inflation and the stock market. American Economic Review, 70, 839–847. [14] Fama, EF (1981). Stock returns, real activity, inflation and money. American Economic Review, 71, 545–565. [15] Friedman, M (1977). Nobel lecture: Inflation and unemployment. Journal of Political Economy, 85, 451–472. [16] Li, DX (2000). On default correlation: A copula approach. Journal of Fixed Income, 9, 43–54. [17] Hull, J and A White (2004). Valuation of a CDO and an nth to default CDS without a Monte Carlo simulation. Journal of Derivatives. [18] Kalemanova, A and R Werner (2006). A short note on the efficient implementation of the Normal Inverse Gaussian distribution. Working Paper. [19] Vasicek, O (1987). Probability of loss on loan portflio. Memo, KMV Corporation. [20] Bluhm, C (2003). CDO modeling: Techniques, examples and applications. Working Paper. [21] Schmid, B., R Zagst and S Antes (2006). Pricing of credit derivatives. Working Paper. [22] Zagst, R (2006). Integrated risk management — distinguished lecure series 2006. Lecture notes. [23] Standard & Poor’s (2006) Annual 2005 global corporate default study and rating transitions.
May 12, 2010
17:47
WSPC/SPIB913
b913ch07
This page intentionally left blank
FA
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
8
CROSS ASSET PORTFOLIO DERIVATIVES
STEPHAN HÖCHT∗,‡ , MATTHIAS SCHERER∗,§ and PHILIP SEEGERER†,¶ ∗
HVBStiftungsinstitut für Finanzmathematik, Technische Universität München, Boltzmannstrasse 3, 85748 Garching bei München, Germany † Assenagon GmbH, Theresienhöhe 13 a, 80339 München, Germany ‡
[email protected] §
[email protected] ¶
[email protected] The dependence of extreme financial events among different asset classes is taken under consideration on a portfolio level. For this, a new product group, called cross asset portfolio derivatives, is introduced and explained in the light of related existing products and pricing methods. A classification is presented and features of these products are described. Finally, two modeling and pricing frameworks using multivariate stochastic processes and (hierarchical) copulas, respectively, are suggested.
8.1. INTRODUCTION TO CROSS ASSET PORTFOLIO DERIVATIVES This chapter focuses on the dependence of tail events, i.e., the risk of rare market events occurring interdependently. A current trend in product development is to apply credit 175
May 12, 2010
17:47
WSPC/SPIB913
176
b913ch08
FA
Höcht et al.
derivative techniques on stylized insurance claims based on various assets. Hence, we aim at extending the well established concepts for the pricing of credit derivatives, such as credit default swaps (CDS), nthtodefault baskets, and collateralized debt obligations (CDO), to other asset classes in the form of new investment products. Asset classes that might serve as underlyings are equity, commodity, interest rates, and foreign exchange. Thereby, we focus on the dependence of tail events among different asset classes in these products. The following definitions were adopted from the world of credit derivatives, see [1], page 8 ff, and generalized to trigger derivatives.
8.1.1. Definitions and Examples A trigger event causes a payment stream in some derivative. The default of a company (credit event) in the case of a CDS or the drop of a stock below a specified trigger level (equity event) in the case of an equity default swap (EDS) are examples. Very unlikely trigger events, such as a steep sudden drop of an equity index (equity tail event) or the sudden default of an AAArated company (credit tail event), are referred to as tail events. A trigger derivative is a bilateral derivative security, where the payoff profile depends on the occurrence of a trigger event. The trigger event is defined with respect to one (or several) reference underlying(s). After the trigger event has occurred, a contingent payment is due by one of the contractual parties. In case of an asset trigger derivative, the trigger event is defined with respect to a tradable reference underlying. Single name asset trigger derivatives (SNATD) stand for only one reference underlying in the contract and are represented by trigger swaps (TS ). These are agreements between two contractual parties to exchange the risk of a reference asset hitting a predefined trigger level. More precisely, one party receives periodic premium payments as long as the trigger event has not taken place. In return, the other party receives a contingent payment at the time of the trigger event. Trigger swaps might be interpreted as EDS (or CDS) extended to other asset classes; the difference of the first to the latter being that the underlying does not default on some equity event (or credit event), instead, the trigger event is defined on a tradable asset. In the context of options, TS might also be seen as digital onetouch options or barrier options, see [2], page 561. In contrast to CDS contracts, where the reimbursement rate (RIR), i.e., one minus the recovery rate, is usually not known prior to the credit event, in a trigger swap the reimbursement rate is fixed as part of the contractual specifications. A range from 50% to 100% seems to be realistic. Whereas SNATD relates to a single reference underlying, portfolio asset trigger derivatives (PATD) have payment streams that are linked to a portfolio of underlyings. At the next level, we distinguish between mono asset portfolio derivatives and cross asset portfolio derivatives. Mono asset portfolio derivatives (MAPD) are derivatives
May 12, 2010
17:47
WSPC/SPIB913
Cross Asset Portfolio Derivatives
b913ch08
FA
177
with a payoff, which is contingent on trigger events referring to an underlying pool of trigger swaps from the same asset class. Generalizing mono asset portfolio derivatives to multiple asset classes leads to cross asset portfolio derivatives (CAPD). Considering the specific payment schemes, we can distinguish between nthtotrigger baskets and collateralized trigger swap obligations (CTSO). The former are portfolio derivatives with a contingent payment depending on the nth trigger event in the underlying portfolio of trigger swaps, the latter are tranched securities which are backed by a pool of TS. In case of a CTSO, the underlying portfolio is composed of I trigger swaps, where I is typically in the range of 50–200. In contrast to synthetic CDO, where the underlying CDS are linked to different credits, two trigger swaps in a CTSO portfolio might refer to the same underlying asset, however, with different trigger events/trigger levels. The portfolio’s notional, i.e., the sum of the TS’s individual notionals, can be divided into different tranches. Similar to a TS providing exposure to the trigger event risk of the reference asset, a tranche of a CTSO provides exposure to the risk of a particular amount of loss on a portfolio of assets. An investor can decide how much trigger event exposure she is willing to take from the portfolio. This is achieved by choosing a specific tranche, i.e., by choosing a particular level of subordination. In the case where only one underlying asset class is considered, these products were already introduced to the market. Examples are collateralized commodity obligations (CCO) and collateralized foreign exchange obligations (CFXO). In the case of a CFXO, with foreign exchange rate TS serving as the underlying portfolio, several TS for each foreign exchange rate (with a cascade of trigger levels) are included in the underlying portfolio. The following example illustrates the general structure of a CTSO and the involved tranche structure. Example 1 (General structure of a CTSO). Consider a CTSO consisting of 100 TS with a notional of one million US$ each. Instead of taking a linear exposure in all TS simultaneously, which corresponds to selling all TS, an investor (insurance seller) decides to invest in the (10%, 20%) tranche of the CTSO with notional size 10 million. The tranche’s boundaries are called attachment and detachment points. The resulting payoff schedule is as follows: the investor receives a regular periodic spread on the remaining outstanding notional of the invested tranche. If one TS in the portfolio is triggered, the portfolio’s notional is reduced to 99 million (assuming a reimbursement rate of 100%) and the subordination of the investor’s tranche is reduced to (9%, 19%). If the 11th trigger in the portfolio occurs, the investor’s tranche is hit for the first time. The tranche’s remaining notional reduces to nine million and a contingent payment of one million has to be paid to the protection buyer. Since spread is paid pro rata to the remaining tranche, spread payments are reduced accordingly in the following. With the 20th trigger occurring in the portfolio, the whole tranche notional is eliminated and further spread payments are stopped. Since the first loss occurs with the 11th trigger
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
178
FA
Höcht et al.
and the total loss is realized with the 20th, the result is an optionlike payoff. This optiontype behavior can be used to introduce the leverage into the investment. Due to the socalled waterfall structure of the tranched CTSO, the investor can choose different subordination levels. This choice determines the real trigger event exposure to the portfolio. A less senior tranche bears a higher trigger risk, but bearing this risk is compensated with higher spread on the outstanding notional. In case of a CDO, one often classifies the subordination in equity, mezzanine, and senior tranches. These names intend to reflect the different risk exposures. Example 2 (General structure of an nthtotrigger basket). Consider a portfolio of 10 TS, with a notional of one million US$ each, and an investor (insurance buyer) who wants to protect herself against the first loss in the portfolio by means of a firsttotrigger swap. For this, she pays periodic premium payments on a prespecified schedule. As soon as one of the TS is triggered, premium payments stop and she receives a contingent payment in the amount of reimbursement rate times the notional. For n > 1, nthtotrigger derivatives are defined similarly, with the first trigger replaced by the nth trigger. Figure 8.1 shows the classification of cross asset portfolio derivatives within the product group of trigger derivatives. The specific products mentioned denote examples of the respective group.
Figure 8.1 The classification of trigger derivatives.
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
Cross Asset Portfolio Derivatives
FA
179
8.2. COLLATERALIZED OBLIGATIONS This section provides an economic classification of cross asset portfolio derivatives within the group of collateralized obligations, opposed to the structural classification in Sec. 8.1. We also stress the risks involved and compare those with existing related products in the context of collateralized obligations. We seek to introduce a simple classification, focusing on the major parameters rather than on pure theoretical constructs. We define collateralized obligations (CO) as structured and tranched investments on an underlying pool of assets. Similar to trigger derivatives we distinguish CO with respect to their underlying collateral. We call a structured and tranched investment on an underlying pool of credit assets (e.g., bonds, loans, CDS, and creditlinked notes) a collateralized debt obligation (CDO). Imposing CDO conventions on other asset classes than debt, e.g., equity, foreign exchange, commodity, and fixed income, results in a collateralized asset obligation (CAO). At the next level, we consider the source of funds for the interest paid to the investor of a CO. On the one hand, we have market value CO, where the underlying collateral is a marktomarket instrument. On the other hand, we have cash flow CO, where the underlying collateral yields a periodic cash flow. Funding is another criterion to classify CO and we distinguish between cash and synthetic funding. The former involves a portfolio of cash assets as underlying and the ownership of the assets is transferred to the legal entity (referred to as special purpose vehicle) issuing the CO’s tranches. The reference assets of a synthetic CO are not owned but swaplike. Using the above classification, a CTSO can alternatively be defined as a synthetic cash flow CAO. Besides CTSO, there exists another class of CAO referencing to other asset classes than credit, namely collateralized fund obligations (CFO). However, as opposed to CTSO, a CFO is a market value cash CAO. Figure 8.2 shows the classification of CO regarding the distinctions made in Sec. 8.2.
8.3. A COMPARISON OF CFO WITH CTSO At first glance, CFO and CTSO share many similarities, but the structural features as well as the involved risks differ quite significantly. The following structural features are presented in [3].
8.3.1. Structural Features of CFO CFO have been known since the beginning of the 90s, referring to underlying indices on a price level, e.g., commodity indices, hedge fund indices, or equity indices. We assume CFO to fit especially the needs of realmoney accounts that are interested in
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
180
FA
Höcht et al.
Figure 8.2 The classification of CO.
new investment technologies based on asset classes, they are familiar with. Traditional CFO, using equity and commodity indices as underlying, did not become as popular as their counterparts: cash flow CO. In 2000–2001 CFO experienced a strong growth, with hedge fund investments being included. The idea was to use a technology, which generates benefits from a combination of highly volatile and uncorrelated trading strategies (longshort equity, convertible arbitrage, eventdriven, and fixed income arbitrage). On the negative side, these CFO investments included multiple fees due to the investment in tradable funds. According to [4], securitized hedge fund investments can provide the following benefits: • Attractive financing terms • Transferability of existing portfolios of fund investments • Diversification of the investor base by providing fixed income investors with market exposure linked to hedge fund returns • New opportunities for investors seeking leveraged returns in this asset class In 2003–2004, according to [3], hedge fundlinked CFO activity slowed down due to accounting issues. Still, market interest remained. The only innovative feature in these transactions, compared to usual CO, was the use of an uncommon underlying.
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
Cross Asset Portfolio Derivatives
FA
181
8.3.2. Structural Features of CTSO From the perspective of portfolio diversification, CTSO offer the opportunity to tap the market for (correlated) trigger event risks. Tail events are the major risks in such structures, while structural features as well as the unfunded nature of such transactions are similar to the synthetic CDO market. Hence, this might be a favorable playing field for hedge funds, proprietary desks, and insurance companies, given the in general unfunded nature of the investments. The major characteristics of CTSO are the following: • • • •
Unfunded nature, trigger swaps as underlying Tail event risk as major risk factor Tail event correlation as major price input No waterfall principle necessary, due to the synthetic structure
8.3.3. The Different Risks In general, while CFO are funded investments, CTSO are unfunded, i.e., swaplike. Whereas the risk in a CFO transaction is basically market risk (volatility of the underlying), it is trigger event risk in case of a CTSO, due to the swap character of the underlying. Hence, a CTSO and a CFO, referring to the same underlyings, bear a different risk/return profile. One can construct a CFO which refers to specific marktomarket instruments such as commodity or equity indices. Using the corresponding swap contracts (with some defined trigger event) on these indices as an underlying in a CTSO, the risk profile of the transaction changes. While the CFO reacts highly sensitive to daily marktomarket changes in the underlying indices, a CTSO investment better withstands small shortterm price fluctuations, but is skewed to trigger events. To sum up, the major difference between a CFO and a CTSO is that volatility is the appropriate risk parameter in the former, and tail event risk (depending on the trigger event of the underlying swap) is suitable to describe the risk profile of the latter.
8.3.4. Correlation of Tail Events in CTSO Since the dependence of tail events is a major pricing input, one needs to stress specific correlation assumptions. For example, in CCO, one might argue that specific macroeconomic shocks lead to a trigger event. Tail event correlation in different hedge funds, for instance, is mostly linked to systematic, regulatory, and legal risks. Consequently, a crossasset CTSO offers the opportunity to benefit from (thus far) nontradable asset classes like intermarket tail event correlation. The traditional correlation pattern between asset classes is changing in a cross asset CTSO regime, with idiosyncratic risk factors losing in importance compared to
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
182
FA
Höcht et al.
systematic risk factors. The risk/return profile of CTSO tranches differs significantly from the one of the underlying assets, shifting risk away from volatility toward tail events. The impact of changes in correlation depends on the specific tranche as the following two examples illustrate: (1) A long position in an equity piece of a CCO, where the trigger levels are set below and above the current spot prices, is long in joint probabilities of large price swings in the commodity market, i.e., the buyer of that piece expects rising joint probabilities of large price movements. (2) A short position in a senior piece of a cross asset CTSO is long in joint probabilities of distortions in the underlying markets, i.e., the short seller expects the likelihood of a market crash to rise. Considering intermarket correlation is not new, but the securitization in terms of the aforementioned products is. However, in order to derive the correlation of trigger events in different financial markets, analyzing historical data is not sufficient. Mathematical models, such as structural and/or copulabased models known from credit markets, have to be implemented in order to get an intuition for the traded tail event correlations in CTSO.
8.4. PRICING CROSS ASSET PORTFOLIO DERIVATIVES 8.4.1. Pricing Trigger Swaps In the following, we introduce the basic notations of TS and state a pricing formula for the annualized fair spread sf , which is also known as par spread. This spread is computed such that both contractual parties can enter the TS at zero cost, i.e., without upfront payment, at initiation. • • • • • •
τ: the trigger time of the underlying N: the nominal of the TS RIR: the reimbursement rate of the TS s: the annualized spread of the TS T : the maturity of the contract rt : the deterministic termstructure of risk free interest rates
At initiation of the contract, a payment schedule T = {0 < t1 < · · · < tn = T } is specified, where t0 = 0 is the settlement date. Assuming, for notational simplicity, the usual (for credit derivatives) quarteryearly premium payments, this implies tk = k/4 for k = 0, . . . , 4T and tk = tk − tk−1 = 1/4. For a given pricing measure Q and the respective risk free interest rates rtk , continuously compounded, the value of the
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
183
Cross Asset Portfolio Derivatives
expected discounted premium leg (EDPL) of the TS is obtained via: EDPL = EQ
n
N · s · tk · e−rtk tk 11{τ>tk }
k=1
N · s −rt tk e k Q(τ > tk ). 4 k=1 n
=
(8.1)
The expected discounted trigger leg (EDTL) of the TS is given by EDTL = EQ RIR · N
n
e−rtk tk 11{tk−1 tk ). 4 n
e−rk tk Q(tk−1 < τ ≤ tk ) −
k=1
k=1
The par spread sf , computed to allow both parties to enter the TS at zero cost, is found by solving TS(0, T) = 0 for s, i.e., sf =
RIR
n −rtk tk Q(tk−1 < τ ≤ tk ) k=1 e . n 1 −rtk tk Q(τ > tk ) k=1 e 4
(8.3)
Equations (8.1) and (8.3) do not account for accrued interest, which might additionally be included as part of the terms of contract. If accrued interest is stipulated, an additional premium payment is added to the premium leg, which is proportional to the time between the trigger event τ and the last payment date tk−1 and is conditional on the event that the trigger falls into the respective period (tk−1 , tk ].
8.4.2. Pricing nthtoTrigger Baskets We consider a basket of I underlying TS and assume a pricing measure Q to be given. If not stated differently, we assume the same reimbursement rate RIR for all underlying TS. The order statistic of the stochastically dependent trigger times τ1 , . . . , τI is denoted by τ (1) ≤ · · · ≤ τ (n) ≤ · · · ≤ τ (I) . The expected discounted trigger and
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
184
FA
Höcht et al.
premium legs of an nthto trigger contract are given by n N · s(n) · tk · e−rtk tk 11{τ (n) >tk } EDPL(n) = EQ k=1
=
(n)
EDTL
N ·s 4
(n)
n
e−rtk tk Q(τ (n) > tk ),
k=1
= EQ RIR · N
n
(8.4)
−rtk tk
e
11{tk−1 tk ) k=1 e 4
8.4.3. Pricing CTSO In what follows, we introduce a mathematical notation of the terms verbally explained earlier in Example 1. Assume the CTSO portfolio consists of I TS contracts, indexed by i. This portfolio is segmented in J tranches, indexed by j. The conventions regarding the payment schedule are similar to the ones in the previous sections. Moreover, we need • • • • • • • • • •
TSi : a TS relating to underlying i τi : the trigger time of underlying i Ni : the nominal value of TSi RIRi : the reimbursement rate of underlying i lj , uj : the attachment and detachment point of tranche j M: the total nominal value of the CTSO j Mt : the remaining nominal of tranche j at time t Lt : the cumulated loss of the CTSO portfolio up to time t j Lt : the cumulated loss in tranche j up to time t j s : the annualized spread of tranche j
For notational simplicity, the structure is simplified by assuming a homogeneous portfolio with respect to the reimbursement rate and nominal of each TS, i.e., RIR = RIRi and N = Ni for all i ∈ {1, . . . , I}. The crucial quantity for the description of all
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
185
Cross Asset Portfolio Derivatives
payment streams is the portfolioloss process Lt , which is defined as Lt = RIR · N
I
t ∈ [0, T ].
11{τi ≤t} ,
(8.7)
i=1
The total nominal value of the CTSO is given by M = I · N. This nominal value is segmented using the attachment and detachment point of each tranche. More precisely, we define the partition 0 = l1 < u1 = l2 < · · · < uJ−1 = lJ < uJ = M. Based on the overall portfolio loss, the loss affecting tranche j results in j
Lt = min{max{0, Lt − lj }, uj − lj },
t ∈ [0, T ],
j ∈ {1, . . . , J}.
(8.8)
The remaining nominal of tranche j is determined by j
j
Mt = (uj − lj − Lt ),
t ∈ [0, T ],
j ∈ {1, . . . , J}.
(8.9)
Given the payment schedule T = {0 = t0 < t1 < · · · < tn = T } and the pricing measure Q, the expected discounted premium and trigger legs of tranche j are given by n j (j) (j) −rtk tk s · tk · e Mtk , j ∈ {1, . . . , J}, EDPL = EQ (8.10) (j)
EDTL
= EQ
k=1 n
−rtk tk
e
j (Ltk
−
j Ltk−1 )
,
j ∈ {1, . . . , J}.
(8.11)
k=1
The fair spread of tranche j of the CTSO results in j j n −rtk tk e (L − L ) EQ t t k=1 k k−1 (j) , sf = j n −rtk tk EQ Mtk k=1 tk · e
j ∈ {1, . . . , J}.
(8.12)
8.4.4. Modeling Approaches For the pricing of derivatives on a portfolio (or basket) of TS, the joint distribution of all trigger times (τ1 , . . . , τI ) is required. On a conceptual level, there are two methods to incorporate dependence among the trigger times that seem promising for an adaption from the world of credit derivatives to the current situation. In what follows, both methodologies are introduced and discussed.
8.4.4.1. The structural approach Dependence is introduced to the model by specifying a suitable Idimensional stochastic process. The univariate marginals either directly represent the respective underlying in the case of a traded asset, e.g., a stock index, or might be interpreted as
May 12, 2010
17:47
WSPC/SPIB913
186
b913ch08
FA
Höcht et al.
a proxy variable. An example for the latter is the firmvalue interpretation used in structuraldefault models. Typically, the trigger time τi is then defined as some firstpassage time of the respective univariate marginal below or above some trigger level. Depending on the choice of univariate model, one has different options to couple the marginals. For instance, if the univariate marginals are driven by Brownian motions, a natural approach is to assume these Brownian components as being correlated. If more general Lévy processes are used, common jumps might also be used to introduce dependence. References for the modeling of stocks and credit products are [2] and [1]. References for the modeling of interest rates and energy, as examples for other underlyings than stocks and credit, are the books [5] and [6], respectively. The major advantage of the structural approach is that it is fully dynamic, since all relevant objects are modeled as marginals of one highdimensional stochastic process. Besides the static pricing of all univariate and portfolio derivatives at a time, this additionally induces a dynamic which can be useful for the pricing of options on these derivatives. However, the structural approach has some drawbacks. Firstly, it is not obvious how dependence shall be introduced to univariate processes from different model classes, e.g., classical diffusions and pure jump processes. Secondly, analytically solving a nontrivial multidimensional firstpassage time problem is virtually impossible. This becomes evident when the example of [7] is considered, who succeeds in presenting a (highly nontrivial) formula for the joint probability of two correlated diffusions to remain above some threshold level for a given time. Hence, pricing portfolio trigger swaps requires a Monte Carlo engine, which is slow and biased if correlated processes are sampled on a discrete grid and monitored for some threshold level. Thirdly, for dependent stochastic processes, it is typically not possible to separate the parameters of the marginals from those specifying the dependence structure; an exception are processes purely driven by correlated Brownian motions. This complicates the calibration of the model. Finally, the dependence structure of the resulting vector (τ1 , . . . , τI ) is unknown for all nontrivial examples.
8.4.4.2. The copula approach Combine the marginal distributions of the univariate trigger times by means of a copula, i.e., by means of an Idimensional distribution function on the unit Icube. Since an introduction to copulas is beyond the scope of this chapter, the interested reader is referred to the textbooks [8]– [10]. The following approach was first used in the context of credit riskmodelling by [11] and [12]. For each trigger time, it is assumed that the (risk–neutral) marginal distribution is known and abbreviated as Q(τi ≤ t) = p¯ i (t) = 1 − pi (t). On a univariate level, each trigger time admits the
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
187
Cross Asset Portfolio Derivatives
canonical construction τi = inf{t ≥ 0 : pi (t) ≤ Ui },
Ui ∼ Uniform(0, 1).
(8.13)
This representation is especially convenient for simulations, since sampling τi boils down to sampling the univariate trigger variable Ui and solving, assuming a continuous and strictly decreasing pi , pi (t) = Ui for t. Considering all random times τ1 , . . . , τt at once it is possible to introduce dependence to the model by assuming the vector of trigger times (U1 , . . . , UI ) to be distributed according to some Idimensional copula C. Note that, by the defining properties of a copula, it is guaranteed that the univariate marginal distributions (of all τi ) are preserved, which is highly convenient in a calibration. If sampling strategies for C are known, it is straightforward to sample from the vector of random times (τ1 , . . . , τI ): sample (U1 , . . . , UI ) ∼ C and locate each univariate trigger time τi . The major advantage of the copula approach is the separation of marginals from the dependence structure. Firstly, this allows the use of different model classes for the univariate marginals. For instance, one might use a classical geometric Brownian motion for one asset class, a purejump Lévy process for a second asset class, and an intensity model for the asset class credit. For each of these classes, one can derive univariate distributions for the respective τi ’s, which are then coupled by means of the copula C. Using copulas with taildependence and singularities allows to include Armageddon scenarios (with multiple trigger events at a time) and trigger clusters. Secondly, the approach is wellsuited for simulations, as long as sampling routines for the copula are known. Thirdly, the calibration of the multivariate model is simplified when the univariate marginals can be calibrated individually, followed by a calibration of the dependence structure in a second step. However, there are some shortfalls of the copula approach that need to be addressed. Firstly, the model is static in the sense that a multivariate distribution of (τ1 , . . . , τI ) is specified, but no dynamic model leading to it. While this is unproblematic for pure pricing problems, it rules out applications such as the simulation of the time evolution of resulting model spreads. Secondly, it is not clear which class of copulas is best suited for modeling the dependence structure of the respective market. As long as there is no liquid market for tail dependence, it is not clear which copula should be preferred and how this copula should be parametrized. Let us briefly give some references for the required sampling of copulas. In high dimensions, sampling strategies for some elliptical copulas (including the Gaussand tcopula) are known, see e.g., [10]. For exchangeable (meaning that all pairwise correlations are identical) Archimedean copulas the standard reference is [13], an important contribution is [14]. Recently, some authors considered hierarchical (nested) Archimedean copulas and sampling strategies for these, see e.g., [15]– [17]. Sampling Marshall–Olkin type copulas in high dimensions is considered in [18].
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
188
Höcht et al.
8.4.5. An Example for an nthto Trigger Basket In the following, we illustrate both pricing methodologies using an example for a trigger derivative of medium complexity. Note that this example is more general than cross asset trigger derivatives as introduced in Sec. 8.1, since it combines asset trigger events with credit events. Example 3 (An nthto trigger basket over two asset classes). In this example, an nthto trigger basket consisting of six underlying TS is considered. The payment frequency for premium payments is quarterly, the contract settles on 10 February 2009, with a maturity of T = 5 years. The reimbursement rate is set to RIR = 50% for all TS and for the CDS calibration. All TS contribute the same unit nominal to the portfolio. (1) Stock indices: • DAX 30, initial value S1,0 = 4636 • EuroStoxx 50, initial value S2,0 = 2338 • SMI, initial value S3,0 = 5219 An equity event is triggered when some index falls below a trigger level, which is defined as 20% of the initial value, i.e., li = 0.2 · Si,0 . (2) CDS: • 5 year CDS on Allianz SE, initial spread 77.50 bps • 5 year CDS on Linde AG, initial spread 90.00 bps • 5 year CDS on E.ON N, initial spread 63.33 bps Credit events are triggered according to the iTraxx CDS convention.
8.4.5.1. A pricing exercise of Example 3 (structural approach) We model the evolution of each stock index i = 1, 2, 3 using a standard geometric Brownian motion Si . Considering CDS spreads, we follow [19] and assume a simple structural default model with geometric Brownian motion as firmvalue process Vi and monitor continuously for default. Since all processes have the same structure and are driven by Brownian motions, it is natural to introduce dependence by assuming correlated Brownian components. Summarizing, we consider the vectorvalued process (S1,t , S2,t , S3,t , V1,t , V2,t , V3,t ), where under Q S ), dSi,t = Si,t (rdt + σS,i dWi,t
Si,0 > 0, i = 1, 2, 3,
(8.14)
dVi,t = Vi,t (rdt +
Vi,0 > 0, i = 1, 2, 3.
(8.15)
V σV,i dWi,t ),
All equity events are defined as τi = inf{t ≥ 0 : Si,t ≤ li }, where li = 0.2 · Si,0 for the respective contract. Credit events are taken as τi = inf{t ≥ 0 : Vi,t ≤ di } for
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
189
Cross Asset Portfolio Derivatives
some suitable default threshold di , i = 1, 2, 3. Solving the model via a Monte Carlo simulation requires the following steps. (1) Initialize the simulation, i.e., choose the number of simulation runs SR and the gridsize . (2) For each simulation run n = 1, . . . , SR do: (a) Simulate the vectorvalued process (S1,t , S2,t , S3,t , V1,t , V2,t , V3,t ) on the grid 0 < < 2 < · · · < T . (b) At each point of the grid, monitor for eventual trigger times below (or above, for other contract specifications) the equity or default thresholds, respectively. (3) Compute and store the discounted premium and trigger leg of the current simulation run n. (4) Estimate the expected discounted premium and trigger leg as the arithmetic mean of the legs in each run. (5) Set the (estimated) fair spread as the quotient of the estimated trigger and premium leg. Note that such a continuous model might underestimate the risk of extreme (joint) movements. This fact is well known in the univariate case, see e.g., Chapter 1 of [20], and is further amplified for joint movements, due to a lack of tail dependence of the multivariate normal distribution. To overcome this, one might try to use jump processes instead of geometric Brownian motions.
8.4.5.2. A pricing exercise of Example 3 (copula approach) We again model the evolution of each stock index using a standard geometric Brownian motion. Considering CDS spreads, we use a simple reducedform model with constant default intensity. Since dependence is introduced in a second step by means of a copula, we use independent processes of the form S dSi,t = Si,t (rdt + σS,i dWi,t ),
Si,0 > 0, i = 1, 2, 3,
(8.16)
to model the stocks under Q. All equity events are again defined as τi = inf{t ≥ 0 : Si,t ≤ li }. Credit events are defined as τi = inf{t ≥ 0 : exp (−λi t) ≤ Ui }, where U1 , U2 , U3 are dependent (univariate uniform) triggers and λ1 , λ2 , λ3 denote the (constant) default intensities. Solving the model via Monte Carlo requires the following steps. (1) Initialize the simulation, i.e., choose the number of simulation runs SR. (2) For each stock index, compute survival probabilities Q(τi ≥ t) from the parameters of the respective process. For this, the required probability of a Brownian motion
May 12, 2010
17:47
WSPC/SPIB913
190
b913ch08
FA
Höcht et al.
with drift not falling below some threshold level for a given amount of time is given in [21], page 61. (3) For each simulation run n = 1, . . . , SR do: (a) Simulate the dependent trigger variables (U1S , . . . , U1C , . . . ) ∼ C. (b) Compute the resulting trigger times. (4) Compute and store the discounted premium and trigger leg of the current simulation run n. (5) Estimate the expected discounted premium and trigger leg as the arithmetic mean of the legs in each run. (6) Set the (estimated) fair spread as the quotient of the estimated trigger and premium leg. More details on the copula approach in the context of credit–risk modeling, including asymptotic confidence intervals and the implementation of several nestedArchimedean copulas, are given in [22].
8.4.5.3. Resulting model spreads The input for our calibration are quotes obtained from Reuters on 10 February 2009 (2 p.m.). The continuously compounded interest rate is chosen as r = 0.024. The marginals of the stock indices and individual stocks are calibrated as follows: the parameter σS,i of stock index i is calibrated as implied volatility from outofthemoney put options. For this, a fiveyear maturity and strike close to li would be ideal. Since these were not liquidly traded, we settled with a one year maturity and a strike in the order of twothirds of the initial value. The results are (σS,1 , σS,2 , σS,3 ) = (38%, 38%, 31%) for the stock indices and (65%, 42%, 47%) for the individual stocks. Note that these outofthe money options typically trade at a higher implied volatility compared to atthemoney options. The resulting triggersurvival probabilities for the stock indices are (90%, 90%, 97%). Considering CDS spreads, the stock price process is taken as a reference for the firmvalue process and is calibrated in the very same way as the stock index. In a second step, the default threshold di is chosen such that the fiveyear CDS spread, being monotonically increasing in di , is matched. Note that CDS spreads are computed under the simplifying assumptions of no accrued interest and a deterministic recovery rate R = (1 − RIR) = 50%. Translated in terms of probabilities, this corresponds to a fiveyear implied survival probability of about (92%, 91%, 94%), respectively, for the three companies. For the copula approach, the marginal intensity λi for the ith CDS is similarly chosen such that the intensity model matches the observed fiveyear CDS spread. Since market prices of correlated tail risk are not (yet) observable, a calibration to existing derivatives is not possible. Hence, we computed model prices for a large
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
Cross Asset Portfolio Derivatives
FA
191
spectrum of dependence structures. For this, we assumed a homogeneous correlation ρ1 among each pair of stock processes and firmvalue processes, i.e., dWi,t dWj,t = ˆ j,t = ρ1 dt, i = j, and a homogeneous correlation ρ2 < ρ1 among stocks and ˆ i,t d W dW ˆ j,t = ρ2 dt, i ∈ {1, 2, 3}. For the copula framework, firmvalue processes, i.e., dWi,t d W we chose the Gauss copula with a blockmatrix as correlation structure such that again the groups stocks and credit are correlated with ρ1 , and stocks to credit are correlated with ρ2 < ρ1 . Moreover, we included a partially nested Archimedean copula of the Gumbel family, see [15] and [16]. We use two levels and three members in each of the two (homogeneous) groups, i.e., C(u1 , . . . , u6 ) = C(C(u1 , u2 , u3 ; ψ1 ), C(u4 , u5 , u6 ; ψ1 ); ψ2 ) = ψ2 (ψ2−1 (ψ1 (ψ1−1 (u1 ) + ψ1−1 (u2 ) + ψ1−1 (u3 ))) + ψ2−1 (ψ1 (ψ1−1 (u4 ) + ψ1−1 (u5 ) + ψ1−1 (u6 )))),
ui ∈ [0, 1], (8.17)
where ψi (t) = exp(−t 1/ϑi ) and ϑi ∈ [1, ∞), i = 1, 2. A sufficient condition for (8.17) being a copula, is that the nodes ψ2−1 ◦ ψ1 have completely monotone derivatives. For the nested Gumbel copula, this condition is equivalent to ϑ2 ≤ ϑ1 . NestedArchimedean copulas might be interpreted as follows: any two members of the groups stocks and credit are coupled via an inner Archimedean copula, and any two members of different groups are coupled via an outer Archimedean copula. Nested Archimedean copulas are appealing for the modeling of cross asset portfolio derivatives, since members of the same asset class can be coupled via inner copulas, which are then connected via one outer copula. For simplicity, we choose a single parameter ϑ1 for both groups (stocks and credit) and a second parameter ϑ2 ≤ ϑ1 for the outer copula. This implies that the dependence within a group is at least as large as the dependence among members of different groups. The resulting model prices as functions of the dependence parameters are presented in Figs. 8.3 through 8.5. An interpretation of the resulting model spreads is given in what follows: • An interesting observation is the monotonicity of spreads with respect to the dependence ordering within a group (stocks and credit) and in between the groups. On a qualitative level, these results are parallel for all pricing methods and therefore interpreted at once — ignoring minor differences. Firsttotrigger spreads are decreasing in dependence, which is consistent over all models and the two parameters of dependence. The reason behind this observation is that multiple trigger events, but also no trigger event at all, are more likely for a large correlation when trigger probabilities are kept fix. The secondtodefault spread is increasing in group dependence but slowly decreasing in dependence among stocks and credit. The thirdtotrigger spread is increasing in group dependence and almost flat in extragroup dependence.
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
192
Höcht et al. 1st to trigger spread (in bps)
2nd to trigger spread (in bps)
1
1
140
350 0.8
130
0.8 300
120 0.6 110
2
250
ρ
ρ2
0.6 0.4
0.4
100
200 0.2
90
0.2
80
150 0
0
0.2
0.4
ρ1
0.6
0.8
0
1
0
3rd to trigger spread (in bps)
0.2
0.4
ρ1
0.6
0.8
1
4th to trigger spread (in bps)
1
1 50
70 0.8
0.8
60
40 0.6 30
40
0.4
ρ
2
50
ρ2
0.6
0.4
20
30 0.2 0
0.2
20 10 0
0.2
0.4
ρ1
0.6
0.8
0
1
10 0
5th to trigger spread (in bps)
0.2
0.4
ρ1
0.6
0.8
1
6th to trigger spread (in bps)
1
1
40
20
0.8
0.8 30
15
20 0.4
10
0.4 10
0.2 0
ρ
2
0.6
ρ2
0.6
0
0.2
0.4
ρ1
0.6
0.8
1
0
5
0.2 0
0
0.2
0.4
ρ1
0.6
0.8
1
Figure 8.3 Model spreads from the structural approach (25,000 runs).
The {4,5,6}thto trigger spreads are all increasing in both dependence parameters, since more dependence (of either sort) makes multiple defaults more likely, with according influence on spreads. • Comparing the structural approach and the Gausscopula approach to the Gumbel copula shows that kthto trigger spreads for higher k are much larger for the latter. This is explained by the positive uppertail dependence of the Gumbel family. Heuristically speaking, this property means that the trigger variables Ui have a positive probability for being simultaneously close to one. Translated into the pricing framework, this implies multiple trigger events within one simulation run, hence, positive spreads for higher kthto trigger spreads. Extending the simple structural
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
193
Cross Asset Portfolio Derivatives 1st to trigger spread (in bps)
2nd to trigger spread (in bps)
1
400
1
0.8
350
0.8
0.6
300
0.6
130 120 110
ρ2
ρ2
140
250
0.4
0.4
200
0.2
100 90
0.2
80
150 0
0
0.2
0.4
ρ1
0.6
0.8
0
1
0
3rd to trigger spread (in bps)
0.2
0.4
ρ1
0.6
0.8
1
4th to trigger spread (in bps)
1
1 70
0.8
50
0.8 60
0.6
40 0.6
40
0.4
ρ2
ρ2
50
30 0.4 20
30 0.2 0
0.2
20 10
0
0.2
0.4
ρ1
0.6
0.8
0
1
10 0
5th to trigger spread (in bps)
0.2
0.4
ρ1
0.6
0.8
1
6th to trigger spread (in bps)
1
1 40
0.8 30
20
0.6
15
0.4
10
0.2
5
ρ2
ρ2
0.6
0.8
20
0.4
10
0.2 0
0
0.2
0.4
ρ1
0.6
0.8
1
0
0
0.2
0.4
ρ1
0.6
0.8
1
0
Figure 8.4 Model spreads from the Gausscopula model (250,000 runs).
model using jump processes might resemble this property. Another observation is that spreads for the limiting case of independence agree within all models (up to a small Monte Carlo noise). • A final remark addresses the required computation time for both approaches. The structural approach is based on the simulation of a path of a multivariate process (on a fine grid) in each simulation run. Compared to the copula approach, which only requires a sample from the respective copula in each run, it is therefore not surprising that the computation time for the evaluation of one option price with 250,000 simulation runs differs massively. Our implementation in Matlab on a standard PC requires about 3813 s for the structural and 11 and 18 s for the two copula approaches, respectively.
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
FA
194
Höcht et al. 1st to trigger spread (in bps)
2nd to trigger spread (in bps)
2
2
1.8
350
ϑ2
300
110
1.6
100
1.4
90
1.2
80
ϑ2
1.6
1.8
1.4 250 1.2 1
200 1
1.2
1.4
ϑ1
1.6
1.8
1
2
1
3rd to trigger spread (in bps)
1.2
1.4
1.6
ϑ1
1.8
2
4th to trigger spread (in bps)
2
60
2
1.8
50
1.8
40
ϑ2
40
1.4
30
1.6
25
ϑ2
1.6
35
20
1.4
30
15 1.2
1.2
10
20 1
1
1.2
1.4
ϑ1
1.6
1.8
1
2
5 1
5th to trigger spread (in bps)
1.2
1.4
1.6
ϑ1
1.8
2
6th to trigger spread (in bps)
2
2 30 25
1.8
1.6
20
1.6
15
1.4
15
10
ϑ2
ϑ2
1.8
1.4
10 1.2
5
1.2 5
1
1
1.2
1.4
ϑ1
1.6
1.8
2
1
1
1.2
1.4
ϑ
1.6
1.8
2
1
Figure 8.5 Model spreads from the Gumbelcopula model (250,000 runs).
8.5. OUTLOOK Innovative instruments in the credit derivatives universe allow investors to construct positions which refer to specific aspects of credit risk, e.g., spread volatility or (correlated) default risk. While the standard valuation approach for these instruments is still skewed to a pure transactiondriven style, i.e., ignoring interdependencies with other instruments and markets, we expect the portfolio perspective to soon become more important. For such a portfolio treatment, cross asset portfolio derivatives become
May 12, 2010
17:47
WSPC/SPIB913
Cross Asset Portfolio Derivatives
b913ch08
FA
195
interesting. Allowing for collateralized products referring to other asset classes than debt offers new opportunities for investments and risk management. Cross asset portfolio derivatives typically smooth the risk/return profile of volatile asset classes. Whereas the classical Capital Asset Pricing Model (CAPM) is still the appropriate model to optimize the underlyings regarding return, risk, and correlation, this portfolio management style changes in the case of cross asset portfolio derivatives. In the CAPM world, the risk/return optimization is already implemented in the market portfolio, which is defined as being efficient, while correlation between the portfolio constituents is defined as the comovement of asset prices. The individual risk perception is simply reflected by the investment share of the market portfolio, and the riskfree asset. In a cross asset portfolio derivatives world, the underlying portfolio is selectable without the usual µ/σconstraints. Correlation is rather joint trigger probability and, e.g., the specific tranche investment in case of a CTSO is done in line with the individual risk perception. Correlation, defined as comovement of assets, is replaced by the correlation of tail events, which induces a different view on the risk/return optimization. Systematic shocks, like the subprime crisis starting in 2007, are a crucial risk factor for a cross asset portfolio derivative, whereas market fluctuations and cyclical moves are of minor importance for the performance of such derivatives. To sum up, cross asset portfolio derivatives might be an interesting investment product or portfolio management tool. There should not be any concerns regarding demand/supply patterns. Both, demand and supply for tail event risk is immense, given the huge variety of players in this market and, e.g., the recent financial crisis or the rise/decline of the oil price. Traditional tailevent riskdriven companies (insurance sector) are probably skewed to selling this kind of risk to players like hedgefunds, banks, and fund managers, who can implement the attractive trading positions, but also optimize the multiasset portfolios, using cross asset trigger derivatives.
8.6. CONCLUSION An extension of popular portfolio credit derivatives to more general trigger derivatives is presented. The introduced class of cross asset portfolio trigger derivatives allows to securize the risk of correlated tail events in the form of swaplike insurance claims. Two pricing methodologies, transferred from the world of credit derivatives, are discussed and illustrated by an example of medium complexity. It is shown how sensitive model prices are with respect to the assumed concept of dependence. On the one hand, this should be seen as a word of warning from the perspective of model risk. On the other hand, one should keep in mind that typical cross asset portfolios contain exactly these correlated tail risks. Hence, a securitization of correlated tail risks might be interesting
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
196
FA
Höcht et al.
for both insurance seller and buyer. A standardized market of these derivatives would then allow to calibrate the models to the market’s perception.
Acknowledgments We thank Dr. Jochen Felsenheimer of Assenagon and Dr. Philip Gisdakis of UniCredit MIB for their initial thoughts on this product group. Moreover, we thank Marius Hofert (Universität Ulm) for valuable remarks on earlier versions of the manuscript. Views expressed in this paper are those of the authors and do not necessarily reflect the positions of the respective employers.
References [1] Schönbucher, PJ (2003). Credit Derivatives Pricing Models: Models, Pricing and Implementation, 1st edn. Wiley Finance Series, John Wiley and Sons, Inc. [2] Hull, J (2004). Options, Futures, and Other Derivatives. London: PrenticeHall International, Inc. [3] Felsenheimer, J (2006). The wonderful world of CAOs (strategy update). Technical report, UniCredit MIB Global Credit Research. [4] Standard and Poor’s (2006). CDO spotlight: Global criteria for securitizations of funds of hedge funds. Standard and Poor’s, Structured Finance. [5] Brigo, D and F Mercurio (2001). Interest Rate Models — Theory and Practice. Springer, Finance, Berlin: Springer. [6] Benth, FE, JS Benth and S Koekebakker (2008). Stochastic Modeling of Electricity and Related Markets. World Scientific. [7] Zhou, C (2001). An analysis of default correlations and multiple defaults. Review of Financial Studies, 14, 555–576. [8] Joe, H (1997). Multivariate Models and Dependence Concepts. New York: Chapman & Hall. [9] Nelsen, RB (1998). An Introduction to Copulas, 1st edn. Berlin: Springer. [10] McNeil, AJ, R Frey and P Embrechts (2005). Quantitative Risk Management. Princeton Series in Finance, Princeton University Press. [11] Li, DX (2000). On default correlation: A copula function approach. The Journal of Fixed Income, 9(4), 43–54. [12] Schönbucher, PJ and D Schubert (2001). Copuladependent defaults in intensity models. URL http://papers.ssrn.com/sol3/papers.cfm?abstract_id= 301968. Working Paper. [13] Marshall, AW and I Olkin (1988). Families of multivariate distributions. Journal of the American Statistical Association, 83(403), 834–841. URL http://www. jstor.org/stable/pdfplus/2289314.pdf. [14] McNeil, AJ and J Neslehova (2009). Multivariate Archimedean copulas, dmonotone functions and l1norm symmetric distributions. The Annals of Statistics, 37, 3059–3097. [15] McNeil, AJ (2008). Sampling nested Archimedean copulas. Journal of Statistical Computation and Simulation, 78(6), 567–581. [16] Hofert, M (2008). Sampling Archimedean copulas. Computational Statistics and Data Analysis, 52(12), 5163–5174. doi:10.1016/j.csda.2008.05.019.
May 12, 2010
17:47
WSPC/SPIB913
Cross Asset Portfolio Derivatives
b913ch08
FA
197
[17] Hofert, M (2008). Efficiently sampling Archimedean copulas. Working Paper. [18] Mai, JF and M Scherer (2009). Lévyfrailty copulas. Journal of Multivariate Analysis, 100(7), 1567–1585. [19] Black, F and J Cox (1976). Valuing corporate securities: Some effects of bond indenture provisions. Journal of Finance, 31, 351–367. [20] Cont, R and P Tankov (2003). Financial Modelling with Jump Processes, 1st edn. Chapman & Hall/CRC Press. [21] Musiela, M and M Rutkowski (2004). Martingale Methods in Financial Modelling (Stochastic Modelling and Applied Probability). Springer. [22] Hofert, M and M Scherer (2010). CDO pricing with nested Archimedean copulas. To appear in Quantitative Finance.
May 12, 2010
17:47
WSPC/SPIB913
b913ch08
This page intentionally left blank
FA
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
Part II
Alternative Strategies
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
This page intentionally left blank
FA
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
9
DYNAMIC PORTFOLIO INSURANCE WITHOUT OPTIONS
DOMINIK DERSCH Rumfordstr. 6, 80469 München, Germany
[email protected] Dynamic portfolio strategies are an interesting alternative to classical optionbased investment and protection strategies. One of the most prominent techniques is Constant Proportion Portfolio Insurance (CPPI). In this chapter, we provide a review of various techniques and formulate a general framework for investment and protection strategies. The common feature of this strategy is that it empowers the investor to replicate various option like payoff profiles without the usage of options. These strategies may replicate a simple floor type or advanced pathdependent lookback options that implement alltimehigh strategies with a given participation rate. We illustrate the different strategies that employ features like various types of lockin, trailing, leverage, and risky portfolio strategies with historical simulations. We include features that allow the simulation under realistic market conditions taking into account transaction costs and the avoidance of excessive rebalancing through transaction filters. We discuss the use of exchange traded funds (ETF) to invest in broadly diversified multiasset portfolios. The goal of this chapter was to illustrate different protection strategies and to show how a practical implementation of these strategies could look like. This chapter can serve as a guideline for simple spreadsheet models.
201
May 12, 2010
17:47
WSPC/SPIB913
202
b913ch09
FA
Dersch
9.1. INTRODUCTION Institutional investors who require portfolio protection or a minimum absolute performance target are facing the problem to trade off risk and return. On the one hand, exposure to high return risky asset classes is desirable; on the other hand, the downside potential that accompanies higher risk should be kept at the investor’s level of comfort. Holistically, the tasks of asset allocation and risk management are closely linked and should not be performed in two separate steps, as this leaves the investor most likely with suboptimal solutions. However, a fairly large arsenal of risk management techniques has been developed that work completely independent of the investment strategy and respective asset classes. The sole requirements are that the asset classes are investible and allow — at least in theory — a liquid market and continuous trading. The selection of the underlying depends on the specific requirements of the investor with respect to holding period, taxation, investment guidelines, etc., and the asset class itself. Examples of different underlyings are direct investments in stocks or bonds, futures, and funds. It is important to note that the investment universe and the respective protection strategy have to be closely aligned. Investments with large bid–offer spreads would require a strategy with a low reallocation frequency. For the sake of simplicity, this chapter focuses on this twostep approach described above: Step 1 asset allocation, Step 2 risk management framework. Throughout this chapter, the first step of asset allocation is simulated by investing in a set of indices. The performance of each investment is assumed to follow the performance of the index timeseries. The considered investment universe includes stock, bond, commodity, and hedge fund indices. Our choice of the investment universe poses no restriction on the second step of the investment process — the risk management framework. Within the fairly general requirements of “investability” and liquidity, any asset classes and (propriety) trading system may be embedded in this risk management framework. The recent history of financial markets posed a huge challenge to portfolio managers as reflected in massive declines in asset values, historical highs in volatility, and a break down in correlations. Dynamic portfolio strategies are an interesting alternative to classical optionbased investment and protection strategies that allow one to cope with the market turmoil. One of the most prominent techniques is Constant Proportion Portfolio Insurance (CPPI) [2, 3]. CPPI went out of fashion because of a number of drawbacks like the fixed time horizon, the inability to take profits and recover from a major drawdown. One of the major shortcomings — the procyclical behavior — has been blamed for huge market movements and the stock market crash of 1987. However, a number of advanced features allow one to overcome the shortcomings of the first generation model and allow a practical application in asset management.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
Dynamic Portfolio Insurance Without Options
FA
203
A comparison of optionbased strategies — mentioned above — and CPPI is shown in [1]. The mature market for exchange traded funds (ETF) gives access to a huge universe of different asset classes that may be traded with high liquidity and low bid–offer spreads. This enables the investor to apply advanced portfolio insurance strategies as described above and investment in a broadly diversified universe. The remainder of this chapter is structured as follows. In the next two sections, we review simple portfolio strategies and plain vanilla CPPI and illustrate them with historical simulations. In Sec. 9.4, we formulate a more general framework of CPPI with various features like different types of lockin, trailing, leverage, and risky portfolio rebalancing strategies. We include features that allow simulation under realistic market conditions taking into account transaction costs and the avoidance of excessive rebalancing through transaction filters. We further show that our framework also includes an extension of CPPI named TIPP [4]. In Sec. 9.5, the strategies are illustrated using historical simulations. In particular, we investigate how the different strategies cope with the historical market evolution. In Sec. 9.6, we discuss the use of exchangetraded funds (ETF) to implement protection strategies. This chapter concludes with final remarks on different risk transfer mechanisms of option strategies versus dynamic portfolio strategies.
9.2. SIMPLE STRATEGIES 9.2.1. BuyandHold Probably the most simple and most common risk management framework is the buyandhold strategy. Not just among retail investors, either deliberately or not deliberately buyandhold is a widespread approach. Many ambitious strategies will eventually drift into passive sitandwait strategies as the investor sits out long periods of negative performance or hesitates to take profits on time. Cheekily, buyandhold or strategic trades very often stem from shortterm tactical trades turned bad. The only reason why this strategy remotely qualifies as some kind of protection strategy is that in the absence of leverage the total loss is limited to the initial investment. However, buyandhold has performed well over long periods and across many asset classes.
9.2.2. StopLoss A stoploss strategy is the first nontrivial step toward a risk management framework. Here, we distinguish between an investment target IT on a present value base and a target on a given time horizon T . In the first case, the position has to be switched into a
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
204
FA
Dersch
riskfree investment matching the investment horizon if the portfolio value falls below the target value at any time PVportfolio (t) > IT .
(9.1)
In the second case, the portfolio value must not fall below PVportfolio (t) > e−r(T −t) · IT
(9.2)
PVportfolio (t) − e−r(T −t) · IT > 0.
(9.3)
or equivalently
Here, T − t is the time until the investment horizon is reached and r is the riskfree rate of this time span. The second case is slightly more complex. Here, we must ensure that the investor reaches his/her investment target at maturity. Therefore, the portfolio value must not fall below a certain floor value given by the right hand side of Eq. (9.2). In case, the stoploss level is reached, the portfolio must be liquidated and invested in the riskfree asset. Investing the floor value with the riskfree rate will ensure the given target value at maturity. The implementation of a stoploss strategy requires monitoring both the portfolio performance and the total return of the riskfree investment. The riskfree investment is usually implemented with treasury bills or bonds with a maturity matching the investment horizon. The reinvestment of coupons paid until maturity must also be considered. In case, the stoploss level is reached, a single portfolio reallocation occurs. Strictly speaking, stoploss is therefore not a dynamic strategy. In the above analysis, we assume that the investments and the target level are in the same consolidation currency. Otherwise, the respective FX spot or forward rates have to be additionally monitored. Please note that a stoploss strategy with a target value of zero is the same as the above buyandhold strategy.
9.2.3. The Bond Floor Strategy The stoploss strategy bears the risk that the portfolio value may be exposed to large volatility. In addition, it carries a short fall or gap risk. This is the risk that the stop level is missed in large market movements and the investor is left with a final portfolio value below the target. The bond floor strategy takes a more cautious approach. Here, we only invest in the risky portfolio the amount we are willing to lose in the first place. Let us take Eq. (9.3) at the beginning of the investment horizon PV portfolio (t = t0 ) − e−rT · IT = C.
(9.4)
C is the amount given by the difference between the initial investment and the present value of the target amount. Here, we assume an investment in a riskfree bond with a
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
Dynamic Portfolio Insurance Without Options
205
maturity matching the investment horizon. The bond floor strategy is usually implemented with a target value close to 100%. According to the above equation, a given investment of EUR 100, a target value of EUR 101 in one year, and a riskfree rate of 2.5% would leave us with a risk budget of EUR 1.5 and a bond investment of EUR 98.5. That means EUR 1.5 is invested with a buyandhold strategy in a risky investment. A total loss of the risky investment would still guarantee the target amount of 101 EUR in one year. For a target value close to 100%, the bond floor strategy — as compared to the stoploss strategy — swings the pendulum in the other direction of extreme risk aversion with the consequences of little upside potential beyond the target value. There are two things worth mentioning: The bond floor must not be larger than the amount that may be earned with the riskfree investment and a bond floor of zero is the same as the above buyandhold strategy.
9.2.4. Plain Vanilla CPPI CPPI tries to bridge the gap between high risk and high risk aversion of the above strategies. The Constant Proportion Portfolio Insurance technique [2] was first introduced by Black and Jones in 1987. It may be seen as a further generalization of stoploss and bond floor and it is literally a dynamic strategy. Similar to both strategies, a target level and an investment horizon are given. In contrast to the bond floor strategy, we assume that a total loss of the risky investment is highly unlikely. We rather accept that the risky investment may fall by a factor of 1/M within a given time horizon. This means that if we invest twice (e.g., M = 2) the amount given by the right hand side of Eq. (9.3) in the risky portfolio, we still meet the investment target if the investment will lose less than 50% of the initial value. This is the key idea of CPPI. The strategy may be summarized by the following set of steps: (1) Calculate the current risk budget C(t) similar to Eq. (9.4) according to C(t) = PV portfolio (t) − e−r(H−t) · T,
(9.5)
with PV portfolio (t = t0 ) = N0 . (2) Calculate the exposure E(t) invested in the risky portfolio according to E(t) = M · C(t).
(9.6)
(3) Rebalance the portfolio by investing the amount E(t) in the risky portfolio and the remaining amount in the riskfree investment. (4) Wait until t has passed and go back to Step (2). (5) Repeat the above steps until the end of the investment horizon T is reached. In case C(t) according to Eq. (9.5) is zero, the portfolio will be allocated in the riskfree investment until the end of the horizon. In that case, only the target value is achieved.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
206 Table 9.1 Strategy
Dersch Comparison of Simple Protection Strategies and Their Relation to CPPI. Risk appetite
Upside potential
Short fall risk
Relation to CPPI
Buyandhold
High
High
No
CPPI with zero target
Stoploss
Medium, depends on stop level
High, depends on stop level
Yes
CPPI with very large multiplier and target equal to stoploss level
Bond floor
Low, depends on bond floor level
Low, depends on bond floor level
No
CPPI with target level equals bond floor and multiplier M = 1
CPPI
Low, depends on target level
Medium, depends on target level
Yes, but small Yes
Depending on the parameterization, CPPI allows one to implement a given risk appetite: The larger the multiplier and the lower the target level, the higher the risk appetite. It is easy to show that CPPI contains the above strategies buyandhold, stoploss, and bond floor for different settings of target level and multiplier. The discounted target level is also called the floor F(t) = E−r(T −t) · IT.
(9.7)
The floor is the present value of the target. Table 9.1 summarizes the risk and reward characteristics for the above strategies and shows the link to CPPI.
9.3. HISTORICAL SIMULATION I In this section, we show sample simulations for the above protection strategies. Our risky portfolio is the Dow Jones Euro Stoxx 50. The simulation covers a period of close to 10 years. For comparative reasons, we use similar protection levels where applicable. • • • • •
Table 9.2 summarizes the simulation parameter. Figure 9.1 shows the buyandhold strategy. Figure 9.2 shows the stoploss strategy. Figure 9.3 shows the bond floor strategy on the DJ Euro Stoxx 50 index. Figure 9.4 shows the CPPI strategy with a target level of 100% and a multiplier of 6. The floor is calculated using a discount rate of 3.25%. • Table 9.3 summarizes the simulation results.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
207
Dynamic Portfolio Insurance Without Options Table 9.2
Simulation Parameters of Simple Strategies.
Simulation period Investment Riskfree and risky investment Target Discount rate Multiplier
4 January, 1999 — 12 December, 2008 EUR 100 mn EUR Overnight liquidity, DJ Euro Stoxx 50 100% (except buyandhold) 3.25% (except buyandhold) 6 (CPPI only)
180 160 140 120 100 80 60 40 20 0 1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
Figure 9.1 Buyandhold strategy on the DJ Euro Stoxx 50 performance index for an initial portfolio value of EUR 100 mn. The performance of the portfolio (in EUR mn) is the performance of the index.
In this market environment, the bond floor strategy seems to perform best over the complete period both in the absolute return and the size of the worst drawdown. The three other strategies all reach the same maximal portfolio amount of EUR 156.3 mn (up 56%), but fail to benefit at maturity. The stoploss and CPPI strategy show a very similar picture. They are both stopped out during the sharp market decline in mid2002 and realize a slight loss as compared to the initial portfolio value. This translates into a slightly negative annual return of −0.17% (CPPI) and — 0.30% (stoploss). Buyandhold ranks last with respect to final portfolio amount annual return, and suffers the worst portfolio drawdown of 65% of the previous alltime high value. Stoploss and CPPI both fail the target by a small amount. This may be due to two reasons. The first is fundamental and is caused by rapid market movements when the position may only be liquidated below the theoretical stop level. We can reduce this fundamental risk by monitoring the position intraday. The second reason is caused by a simplification of our simulation framework. Monitoring the stoploss or floor level requires one to monitor the zero coupon bond with a maturity equal to the remaining
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
208
Dersch 180 DJ EUR STOXX 50
riskfree
160 140 120 100 80 60 40 20 0 1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
Figure 9.2 Stoploss strategy on the DJ Euro Stoxx 50 performance index with a target level of 100% at maturity. The riskfree investment is EUR overnight liquidity. The yaxis shows the portfolio value in EUR mn.
160 DJ EUR STOXX 50
riskfree
140 120 100 80 60 40 20 0 1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
Figure 9.3 Bond floor strategy on the DJ Euro Stoxx 50 index with a target level of 100% at maturity. The riskfree investment is EUR overnight liquidity. The yaxis shows the portfolio value in EUR mn.
investment horizon. As a simplification, a fixed discount rate is used in our simulation framework to calculate the stoploss level and floor. Our simplified riskfree investment is EUR overnight liquidity rather than the corresponding zero bond with matching investment horizon. We may fail to reach the target if the overnight investment fails to earn — over the remaining investment horizon — the return implied by the floor level.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
209
Dynamic Portfolio Insurance Without Options 180 DJ EUR STOXX 50
riskfree
floor
160 140 120 100 80 60 40 20 0 1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
Figure 9.4 CPPI strategy on the DJ Euro Stoxx 50 performance index with a target level of 100% at maturity. The riskfree investment is EUR overnight liquidity. The floor (black line) is assumed to follow an annual rate of 3.25%. The yaxis shows the portfolio value in EUR mn. Table 9.3 Historical Simulation of Four Different Protection Strategies Using the Dow Jones Euro Stoxx 50. We Calculate the Annual Return Over the Simulation Period and the Worst DrawDown. A DrawDown of, e.g., 65% Means That the Portfolio Lost 65% of Its Previous AllTimeHigh. Strategy
Buyandhold
Stop loss
Bond floor
Ann. return Reached protection Worst drawdown
−1.69% NA 64.64%
−0.30% Yes 48.47%
2.03% No 27.26%
CPPI −0.17% Yes 47.34%
9.4. ADVANCED FEATURES The above simulations illustrate different protection strategies. However, they are not suitable for practical use for a number of reasons: • Transaction costs may have an impact on the real world performance. • Frequent rebalancing may cause excessive transaction costs and should therefore be constrained by transaction filters. • Investors require more sophisticated protection strategies like lockin of gains. • Investors may wish a leveraged exposure to the risky portfolio. • For nontrivial risky portfolios, e.g., more than one risky asset, different rebalancing strategies for the risky portfolio may be applied.
May 12, 2010
17:47
WSPC/SPIB913
210
b913ch09
FA
Dersch
In our simulation environment, we implemented a fairly general framework that takes into account the above features.
9.4.1. Transaction Costs Independent of asset classes and markets, transaction costs is a relevant factor that may significantly impact the performance of a trading system. Trading systems that look good on paper, e.g., paper trading, may fail in reality because transaction costs were not considered. The impact of transaction costs is more dominant for frequent trading and for long investment periods. The latter is important because of the compounding effect. Transaction costs that reduce the portfolio amount in an early stage are no longer available for future investment. The impact is hard to estimate and therefore has to be simulated. In our simulation environment, we model transaction costs as a percentage of the transaction volume. Different transaction costs can be set for the riskfree and the risky assets and for buying and selling the asset. The bid–offer spreads of ETF, for example, vary widely from 2 bp to up to 100 bp depending on the asset class and time. This corresponds to a transaction cost of 0.0001–0.005 times the transaction volume. The transaction cost is only half because the bid–offer spread is paid on the full round trip to get in and out of the asset.
9.4.2. Transaction Filter With the ability to model transaction costs, the impact can be analyzed and optimized. Here, we have to trade off flexibility versus rigidity — rapid adaptation to changes in the market environment on the one hand with the downside of a large number of transactions and high transaction costs. On the other hand, less frequent trading reduces the transaction costs but poses the risk that the system is not flexible to respond to large market moves. The introduction of transaction filters , λ, and t allows the investor to trade off these two effects. Strictly speaking a rebalancing in the plain vanilla CPPI is required whenever Eq. (9.6) is violated. We calculate the deviation after time t = t + t has passed according the following equation: E(t) (9.8) 1 − buy ≤ target ≤ 1 + sell . E (t) A rebalancing is performed only if the target exposure Etarget (t) deviates from the current exposure E(t) beyond the given boundaries. Typically, the parameter for buy and sell are in the range of [0,0.2], with sell slightly smaller than buy , in order to react quicker to a market downturn. Larger defines a higher threshold for rebalancing, = 0 requires instant rebalancing.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
Dynamic Portfolio Insurance Without Options
FA
211
The volatility filter λ scales the size of the rebalance — either buy or sell — if a rebalancing is triggered according to Eq. (9.8). E = λ · Eold (t) − Enew (t),
(9.9)
with λ in the range of [0.7,1.0]. For λ smaller than one, we follow only a fraction of the rebalancing amount. The idea behind the volatility filter is that the market trades in a range rather than follows a trend. The reobservation period t is also a transaction filter: The larger t, the more time elapses between a test of Eq. (9.8). The reobservation period has to be in line with . Small and a large t or very close to one and small t do not fit well together. Typically, t is in the range of one day. This implies that we test Eq. (9.8) at close, but ignore intraday movements outside the range defined by . This reflects the observation that the intraday volatility is typically larger than the daytoday volatility.
9.4.3. Lockin Levels The plain vanilla CPPI does not protect any gains. To remedy this weakness, different strategies to lockin gains by raising the target level have been proposed. The observation period for lockin may be given by t, its multiples (k · t), or by any other discrete lockin dates like every week or month. To implement lockin, we have to distinguish between the lockin trigger and the lockin action. The trigger can be a simple trigger in time as described above, or a trigger caused by a certain portfolio level or given by both. The different lockin actions are performed if a lockin trigger is reached. They are summarized as follows: • Discrete lockin steps: X% gain in the portfolio amount is locked in by raising the floor defined in Eq. (9.7). The discrete lockin steps may refer to the fraction of the initial notional at the beginning of the investment period, like every EUR 100,000 or to the notional at the previous lockin level (compounding lockin), like 10% of the portfolio amount at the last lockin. • Newly reached alltimehigh levels are locked in by raising the floor with respect to the previous high level. • Trailing: Y % gain in the portfolio amount is locked in by raising the target level by the given gain. The discrete lockin steps may refer to the initial target level at the beginning of the investment period or refer to the previous lockin level (compounding lockin). • Alltimehigh trailing: Newly reached alltimehigh levels are locked in by raising the target level with respect to the previous high level. The difference between the first two and the last two lockin actions is that the lockin is applied to the floor level rather than the target level.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
212
FA
Dersch
The advantage of a lockin is that gains that have been reached in the past are protected. On the other hand, a lockin may also cause a reduction in the exposure due to a rising floor level and therefore limits participation in a future rise in the risky portfolio.
9.4.4. Leverage and Constrain of Exposure The lockin levels discussed above support the requirement to protect gains in the risky portfolio. On the other hand, investors may request further upside potential through higher exposure to the risky portfolio. We may introduce leverage by rewriting Eq. (9.6) E(t) = min{Emax (t), M · C(t)}.
(9.10)
In the absence of leverage, we define Emax (t) = PV portfolio (t).
(9.11)
This means that the exposure to the risky portfolio may not be larger than the current portfolio value. We cannot invest more than our current notional amount. More generally, we may write Emax (t) = K · PV portfolio (t).
(9.12)
For K > 1, we allow leverage up to a certain level. For 0 < K < 1, we constrain the maximum exposure to the risky portfolio. The latter has a similar effect as a lockin. In addition, we may scale Emax with respect to the initial investment PV (t0 ) rather than the current portfolio value. The leverage and constraint in the above description refer to the risky portfolio. Similarly, we may define leverage with respect to the riskfree investment defined by Riskfree(t) = max{Riskfreemin (t), PV portfolio (t) − E(t)}.
(9.13)
In the absence of leverage, we find Riskfreemin (t) = 0.
(9.14)
This means that it is not allowed to borrow funds in order to invest in the risky portfolio. More generally, we may write: Riskfree(t) = k · PVportfolio (t).
(9.15)
For k < 0, we allow borrowing and thus leverage up to a certain level. For 0 < k < 1, we request a minimum amount to be invested in the riskfree asset, which constrains the exposure to the risky portfolio. This has a similar effect as a lockin. As proposed above, we may also scale the riskfree exposure with respect to the initial investment PV (t0 ) rather than the current portfolio value.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
Dynamic Portfolio Insurance Without Options
FA
213
In our framework, we implemented leverage and constraint of the risky portfolio and the riskfree asset. Both features may be used at the same time. The weaker criterion determines the overall portfolio leverage or constraint.
9.4.5. Rebalancing Strategies for the Risky Portfolio Up to now, we have not discussed how a rebalancing of the risky portfolio is propagated to the portfolio constituents. In principle, we could perform a complete portfolio optimization in each step and adjust the portfolio weights accordingly. This may include Markowitz, VaR, or CVaR optimization with different types of linear, nonlinear constraints, or boundary conditions. For the sake of simplicity, we show three types of simple portfolio rebalancing strategies: • Balanced: Here, the notional of all N risky portfolio constituencies is rebalanced in accordance with the initial weights defined at the start of the strategy. This approach reduces the amount of the aboveaverageperforming risky assets and increases the amount of the belowaverageperforming risky assets. The balanced approach is an anticyclical profittaking strategy that assumes a meanreverting market within the universe of the risky assets. • Proportional: Here, the risky portfolio constituencies are rebalanced proportional to current weights. This strategy implements a simple trendfollowing approach with a soft competition among the risky assets. • Squared: Here, the risky portfolio constituencies are rebalanced proportional to the square of the current weight. Similar to the proportional strategy, this approach implements a trend follower but with fierce competition among the risky assets due to the squaredweighting factor.
9.4.6. CPPI and Beyond The described framework allows implementation of a wide range of different strategies customized to the risk appetite and investment guidelines of the investor. These strategies are independent of the invested asset classes. The advanced features may also make the simple strategies buyandhold, stoploss, and bond floor more flexible. When these strategies are extended from a static to a dynamic portfolio approach, the transaction filters are also very useful. The advanced features also offer the opportunity to remedy the weaknesses of the original CPPI, like the lack of protecting gains, the fixed time horizon, and the limited potential to recover from a large drawdown. The guaranteed target level at a fixed time horizon is a feature that may not be necessarily required by an investor as this requirement has also a significant downside: There is little upside potential to take
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
214
FA
Dersch
advantage of a market recovery if the risk budget is strongly depleted after a downturn in the risky portfolio. The situation is worse if this happens in the early stage of the strategy. One possibility to remedy this weakness is the minimum exposure CPPI. Here, the allocation in the risky portfolio is held at a minimum guaranteed level. This ensures participation over the complete lifetime of the strategy. But the participation comes at a cost. An additional option has to be purchased to guarantee the target level at maturity. In this chapter, we focus on portfolio strategies without options and therefore follow a different approach. If we forgo the requirement of a target level at a fixed time horizon, we can instead attempt to secure — at any time — a fraction, e.g., 80% of a past portfolio value. In our advanced CPPI framework, this approach can be implemented by setting the discount rate in Eq. (9.7) to zero. Now target and floor are the same. In order to provide a risk budget, the initial target value has to be below 100% of the notional amount invested. Adding the lockin type alltimehigh trailing results in a strategy that has been described as TimeInvariant Portfolio Protection (TIPP) [4]. The main characteristics are • No fixed investment horizon is required. • Guaranteed instantaneous target level with an initial value below the investment amount. • Guarantee to recover from a market rebound because the riskfree portfolio continuously generates a new risk budget. • Implements a sequence of alltimehigh lookback options with a participation rate, which equals the target level but without the requirement of a fixed expiry date. • The present value of the portfolio may fall with a growing investment horizon in falling or sidewardmoving markets. The last point is a drawback of the approach. However, the investor can redeem the investment at any time.
9.5. HISTORICAL SIMULATION II In this section, we illustrate the aboveadvanced features with individual simulations and analyze the impact on the performance in the current volatile market environment.
9.5.1. Transaction Costs and Transaction Filter We mentioned above that it is important to include transaction costs in a simulation in order to get an idea of the performance under realistic conditions.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
215
Dynamic Portfolio Insurance Without Options Table 9.4 Simulation Parameters to Study the Impact of Transaction Filters on Transaction Costs. Four Different Simulations are Carried Out. Simulation period Investment Target Multiplier Riskfree investment Risky investment Discount rate Trans. filter (Sim. 1 and 3) Trans. cost (Sim. 1 and 2)
4 January, 1999–26 December, 2006 EUR 100 mn 100% 3 EUR overnight liquidity DJ Euro Stoxx Select Dividend 30 3.25% (buy) = 0.12, (sell) = 0.08, λ = 0.85, t = 1 day Risky investment 60 bp, riskfree investment 1 bp
In the following, we carry out four simulations with and without transaction filters combined with and without transaction costs. The risky asset is the DJ Euro Stoxx Selected Dividend 30 index. The simulation parameters are summarized in Table 9.4. To illustrate the impact of transaction costs, we selected a period of a rising market. The transaction costs are set to 60 bp of the transaction volume for the risky investment and to 1 bp for the riskfree investment. The transaction costs are deducted from the portfolio at the time of the transaction. In our simulation, we ignore the effect of slippage and partial execution. Figure 9.6 shows one out of the four simulations using transaction filters and considering transaction costs. Table 9.5 summarizes a comparison of different simulations with and without transaction filters and transaction cost. Without the transaction filter, there is a huge impact of transaction costs. The difference is EUR 54.45 mn. This means that the naïve approach — Simulation 4 — would suffer a drop of more than 20% of the final portfolio value in case transaction costs have to be taken into account (Simulation 2). Using the transaction filter results in much smaller dependency on transaction costs. The difference here is only Table 9.5 Comparison of Different Historical CPPI Simulations With (Without) Transaction Filters, Simulations 1 and 3 (Simulations 2 and 4) and With (Without) Transaction Costs, Simulations 1 and 2 (Simulations 3 and 4). For the Different Combinations, We Show the Total Number of Rebalancing Steps, the Portfolio Value at Maturity and the Total Transaction Costs. Simulation
Trans. filter
Trans. costs
# of transactions
PV final in EUR (mn)
Transaction cost in EUR (mn)
1 2 3 4
Yes No Yes No
Yes Yes No No
66 1923 52 1688
230.64 215.40 254.53 269.85
−2.04 −8.84 — —
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
216
FA
Dersch
EUR 23.90 mn (9.4%). The above simulation clearly reveals the damaging effect of transaction cost in the absence of the transaction filter. There is a primary effect caused by pure transaction costs that are deducted from the portfolio. This accounts for EUR −2.04 mn (Simulation 1, with transaction filter) and EUR −8.84 mn (Simulation 2, without transaction filter). There is a secondary effect that shows the total effect of transaction cost. A comparison of the final portfolio values shows that in case of transaction filters the difference between portfolio values of EUR −23.90 mn contains transaction costs in the amount of EUR −2.04 mn. In the absence of transaction filters, the performance gap is EUR −54.45 mn, where EUR 8.84 mn is the pure transaction cost. In this context, the transaction cost ratio tcr =
PV (tc) − PV (ntc) TC
(9.16)
is a useful ratio to measure the secondary effect of transaction cost. Here, PV (tc) is the performance with transaction cost and PV (ntc) without transaction cost and TC the pure transaction cost incurred over the observation period. A tcr close to one means that the performance difference with and without transaction cost is mainly caused by pure transaction cost. A tcr of two means that the transaction cost causes a performance reduction of twice the amount of the pure transaction cost. In our simulation, the tcr is equal to 11.71 (with transaction filter) and 6.16 (without transaction filter). The tcr depends on the compounding effect and the leverage of exposure. In a CPPI simulation with a multiplier of 3, every EUR of risk capital (cushion) changes the exposure to the risky portfolio by EUR 3. Transaction costs that reduce the exposure early in the investment period may have a tremendous impact on the performance later on. Ironically, in a falling market, transaction costs may even have a positive impact in the presence of a high multiplier as they may force the early reduction of the exposure and save the portfolio from otherwise higher losses. We also found that if the floor level is reached early in a simulation, transaction costs become less significant as it makes no difference whether a further falling market or transaction costs are the cause for a decline in portfolio value. In general, we conclude that the effect of transaction costs is not very easy to estimate. Here, simulations shed light on the effect.
9.5.2. Lockin Levels The sample simulation of the simple strategies illustrated a fundamental drawback of the protection strategies: the inability to take profits. An historical simulation on the DJ Euro Stoxx 50 index illustrated how the gains of more than 50% that were accumulated in the year 2000 are wiped out again (please compare Fig. 9.4). In order to remedy this weakness, we introduced a number of different lockin strategies that are designed to protect gains. In this section, we perform historical simulations to evaluate the performance of different lockin strategies. In order to get a direct comparison, we
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
217
Dynamic Portfolio Insurance Without Options
Table 9.6 Parameter Settings for Three Different Advanced CPPI Simulations. The Simulation Parameters Differ only in the Discount Rate for the Floor, the LockIn Trigger, and LockIn Action (the Last Three Rows). Simulation period Investment Target Multiplier Riskfree Risky investment Transaction filter Transaction cost Discount rate Lockin trigger Lockin action
4 January, 1999–12 December, 2008 EUR 100 mn 80% 6 EUR overnight liquidity DJ Euro Stoxx 50 (buy) = 0.12, (sell) = 0.08, λ = 0.85, t = 1 day Risky investment 8 bp, riskfree investment 1 bp 3.25% (Simulations 1 and 2), 0% (Simulation 3) NA (Simulation 1, no lockin), monthly, alltimehigh (Simulations 2 and 3) NA (Simulation 1, CPPI without lockin), trail alltimehigh by moving up the floor (Simulations 2 and 3)
perform our simulation on the Dow Jones Euro Stoxx 50 index. Table 9.6 shows the simulation parameters. For comparative reasons, we first show the CPPI simulation without lockin (Fig. 9.5). As compared to the previous simulation shown in Fig. 9.4, we use a transaction filter, consider transaction costs, and a target of only 80% as compared to 100%. 250 DJ EUR STOXX Select Dividend 30
riskfree
floor
200
150
100
50
0 1999
2000
2001
2002
2003
2004
2005
2006
Figure 9.5 Historical CPPI simulation under realistic condition on the DJ Euro Stoxx Selected Dividend 30 index using transaction filters and transaction costs. For a detailed specification, please refer to the text above. The yaxis shows the portfolio value in EUR mn.
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
218
Dersch 180 DJ EUR STOXX 50
riskfree
floor
160 140 120 100 80 60 40 20 0 1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
Figure 9.6 Historical CPPI simulation on the DJ Euro Stoxx 50. The simulation parameters are shown in Table 9.6. The areas show the allocation of the two assets. The black line is the floor level. The yaxis shows the portfolio value in EUR mn.
Qualitatively, we find a similar result in Fig. 9.6 as before. The decline in the stock market leads to a complete switch into the riskfree asset at the end of the investment horizon. All previous gains are wiped out again. Due to the lower target of 80%, the portfolio is still invested in the risky asset after the sharp decline in 2002 as compared to a target level of 100%. Figure 9.7 illustrates the impact of the alltimehigh trailing. The floor is moved upwards with a rising market. The initial target level of 80% trails the complete upward market movement of 56%, resulting in a final portfolio amount of 133.8% of the initial value. The simulation clearly demonstrates that trailing allows one to protect gains that have been previously accumulated. The floor value is slightly missed for the reasons already mentioned above. As a consequence, the increased target level causes a complete exit from the risky investment until the end of the investment horizon. The strategy protects gains, but is unable to recover from the drawdown. This weakness is removed in the next simulation shown in Fig. 9.8. Compared to the previous simulation, we now discount the floor level with zero. As a result, the floor is a horizontal line shifted upwards when new alltimehighs are reached. The floor level marks the guarantee level. Discounting the floor with a rate of zero reduces the risk budget because the floor is guaranteed instantaneously and not at maturity. The first half of the simulation period shown in Fig. 9.8 is similar to the simulation in Fig. 9.7. The floor trails the rising index value and the following decline results in a
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
219
Dynamic Portfolio Insurance Without Options 180 DJ EUR STOXX 50
riskfree
floor
160 140 120 100 80 60 40 20 0 1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
Figure 9.7 Historical CPPI simulation on the DJ Euro Stoxx 50 using the alltimehigh trailing with monthly lockin triggers. The further simulation parameters are shown in Table 9.6. The areas show the allocation of the two assets. The yaxis shows the portfolio value in EUR mn. The black line is the floor level. The floor value is slightly missed for reasons mentioned above. 200 DJ EUR STOXX 50
riskfree
floor
180 160 140 120 100 80 60 40 20 0 1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
Figure 9.8 Historical CPPI simulation on the DJ Euro Stoxx 50. The simulation parameters are shown in Table 9.6. The areas show the allocation of the two assets. The black line is the floor level. The yaxis shows the portfolio value in EUR mn.
reallocation to the riskfree investment. Here, this reallocation occurs earlier because of the reduced risk budget and we therefore do not reach the same alltimehigh in the portfolio value. The second half of the simulation shows a significantly different picture as compared to Fig. 9.7. The strategy participates in the rising market starting
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
220
Dersch
in 2003. In 2005, the strategy is again allocated in the risky investment to a very high proportion for a very short period of time. The guarantee level is moved upwards again. As a result, we reach a final portfolio value of EUR 167.50 nm. For the given example, the simulation shown in Fig. 9.8 is superior to the other strategies because of two reasons: • The ability to recover from the floor. • The target value is guaranteed instantaneously and not at a preset maturity. The reason for the rising from the dead like behavior of this strategy is the fact that even on the floor the riskfree asset continuously regenerates a cushion. Supported by a large multiplier, the exposure is quickly scaled up again. By the end of 2005, a new alltimehigh levels are reached. They are trailed with a participation rate which equals the target level (80%). Table 9.7 summarizes the results for the three different simulations. The TIPP strategy performs best as compared to the two other strategies.
9.5.3. The Use of Leverage Lockin is a conservative feature. It allows the protection of past gains. On the other hand, each lockin reduces the risk budget and therefore may constrain future exposure to the risky asset. Leverage has an opposite effect. It increases the exposure by borrowing riskfree and investing it in the risky portfolio. It therefore allows a leveraged investment in the risky portfolio. In the following, we illustrate leverage with one example on the Credit Suisse/Tremont Investable Hedge Fund Index. The index consists of 60 different hedge funds that represent 10 different strategies. For more information on the index, please see [6]. There exist institutional and retail products on this index. Table 9.8 shows the parameter settings for this simulation. Please note that the rebalancing frequency is monthly and we assume borrowing at the riskfree overnight rate. Table 9.7 Comparison of Different Historical CPPI Simulations With and Without LockIn (Columns 2 and 3) and with LockIn and Flat Floor (Column 4). For the Three Combinations, We Show the Portfolio Return, the Final Value and the Final Floor Value. The Floor Value for Simulation 2 is Slightly Missed for the Reasons Mentioned Above. Strategy
(1) CPPI
(2) CPPI with lockin
Ann. return End value EUR mn Floor Final in %
−2.25% 79.76 80.00%
2.97% 133.78 136.26%
(3) TIPP: lockin and flat floor 5.33% 167.53 166.80%
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
221
Dynamic Portfolio Insurance Without Options Table 9.8
Parameters for an Historical Advanced CPPI Simulation With Leverage.
Simulation period Investment Target Multiplier Riskfree Risky investment Transaction filter Transaction cost Discount rate Lockin trigger Lockin action Leverage
4 January, 2000–4 December, 2008 USD 1 mn 80% 6 USD overnight liquidity Credit Suisse Tremont Investable index (buy) = 0.12, (sell) = 0.08, λ = 0.85, t = 1 day Riskfree investment 1 bp, risky investment 50 bp 0% Portfolio PV increased by 10% of previous value tested on a monthly base Increase target by 10% of initial portfolio PV Up to 60% of the initial investment of USD 1 mn may be borrowed at any time to increase the exposure to the risky asset
2.5 2.0 1.5 1.0 0.5 0.0 0.5 1.0 2000
CS Trem. Inv. 2001
2002
riskfree 2003
2004
floor 2005
portfolio 2006
2007
index 2008
Figure 9.9 Historical advanced CPPI simulation on the Credit Suisse Tremont Investable Index. The yaxis shows the portfolio value in EUR mn. The leverage is limited to USD 0.6 mn. The areas show the allocation of the two assets. Leverage shows up in a negative allocation of the riskfree asset (light gray area). The grayandblackdashed lines are the performance of the index and portfolio, respectively. The black line is the floor.
Figure 9.9 shows an historical simulation of the above strategy. The light gray area represents the riskfree investment. Until 2008, we find a leveraged investment. The allowed leverage of USD 600,000 is utilized to the maximum level at certain times in 2000, 2005, 2006, 2007, and January 2008. Each increase of the target level
May 12, 2010
17:47
WSPC/SPIB913
222
b913ch09
FA
Dersch
(black line) is followed by a reduction of leverage. In this example, the leveraged investment results in an outperformance of the hedge fund index (compare dashed black line versus dashed gray line). That means that the proportion based on leverage yields an additional positive return after borrowing and transaction costs. The final portfolio value is USD 1.57 mn. This corresponds to an annual return of 5.19%. The current floor level is USD 1.52 mn. There has been no short position in the riskfree investment since October 2008. Currently, the portfolio still holds a 30% investment in hedge funds. In this example, the lockin feature mitigates the effect of leverage. Without lockin, the leverage budget of USD 600,000 mn would be fully utilized from 2001 until the end of the simulation period. A higher leverage together with lockin would not make a significant difference as the lockin and multiplier affect the maximum amount to be borrowed (data for both simulations are not shown). This emphasizes the requirement that strategy parameters are interdependent and have to be carefully adjusted.
9.5.4. CPPI on a MultiAsset Risky Portfolio In this example, we illustrate a risky portfolio of different asset classes represented by different performance indices, namely, equity (DAX), fixed income (Rex), and commodities (Dow Jones AIG Commodity Index) classes. For the three assets, there exist exchangetraded funds. The three asset classes have been selected based on the low historical correlation of daily logreturns. In this example, we make use of the portfolio rebalancing feature proportional. The simulation parameters are shown in Table 9.9. Table 9.9 Parameter Settings for an Historical Advanced CPPI Simulation on a Multi Asset Portfolio Including the DAX, Rex, and Dow Jones AIG. Simulation period Investment Target Multiplier Riskfree Risky investment
Transaction filter Transaction cost Discount rate Lockin trigger Lockin action Rebalancing strategy
3 January, 2000–12 December, 2008 EUR 100 mn 80% 6 EUR overnight liquidity, DAX (initial weight 30%), Rex (initial weight 50%), Dow Jones AIG Commodity (initial weight 20%) (buy) = 0.12, (sell) = 0.08, λ = 0.85, t = 1 day Risky investment 20 bp, riskfree investment 1 bp 0% Monthly, alltimehigh Trail alltimehigh by moving up the floor Proportional
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
FA
223
Dynamic Portfolio Insurance Without Options 160 REX
DAX
riskfree
DJ AIG Commodity Index
floor
140 120 100 80 60 40 20 0 2000
2001
2002
2003
2004
2005
2006
2007
2008
Figure 9.10 Advanced historical CPPI simulation on a risky portfolio of three different lowcorrelated asset classes. The four different shaded areas indicate the allocation over time. The black line is the floor level.
Figure 9.10 illustrates how the allocations in different asset classes evolve over time. From the end of 2000 until the beginning of 2005, the target level is constant. There are two periods with a full investment in the risky asset (2000 and the end of 2004 until the end of 2005). The final portfolio value is EUR 131.32 mn. This corresponds to a return of 3.0% p.a. mainly attributed to the years 2000 and 2005–2007. The present value time weighted asset allocation is 45% Rex, 21% AIG Commodity Index, 17% DAX, and 17% riskfree. This contrasts the current allocation of 22% Rex, 7% AIG Commodity Index, 6% DAX, and 66% riskfree. This is intuitive as we are currently very close to the floor level of EUR 123.95 mn.
9.6. IMPLEMENT A DYNAMIC PROTECTION STRATEGY WITH ETF The strategies shown in this chapter may be implemented by a direct investment in shares or baskets of shares, bonds, or by an indirect investment via Futures. Exchange Traded Funds (ETF) are an attractive alternative investment vehicle. ETF offers an investment in a wide range of different asset classes. All sample investments used in this chapter — except the Hedge Fund and Commodities index — may be implemented by investing in a corresponding ETF. The ETF on the Dow Jones Stoxx Selected Dividend 30 closely replicates the corresponding index. The index composition implements a portfolio strategy in its own sense as the portfolio constituencies are dynamically
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
224
FA
Dersch Table 9.10
Comparison of Various Characteristics of ETF’s Please See Also [5].
ETF
ISIN
Bidoffer spread in bp
Management fee in bp pa
eb.rexx Goverm. Germany 5.5–10.5 DJ Euro Stoxx 50 DJ Stoxx 50 DJ Euro Stoxx Select. Dividend 30 DJ Stoxx Selected Dividend 30 Dow JonesAIG Commodity
DE 000 628 949 9 DE 000 593 395 6 DE 000 593 394 9 DE 000 263 528 1 DE 000 263 529 9 DE 000 A0H 0728
7 8 43 56 72 99
15 15 50 30 30 45
adjusted from the Dow Jones Stoxx 600 universe based on their dividend yield and dividend consistency. For an exact definition of the index, please refer to the corresponding description of Dow Jones. Due to the nature of the index creation, the portfolio follows a more conservative anticyclical profittaking strategy. Generally, the underlying rational of the index, its dynamic and characteristics have to be taken into account when setting the parameters of the dynamic protection strategy. Table 9.10 compares typical bid–offer spreads and management fees of various ETF. The shown bid–offer spreads are snapshots and may vary from day to day. For further information, please compare [5]. Bid–offer spreads reflect the characteristics of the underlying and are influenced among other factors by liquidity and taxation issues. Larger bid–offer spreads in the investment would favor the transaction filter with larger , t, and small λ.
9.7. CLOSING REMARKS We have demonstrated how the main drawback of CPPI, namely, its procyclical behavior, the lack of recovering potential once the floor has been hit, and the fixed investment horizon can be overcome by introducing a number of advanced features. The above strategies may be implemented with simple spread sheet models avoiding the usage of options. However, the investor must be aware of the different risk aspects related to this approach. We illustrated how dynamic portfolio strategies empower an investor to replicate fairly complex option profiles including path dependent lookback options. Under certain conditions, the investor may save hedging costs, e.g., the option premium. But the savings come at a price. The investor is left with the risk that the strategy will miss the investment target. Buying an option allows one to lock in the implied volatility at the time of purchase. If the realized volatility over the lifetime is higher than the implied volatility of the option, the option is superior. Buying insurance — in the form
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
Dynamic Portfolio Insurance Without Options
FA
225
of options — is therefore fairly expensive in the current market environment; however, if such a strategy was implemented one and a half years ago, an optionbased strategy back then looks very cheap now in terms of strikes and volatility.
Acknowledgment The author is grateful to Thorsten Weinelt and UniCredit Research to support this work, Peter Hieber for layout and formating, and to David Dakshaw for proof reading the manuscript. Views expressed in this chapter are those of the author and do not necessarily reflect positions of UniCredit Research.
References [1] Bertrand, P and JL Prigent (2001). Portfolio insurance strategies: Obpi versus cppi. CERGY Working Paper, 30. [2] Black, F and R Jones (1987). Simplifying portfolio insurance. The Journal of Portfolio Management, 14(1), 48–51. [3] Black, F and AF Perold (1992). Theory of constant proportion portfolio insurance. Journal of Economic Dynamics and Control, 16, 403–426. [4] Estep, T and M Kritzman (1988). Tipp: Insurance without complexity. The Journal of Portfolio Management, 14(4), 38–42. [5] iShares (2008). [6] Credit Suisse Tremont (2008).
May 12, 2010
17:47
WSPC/SPIB913
b913ch09
This page intentionally left blank
FA
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
10
HOW GOOD ARE PORTFOLIO INSURANCE STRATEGIES?
SVEN BALDER∗ and ANTJE MAHAYNI† Mercator School of Management, University of DuisburgEssen, Lotharstr. 65, 47057 Duisburg, Germany ∗
[email protected] †
[email protected] Portfolio insurance strategies are designed to achieve a minimum level of wealth while at the same time participating in upward moving markets. The most prominent examples of dynamic versions are optionbased strategies with synthetic put and constant proportion portfolio insurance strategies. It is well known that, in a Black/Scholes type model setup, these strategies can be achieved as optimal solution by forcing an exogenously given guarantee into the expected utility maximization problem of an investor with CRRA utility function. The CPPI approach is attained by the introduction of a subsistence level, the OBPI approach stems from an additional constraint on the terminal portfolio value. We bring these results together in order to explain when and why OBPI strategies are better than CPPI strategies and vice versa. We determine the utility losses, which are caused by introducing a terminal guarantee into the unconstrained maximization approach. In addition, we focus on utility losses, which are due to market frictions such as discretetime trading, transaction costs, and borrowing constraints.
227
May 12, 2010
17:47
WSPC/SPIB913
228
b913ch10
FA
Balder and Mahayni
10.1. INTRODUCTION Portfolio strategies which are designed to limit downside risk and at the same time to profit from rising markets are summarized in the class of portfolio insurance strategies. Among others, [6] and [26] define a portfolio insurance trading strategy as a strategy which guarantees a minimum level of wealth at a specified time horizon, but also participates in the potential gains of a reference portfolio. Principally, one can distinguish between two types of portfolio insurance strategies. A risky portfolio (or benchmark index) is combined either with a riskfree asset or with a financial derivative. In particular, the first class includes dynamic versions of optionbased portfolio insurance (OBPI), stoploss strategies, buyandhold strategies and constant proportion portfolio insurance. The second class is mainly characterized by protective put strategies, either in a static or rolling sense. Notice that, with the exception of the buy and hold and the protective put, the above strategies are all dynamic in the sense that they afford portfolio adjustments during the investment horizon. The concept of (synthetic) optionbased portfolio insurance is already introduced in [15] and [33]. The constant proportion portfolio insurance (CPPI) is introduced in [12]. For the evolution of portfolio insurance, we refer to [33]. The popularity of portfolio insurance strategies can be explained by various reasons. On the side of institutional investors, there are regulatory requirements including return guarantees as well as requisitions on the risk profile. For example, [1] considers the problem of an institution optimally managing the market risk of a given exposure by minimizing its ValueatRisk using options. Amongst early papers on the optimality of portfolio insurance are also [7] and [31]. More recently, [22] justifies the existence of guarantees from the point of an investor through behavioral models. In particular, they use cumulative prospect theory as an example, where guarantees can be explained by a different treatment of gains and losses, i.e., losses are weighted more heavily than gains, cf. [30] and [42]. Unfortunately, the justification of guarantees is less clear assuming that the investor’s preferences can be described using the [43] framework of expected utility. Dating back to [37], it is well known that in a Black/Scholes model setup and a constant relative risk aversion (CRRA) utility function, the expected utility maximizing trading rule is a constant mix strategy, i.e., a strategy where a constant fraction of wealth is invested into the risky asset. In this case, an investment weight below one implies that assets are bought when the asset price decreases. This is in sharp contrast to portfolio insurance. In order to honor a terminal guarantee, the asset exposure is to be reduced if the price of the risky asset decreases. Technically, it is straightforward to achieve CPPI and OBPI strategies as the optimal solution of a modified utility maximization problem, which is based on an exogenously given guarantee. CPPI strategies are optimal for an investor, who derives utility from the difference between the terminal strategy value
May 12, 2010
17:47
WSPC/SPIB913
How Good are Portfolio Insurance Strategies?
b913ch10
FA
229
and a given subsistence level. In contrast, the OBPI is optimal for a CRRA investor if one exogenously adds the restriction that the terminal portfolio value is above the floor. These results are well known in the literature. Without postulating completeness, we refer to the works of [5, 6, 13, 16, 17, 21, 24, 26–28, 40, 41]. [20] considers the inverse problem. They analyze if a specific dynamic strategy can be explained by solving the maximization problem of an expected utility maximizing investor, i.e., they analyze if a given investment strategy is consistent with expected utility maximization. In particular, they show that a strategy which implies a pathdependent payoff is not consistent with utility maximization in a Black/Scholestype model. Another strand of the literature analyzes the robustness properties of stylized strategies. Concerning the robustness of option hedges, we refer the reader to [2, 8, 23, 25, 29, 34, 35]. The properties of continuoustime CPPI strategies are also studied extensively in the literature, cf. [13] or [14]. A comparison of OBPI and CPPI (in continuous time) is given in [9]. [44] also compares OBPI and CPPI strategies. In particular, they derive parameter conditions implying second and thirdorder stochastic dominance of the CPPI strategy. The literature also deals with the effects of jump processes, stochastic volatility models and extreme value approaches on the CPPI method, cf. [10, 11]. An analysis of gap risk, i.e., the risk that the guarantee is violated, is provided in [3] and [19]. Gap risk is implied by introducing jumps into the model or by market frictions such as discrete time trading. In practice, the gap risk was already observable during the 1987 crash. In addition, the crash is sometimes even explained or seen to be supported by the portfolio protection mechanisms. However, there are also contradicting opinions, cf. [32]. Finally, there is also a a wide strand of empirical papers, which measure the performance of portfolio insurance strategies. For example, we refer to [18] who give an extensive simulation comparison of popular dynamic strategies of asset allocation. The following paper mitigates between expected utility maximization and the comparison of stylized strategies. We start with an exposition of the three optimization problems which imply constant mix, CPPI, and OBPI strategies as optimal. Instead of giving a further justification for the existence of guarantees, we use the (wellknown) results of the optimization problems to explain the main differences between portfolio insurance mechanisms. Comparing the terminal payoffs shows that both portfolio insurance strategies, CPPI and OBPI, result in payoffs, which consist of a fraction of the payoff of a constant mix strategy (which is optimal for the unconstrained CRRA investor) and an additional term due to the guarantee. The additional term provides an intuitive way to explain the main advantage of the OBPI approach as compared to the CPPI approach. Intuitively, it is clear that the fraction of wealth which is put into the optimal unconstrained strategy is linked to the price of the guarantee, i.e., the fraction is less than one. In the case of the CPPI approach, the additional term is simply the guarantee itself, i.e., the payoff of an adequate number of zero bonds. In contrast, the
May 12, 2010
230
17:47
WSPC/SPIB913
b913ch10
FA
Balder and Mahayni
additional term implied by the OBPI is a put option where the (synthetic) underlying is given by a fraction of the constant mix strategy and the strike is equal to the guarantee. Obviously, the put is cheaper than the zero bonds. Therefore, an investor who follows the OBPI approach puts a larger fraction of his wealth into the unconstrained optimal portfolio than an investor who follows the CPPI approach. To asses the utility costs of forcing a guarantee into the unconstrained problem, i.e., the utility costs from having to use a suboptimal strategy, we compare the certainty equivalents of the different strategies and calculate the loss rates. In addition, we explain one major drawback of the OBPI method, which is due to the kink in the payoffprofile caused by the option component. The terminal value of the OBPI is equal to the guarantee if the put expires in the money. In contrast to the CPPI method, this implies a positive point mass for the event that the terminal value is equal to the guarantee. The probability is given by the real world probability that the terminal asset price is below the strike of the put. Intuitively, it is clear that this can cause a high exposure to gap risk, i.e., the risk that the guarantee is violated, if market frictions are introduced. We illustrate this effect by taking trading restrictions and transaction costs into account. It turns out that the guarantee implied by the CPPI method is relatively robust. In contrast, the probability that the guarantee is not reached under the corresponding synthetic discretetime OBPI strategy is rather high. The outline of the paper is as follows. In Sec. 10.2, we review the wellknown optimization problems yielding constant mix, CPPI, and OBPI strategies as optimal solutions. We compare the optimal strategies and resulting payoffs, and we discuss some advantages (disadvantages) of the different portfolio insurance methods. In Sec. 10.3, we consider the utility losses caused by the introduction of strictly positive terminal guarantees for a CRRA investor. In particular, we compare CPPI and OBPI strategies according to their implied loss rate. We consider the effects of market frictions in Sec. 10.4 where we focus on the loss rates, which are implied by discretetime trading and transaction costs. In addition, we compare the effects of these market frictions on the protection mechanisms of CPPI and OBPI. In Sec. 10.5, we address the topic of borrowing constraints and consider the capped version of CPPI strategies. Section 10.6 concludes the chapter.
10.2. OPTIMAL PORTFOLIO SELECTION WITH FINITE HORIZONS All stochastic processes are defined on a stochastic basis (, F, F, P), which satisfies the usual hypotheses. We consider two assets. The riskless zero bond B with maturity T grows at a constant interest rate r, i.e., dBt = Bt rdt where BT = 1. The evolution of the risky asset S, a stock or benchmark index, is given by a geometric Brownian
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
231
How Good are Portfolio Insurance Strategies?
motion dSt = St (µdt + σdWt ),
S0 = s,
(10.1)
where W = (Wt )0≤t≤T denotes a standard Brownian motion with respect to the real world measure P. µ and σ are constants and we assume that µ > r ≥ 0 and σ > 0. In particular, dSt = St (rdt + σdWt∗ ),
S0 = s,
(10.2)
where W ∗ = (Wt∗ )0≤t≤T denotes a standard Brownian motion with respect to the equivalent martingale measure Q, i.e., µ−r dQ 1 µ−r 2 = e− 2 ( σ ) t− σ Wt . (10.3) dP t A continuoustime investment strategy or saving plan for the interval [0, T ] can be represented by a predictable process (πt )0≤t≤T . πt denotes the proportion of the portfolio value at time t, which is invested in the risky asset S. In the following, we also refer to πt as the portfolio weight at time t. W.l.o.g., we consider strategies which are selffinancing, i.e., money is neither injected nor withdrawn during the investment horizon [0, T ]. Thus, the fraction of wealth which is invested at time t in the zero bond B is given by 1 − πt . Let V = (Vt )0≤t≤T denote the portfolio value process associated with the strategy π, then the dynamics of V are given by dSt dBt where V0 = x. + (1 − πt ) (10.4) dVt (π) = Vt πt St Bt For the above model assumptions, it follows dVt (π) = Vt [(πt (µ − r) + r)dt + πt σdWt ]
where V0 = x.
(10.5)
Alternatively, the strategies can be represented by the number of shares. Let ϕt = (ϕt,S , ϕt,B )0≤t≤T where ϕt,S denotes the number of risky assets and ϕt,B the number of zero bonds with maturity T , which are held at time t. In particular, we have ϕt,S =
πt Vt St
and
ϕt,B =
(1 − πt )Vt , Bt
(10.6)
where Bt denotes the tprice of the zero bond maturing at T . Traditionally, a strategy specification via the portfolio weights is used in the context of portfolio optimization while the convention of stating the number of shares is normally preferred in the context of hedging. In the case of a finite investment horizon T and no intermediate consumption possibilities, the relevant optimization problem is given by sup EP [u(VT (π))] subject to Eq. (10.5),
π∈
(10.7)
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
232
FA
Balder and Mahayni
where denotes the set of all selffinancing trading strategies. The utility function u (u ∈ C2 ) is assumed to be strictly increasing and concave, i.e., u > 0 and u < 0. In the following, we recall the wellknown optimization problems, which justify three basic strategy classes: constant mix (CM) strategies, constant proportion portfolio insurance (CPPI) strategies, and optionbased portfolio insurance (OBPI) strategies. In contrast to a CM strategy, which is exclusively specified by a constant portfolio weight m, i.e., πtCM = m, portfolio insurance strategies incorporate a guarantee component which, in the simplest case, is given by an amount GT , which is to be honored at the end of the investment horizon T (T > 0). While a CPPI strategy is valuebased in the sense that the portfolio weights are exclusively specified by the current portfolio value (and the present value of the guarantee), the (dynamic) OBPI approach is payoff and model dependent. Here, the investment decisions are, in a complete model, given in terms of the delta hedge of an option payoff. Formally, the three strategy classes are CM = {π ∈ πt = m, m ≥ 0} Vt − e−r(T −t) GT CPPIG = π ∈ πt = m ,m ≥ 0 Vt t St , OBPIG = π ∈ πt = Vt ∂ t = EQ [e−r(T −t) (h(ST ) − GT )+ Ft ] . ∂St
(10.8) (10.9)
(10.10)
Notice that for GT = 0, we have CM = CPPI G . However, we refer to CPPI and OBPI versions, where the guarantee is not a strategy parameter but GT > 0 is exogenously given. In particular, we assume that GT < V0 erT . For the OBPI approach, h is a functional which has to reflect the initial budget constraint. In practice, h is typically a linear function of ST and the resulting payoff equals a protective put strategy, where an initial investment in risky assets is protected by corresponding puts. In general, h can be an arbitrary function, e.g., a power function. In [39], a general discussion of how to choose h optimally with respect to different utility functions is given. Table 10.1 summarizes the optimization problems, which are suited to justify the three strategy classes. Problem (A) is derived from the classic Merton problem, Table 10.1 Problem (A) (B) (C)
Benchmark Optimization Problems.
Utility function (γ > 0, γ = 1) 1−γ
uA (VT ) = VT /(1 − γ) uB (VT ) = (VT − GT )1−γ /(1 − γ) 1−γ uA (VT ) = VT /(1 − γ)
Additional constraint
Optimal strategy
None None VT ≥ GT
CM CPPI OBPI
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
How Good are Portfolio Insurance Strategies?
233
cf. [37]. While Merton additionally considers optimal consumption, the optimization problem (A) is limited to the terminal wealth. Nevertheless, due to the separability of the investment and the consumption decision, the optimal investment strategy is equal. Problem (B) introduces a subsistence level GT such that uB belongs to the class of HARA utility functions. Problem (C) consists of the CRRA utility function uA which is also used in problem (A) but poses an additional constraint on the terminal value of the strategy, i.e., the constraint that the terminal strategy value must be above or equal to the terminal guarantee GT . The solutions of the optimization problems and their corresponding proofs are well known in the literature such that we omit some technical parts of the proofs and refer to the literature given in the introduction. Intuitively, it is clear that the solutions of problems (B) and (C) are modifications of the classic Merton problem (A), where a guarantee is exogenously forced into the optimization problem, respectively the solution. Basically, the subsistence level in (B) results in the optimization problem of (A) if the value process is reduced by the present value of the terminal guarantee, i.e., the optimization problem (A) is given in terms of the cushion process. Technically, the solution of problem (C) is more involved. However, the solution of (C) is intuitive in the sense that the constraint on the terminal value features an European option on the optimal payoff of (A), where the initial investment must take into account the price of the option.
10.2.1. Problem (A) Two observations simplify the optimization problem (A) to a large extend. (i), the constant relative risk aversion implies that the attitude toward financial risk is independent of the initial wealth level. (ii), the problem is independent of the current asset price St . This follows from the stationarity of the increments of St . From (i), one can conclude that the optimal terminal value VT (π∗ ) is linear with respect to the initial wealth. Let Vˆ 1 (t, T ) denote the optimal terminal value at T for current wealth Vt = 1. Consider now two dates t1 < t2 . Then, EP [uA (Vt2 (π∗ ))] = EP [EP [uA (Vt2 (π∗ ))Vt1 (π∗ )]] = EP [EP [uA (Vt1 (π∗ ) · Vˆ 1 (t1 , t2 ))]Vt1 (π∗ )] = EP [uA (Vt1 (π∗ ))]EP [(Vˆ 1 (t1 , t2 ))1−γ ], i.e., the optimal portfolio weight in t < t1 must be equal for the two investment horizons. This implies that the solution of problem (A) can be obtained by restricting the strategy set to constant mix (CM) strategies such that it is enough to consider the
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
234
FA
Balder and Mahayni
maximization problem sup EP [uA (VT (π))] subject to Eq. (10.5).
π∈CM
Notice that for π ∈ CM , Eq. (10.5) simplifies to dVtCM = VtCM [(r + m(µ − r))dt + mσdWt ] r+m(µ−r)− 12 m2 σ 2
VTCM = V0CM e(
i.e., Inserting σWT = ln
ST S0
VTCM
)T +mσWT .
(10.11) (10.12)
− (µ − 21 σ 2 )T gives
=
1 2 2 1 2 V0CM e(m(µ−r)+r− 2 m σ )T −m(µ− 2 σ )T
ST S0
m
= φ(V0CM , m)STm where
1 φ(x, m) := x S0
(10.13) m
1 2 e(1−m)(r+ 2 mσ )T .
(10.14)
Notice that, as a function of the terminal asset price ST , the payoff VTCM is concave for m < 1, linear for m = 1 and convex for m > 1. The expected utility is equal to EP [uA (VTCM )] = =
φ(V0CM , m)1−γ (1−γ)m ] EP [ST 1−γ (V0CM )1−γ (1−γ)(r+m(µ−r)− 1 γm2 σ 2 )T 2 e . 1−γ
(10.15)
Finally, it is straightforward to show that argmaxm EP [uA (VTCM )] =
µ−r =: m∗ . γσ 2
(10.16)
10.2.2. Problem (B) Consider now the (modified) portfolio planning problem of an investor, who derives the utility from the difference between the portfolio value and a given subsistence level. Let C = (Ct )0≤t≤T denote the cushion process, where Ct := Vt − e−r(T −t) GT . If a constant proportion m of the cushion is invested in the risky asset, one obtains analogously to the Eqs. (10.11) and (10.13) dCt = Ct [(r + m(µ − r))dt + mσdWt ], C0 = V0 − e−rT GT
(10.17)
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
How Good are Portfolio Insurance Strategies?
235
and CT = φ(C0 , m)STm , as well as argmaxm EP [uA (CT )] = argmaxm EP [uA (φ(C0 , m)STm )] =
µ−r = m∗ . γσ 2
Notice that EP [uA (CT )] = EP [uB (VT )]. Using the fact that the optimal payoff does not depend on the asset price path (cf. for example [20]) implies that the optimal ∗ investment proportion πt,B of problem (B) is ∗ = πt,B
m∗ Ct m∗ (Vt − e−r(T −t) GT ) = . Vt Vt
(10.18)
Obviously, problem (A) and (B) coincide in the case that GT = 0. However, GT > 0 implies a reduction in the investment proportion, i.e., ∗ ∗ ≤ m∗ = πt,A . πt,B
(10.19)
The cushion dynamics given by Eq. (10.17) immediately implies Vt = e−r(T −t) GT + φ(C0 , m; t)Stm , i.e., Vt ≥ e−r(T −t) GT .
10.2.3. Problem (C) The additional constraint VT ≥ GT together with the pathindependency of the optimal ∗ can be represented as follows solution implies that the optimal payoff VT,C ∗ = max{h(ST ), GT } VT,C
= h(ST ) + [GT − h(ST )]+ = GT + [h(ST ) − GT ]+ .
(10.20)
Thus, the terminal value of the strategy can be interpreted in terms of an option on the payoff h such that the optimal strategy π∗ is given in terms of the delta hedge, i.e., π ∗ ∈ OBPI . In the special case that GT = 0, the optimization problem reduces to the ∗ classic Merton case, i.e., it holds h(ST ; GT = 0) = φ(V0 , m∗ )STm . For GT > 0, the optimal solution affords a reduction of the initial investment from V0 to V˜ 0 (V˜ 0 < V0 ) ∗ such that h(ST ) = φ(V˜ 0 , m∗ )STm . The remaining money V0 − V˜ 0 is then used to buy the put with payoff [GT − h(ST )]+ . A proof for the optimality of the payoff
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
236
FA
Balder and Mahayni
∗ ∗ VT,C = GT + [φ(V˜ 0 , m∗ )STm − GT ]+ is given in [24]. The concavity of uA (x) = (x1−γ )/(1 − γ) implies that for any payoff Vˆ T with Vˆ T ≥ GT
∗ ∗ ∗ uA (Vˆ T ) − uA (VT,C ) ≤ uA (VT,C )(Vˆ T − VT,C ) ∗ ∗ m∗ ∗ −γ −γm ˆ ˜ φ(V0 , m ) ST (VT − VT,C ) for ST >
=
−γ GT (Vˆ T − GT )
∗
for STm
GT ˜ φ(V0 , m∗ ) , GT ≤ φ(V˜ 0 , m∗ )
i.e., −γm∗
∗ uA (Vˆ T ) − uA (VT,C ) ≤ φ(V˜ 0 , m∗ )−γ ST
∗ (Vˆ T − VT,C ) ∗
−γm −γ − GT )+ (Vˆ T − GT ). − (φ(V˜ 0 , m∗ )−γ ST
(10.21)
∗ Now, consider a portfolio VT (ε) = εVT,A + (1 − ε)V˜ T , where V˜ T is the terminal wealth of any other strategy with the same initial wealth. The firstorder condition of optimization problem (A) implies ∂EP [u(VT (ε))] ∗ = EP [u (VT (ε))(VT,A − V˜T )]ε=1 ∂ε ε=1 −γm∗
= EP [φ(V0 , m∗ )−γ ST
∗ (VT,A − V˜ T )] = 0.
∗ Therefore, adding and subtracting the optimal unconstrained solution VT,A to the first term on the right hand side of inequality (10.21) and taking expectations gives −γm∗
φ(V˜ 0 , m∗ )−γ (EP [ST
−γm∗
∗ (Vˆ T − VT,A )] + EP [ST
∗ ∗ (VT,A − VT,C )]) = 0.
∗ ∗ Together with Vˆ T ≥ GT a.s. it follows EP [uA (Vˆ T ) − uA (VT,C )] ≤ 0 such that VT,C is indeed the optimal solution w.r.t. problem (C). ∗ ∗ = GT + [φ(V˜ 0 , m∗ )STm − GT ]+ can be replicated by a selfThe payoff VT,C financing strategy where the initial investment can be represented by the expected discounted payoff under the uniquely defined equivalent martingale measure Q, i.e., ∗ V0 = e−rT EQ [GT + (φ(V˜ 0 , m∗ )STm − GT )+ ]
∗ −rT ∗ = e GT + φ(V˜ 0 , m )EQ e−rT STm −
GT φ(V˜ 0 , m∗ )
+ .
V˜ 0 has to be determined such that the initial cushion C0 = V0 −e−rT GT exactly finances φ(V˜ 0 , m∗ ) power call options with power p = m∗ and strike K = (GT )/(φ(V˜ 0 , m∗ )),
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
237
How Good are Portfolio Insurance Strategies?
i.e., V0 − e−rT GT = φ(V˜ 0 , m∗ )PO 0, S0 ; m∗ ,
G
(10.22)
φ(V˜ 0 , m∗ )
where PO(t, St ; p, K) denotes the tprice of a power call with power p, strike K, and maturity T . The pricing formula for power calls, cf. [45] or [36], is given by PO(t, St ; p, K) := e−r(T −t) EQ [(ST − K)+ Ft ] p 2 − (1−p)σ (T −t) 2 −r(T −t) St e = e e−r(T −t) p
√ St × N h1 t, √ T − t − (1 − p)σ p K St − KN h2 t, √ . p K
(10.23)
N denotes the onedimensional standard normal distribution function. The functions h1 and h2 are given by h1 (t, z) =
ln z + (r + 12 σ 2 )(T − t) ; √ σ T −t
√ h2 (t, z) = h1 (t, z) − σ T − t.
(10.24)
Recall that V˜ 0 ≤ V0 . A lower bound on V˜ 0 follows with [STm − GT /φ(V˜ 0 , m∗ )]+ ≤ ∗ STm , i.e., ∗
φ(V˜ 0 , m∗ )PO 0, S0 ; m∗ ,
G φ(V˜ 0
, m∗ )
∗
≤ φ(V˜ 0 , m∗ )EQ [e−rT STm ] = V˜ 0
and Eq. (10.22). This implies V0 − e−rT GT ≤ V˜ 0 ≤ V0 .
(10.25)
∗ Consider now the optimal portfolio weight πt,C of the dynamic strategy, i.e.,
∗ πt,C
=
T St φ(V˜ 0 , m∗ )PO t, St ; m∗ , φ(V˜G,m ∗) 0
Vt
T φ(V˜ 0 , m∗ )PO t, St ; m∗ , φ(V˜G,m St ∗) 0 = , e−r(T −t) GT + φ(V˜ 0 , m∗ )PO t, St ; m∗ , φ(V˜ G,m∗ ) 0
(10.26)
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
238
Balder and Mahayni
where PO :=
∂PO (t, St ; p, K). ∂St
FA
PO
Differentiating (10.23) immediately gives
σ2
St e 2 p(T −t) (t, St ; p, K) = p e−r(T −t)
p−1
√ St N h1 t, √ T − t . − (1 − p)σ p K (10.27)
Finally, we compare the optimal investment weight implied by problem (C) with the one of the unconstrained solution given by m∗ . Notice that
∗
STm −
GT φ(V˜ 0 , m∗ )
+
∗
≥ STm −
GT φ(V˜ 0 , m∗ )
implies that the denominator of the right hand side of Eq. (10.26) is larger than 1 ∗ 2 ∗ ∗ ∗ φ(V˜ 0 , m∗ )EQ [e−r(T −t) STm Ft ] = φ(V˜ 0 , m∗ )Stm e(m −1)(r+ 2 m σ )(T −t)
such that ∗ ≤ PO t, St ; m∗ , πt,C
∗ GT 1 ∗ 2 ∗ St(1−m ) e(1−m )(r+ 2 m σ )(T −t) . φ(V˜ 0 , m∗ )
In addition, Eq. (10.27) immediately gives PO
GT t, St ; m , ˜ φ(V0 , m∗ ) ∗
≤m
∗
St e−r(T −t)
m∗ −1
e− 2 m 1
∗
(1−m∗ )σ 2 (T −t)
.
Together, we have ∗ πt,C ≤ m∗ .
(10.28)
Analogously to the problem (B), the terminal constraint in problem (C) also gives rise to a reduction of the optimal unconstrained portfolio weight m∗ .
10.2.4. Comparison of Optimal Solutions Recall that the optimal terminal payoffs VT∗ do not depend on the asset price path, but can be specified as a function of the terminal asset price ST . The optimal payoffs for
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
239
How Good are Portfolio Insurance Strategies?
the optimization problems (A), (B), and (C) are summarized as follows ∗
∗ = φ(V0 , m∗ )STm VT,A
(10.29) ∗
∗ VT,B = GT + φ(V0 − e−rT GT , m∗ )STm
V0 − e−rT GT ∗ VT,A V0 + ∗ = GT + φ(V˜ 0 , m∗ )STm − GT = GT +
∗ VT,C
∗ ∗ = φ(V˜ 0 , m∗ )STm + [GT − φ(V˜ 0 , m∗ )STm ]+ +
V˜ 0 ∗ V˜ 0 ∗ = V + GT − VT,A . V0 T,A V0
(10.30) (10.31)
(10.32)
∗ corresponds to φ(V0 , m∗ ) power claims with power m∗ , where the The payoff VT,A ∗ number φ(V0 , m ) depends on the initial investment and the optimal investment weight m∗ . The optimization problem (B) introduces a subsistence level which implies that the number of power claims with power m∗ must be reduced to afford the riskfree investment, which is necessary to honor the guarantee. In consequence, the portfolio weight is lower than in the case of problem (A), cf. Inequality (10.19). The link between the solutions of (A) and (B) is even more explicit if one considers the number of shares in the asset which are held. Notice that the cushion dynamics, cf. Eq. (10.17) implies that CtCPPI /C0CPPI = VtCM /V0 if the multiplier m of the CPPI is equal to the portfolio weight of the CM strategy. In particular, this implies that the value of the cushion is proportional to the value of the CM strategy, i.e., CtCPPI = (C0 /V0 )VtCM . This is also true for the number of assets ϕt,S , i.e., CPPI ϕt,S =
C0 CM ϕ V0 t,S
CPPI = GT + ϕt,B
C0 CM ϕ , V0 t,B
(10.33) (10.34)
where ϕt,B denotes the number of zero bonds with maturity T . In particular, it holds that C0 CM V . (10.35) V0 t The CPPI strategy can thus be interpreted as a buyandhold strategy of a constant mix strategy with an additional investment into GT zero bonds, cf. also Eq. (10.31). A similar reasoning applies to the solution of (C), which can be interpreted as a buyandhold strategy of a constant mix strategy with an additional investment into a put with strike GT , cf. Eq. (10.32). Obviously, the put is worth less than GT zero bonds such that one can buy and hold more CM strategies in the case of the option based approach, VtCPPI = e−r(T −t) GT +
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
240
FA
Balder and Mahayni
i.e., V˜ 0 /V0 ≥ C0 /V0 . Intuitively, it is thus clear that the OBPI approach gives a better result than the CPPI approach with respect to a utility function, which favors the CM strategy with portfolio weight m∗ . ∗ which honors the guarantee GT and In general, a modification of the payoff VT,A with t0 price equal to V0 can be represented by
+ + G G T T ∗ ∗ = V0 . −β subject to EQ e−rT GT + α VT,A −β GT + α VT,A α α (10.36) The CPPI approach corresponds to β = 0 and gives a smooth payoffprofile. β = 1 results in the OBPI approach with a kinked payoffprofile. As a consequence of the ∗ ∗ ∗ guarantee GT , both payoffs VT,B and VT,C are higher (lower) than VT,A for low (high) ∗ terminal asset prices. However, the smooth solution VT,B implies that the intersec∗ ∗ occurs at a higher asset price ST if compared to the intersection of VT,C tion with VT,A ∗ and VT,A . Let si,j (i = j, i, j ∈ {A, B, C}) denote the terminal asset price ST such that ∗ ∗ . Equation (10.30) immediately gives VT,i = VT,j ∗ ∗ = VT,B VT,A
⇔
∗ VT,A = V0 erT
such that 1 2 sA,B = S0 e(r+ 2 (m−1)σ )T .
(10.37)
With Eq. (10.32), V˜ 0 ≤ V0 and V0 ≥ e−rT GT it follows 1
1
sA,C = S0 e m (g−r)T e(r+ 2 (m−1)σ
2
)T
1
= e m (g−r)T sA,B
(10.38)
where g := T −1 (ln GT − ln V0 ) ≤ r. Finally, one obtains 1 1 1 2 sB,C = S0 e m (ν−r)T e(r+ 2 (m−1)σ )T = e m (ν−r)T sA,B
(10.39)
where ν := T −1 (ln GT − ln(V˜ 0 − C0 )) ≥ r. Therefore sA,C ≤ sA,B ≤ sB,C . An illustration of the payoffs and the intersection points is given in Fig. 10.1. Table 10.2 Model paramter S0 = 1 σ = 0.15 r = 0.03 µ = 0.085
Basic Parameter Constellation. Strategy parameter V0 = 1 T = 10 γ = 1.2 m = m∗ = 2.037
Terminal guarantee GT = 1
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
241
How Good are Portfolio Insurance Strategies?
5
Payoff
4
3
2
1
0 0.0
0.5
1.0
1.5
2.0
2.5
3.0
Terminal asset price ∗ (solid line), V ∗ (dotted line), and V ∗ (dashed line), Figure 10.1 Optimal payoffs VT,A T,B T,C where the parameters are given as in Table 10.2.
We end this section by emphasizing one important consequence for the two protection mechanisms implied by the smooth and the kinked solutions, i.e., implied by the assumptions that marginal utility jumps gradually or discontinuously to infinity. ∗ Notice that the smooth payoff VT,B implies that there is no probability mass on the ∗ = GT ) = 0. In event that the terminal value is equal to the guarantee GT , i.e., P(VT,B ∗ contrast, for the kinked payoff VT,C , it holds GT ∗ m∗ P(VT,C = GT ) = P ST ≤ φ(V˜ 0 , m∗ ) ST = P ln ≤ ln S0
m∗
GT φ(V˜ 0 ,m∗ )
S0
.
Using the definition of φ, cf. Eq. (10.14), and m∗ = (µ − r)/(γσ 2 ) yields 2 ˜ ln e−rTV0GT − 12 γλ T √ ∗ + λ T, P(VT,C = GT ) = 1 − N √ λ T γ
(10.40)
(10.41)
where λ := (µ − r)/σ. Intuitively, it is clear that a positive point mass on the event {VT = GT } might indicate that the corresponding strategy is more sensitive to the introduction of gap risk, which is caused by asset price jumps. This problem is considered in Sec. 10.4,
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
242
FA
Balder and Mahayni
where trading restrictions in the sense of discrete time trading and transaction costs are introduced.
10.3. UTILITY LOSS CAUSED BY GUARANTEES 10.3.1. Justification of Guarantees and Empirical Observations There are a few comments necessary concerning the justification of guarantees. The optimization problems (B) and (C) are already based on an exogenously postulated guarantee such that one might doubt their capacity to give a meaningful justification of guarantees. However, there are some arguments, which are in favor of a subsistence level. Similar reasonings are true with respect to optimization problem (C). In consequence, to some extent, guarantees can be explained with respect to the assumption that the investor’s preferences can be described using the [43] framework of expected utility. More recently, [22] justifies the existence of guarantees through behavioral models. In particular, they use cumulative prospect theory as an example. In the following, we do not give further justifications for the existence of guarantees or the popularity of portfolio insurance strategies but take them as given. However, we think in terms of utility losses caused by guarantees and compare the smooth and kinked payoff solutions. One possibility which is consistent with empirical observations is given by measuring the utility losses of guarantees with respect to a CRRA utility function, where the parameter of risk aversion is assumed to be above 1 (γ > 1). For the validity of CRRA utility functions and the parameter of risk aversion, we refer to [38] and the literature given herein.
10.3.2. Utility Loss The performance of the strategy π can be measured by its associated expected utility, which can in turn be described by the certainty equivalent. It is defined as the certain amount, which makes the investor indifferent between achieving this certain amount (at T ) or using the strategy π, i.e., the time T certainty equivalent CET of the strategy π is defined by u(CET (π)) = EP [u(VT (π))].
(10.42)
Consider for example the utility function uA (x) = x1−γ /(1 − γ) and a CM strategy with optimal investment proportion m∗ , i.e., (1−γ)m∗
CE T,A (πCM ) = φ(V0 , m∗ )(EP [ST
1
]) 1−γ .
(10.43)
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
243
How Good are Portfolio Insurance Strategies? p
p
With EP [ST ] = S0 exp{(pµ − 12 p(1 − p)σ 2 )T } it follows ∗
CET,A (πCM ) = φ(V0 , m∗ )S0m e(m ∗
= V0 e(r+m
∗
µ− 12 γσ 2 m∗ (1−m∗ (1−γ)))T
(µ−r)− 12 γ(m∗ )2 σ 2 )T
1
= V0 erT e 2γ (
µ−r 2 ) T σ
(10.44)
.
Analogously, one obtains CE T,B (πCPPI ) = GT + (V0 − e−rT GT )e(r+m 1
∗
= CE T,A (πCM ) + GT (1 − e 2γ (
(µ−r)− 12 γ(m∗ )2 σ 2 )T µ−r 2 ) T σ
(10.45)
).
Thus, the certainty equivalent is lower for an investor with subsistence level than an investor without. We can also calculate the utility loss of an investor, who follows a suboptimal strategy. It is described by the loss rate lT,i (π) of the strategy π and the utility function i (i ∈ {A, B, C}) where lT,i (π) :=
ln
∗ CET,i CET,i (π)
T
(10.46)
.
∗ CE ∗T,i denotes the certainty equivalent of the optimal strategy πi∗ = (πt,i )0≤t≤T while CE T,i (π) the one of the suboptimal strategy π = (πt )0≤t≤T . The loss rates with respect to the utility function uA (x) = x1−γ /(1 − γ) are summarized in Table 10.3 for CM, CPPI, and OBPI strategies with strategy parameter m. Notice that a loss rate, which is higher than the one of a riskfree investment, i.e., πtCM = 0, implies that the associated strategy is prohibitively bad. This critical loss rate is equal to 12 γ(σm∗ )2 , cf. Table 10.3. For m = m∗ , it is obvious that lT,A (πOBPI ) < lT,A (πCPPI ) because πCPPI is suboptimal w.r.t. the optimization problem (C) where
Table 10.3
Loss Rates with Respect to uA (x) = x 1−γ /(1 − γ ).
Strategy π
Loss rate lT,A (π) 2 1 2 ∗ γσ m − m 2
πtCM = m Vt − e−r(T −t) GT Vt T PO t, St ; m, G ˜
∗
EP [(φ(V0 , m∗ )STm )1−γ ] 1 ln (1 − γ)T EP [(GT + φ(V0 − e−rT GT , m)STm )1−γ ]
πtCPPI = m
πtOBPI =
φ(V0 ,m)
Vt
St
∗
EP [(φ(V0 , m∗ )STm )1−γ ] 1 ln (1 − γ)T EP [(GT + [φ(V˜ 0 , m)STm − GT ]+ )1−γ ]
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
Balder and Mahayni
0.06
0.06
0.05
0.05
0.04
0.04
Lossrate
Lossrate
244
FA
0.03
0.03
0.02
0.02
0.01
0.01
0.00
2.04
4
m
6
0.00
2.04
4
6
m
Figure 10.2 Loss rates w.r.t. u = uA for CPPI (solid lines), OBPI (dashed) and CM (dotted) strategies with varying parameter m. The parameters are given in Table 10.2. The investment horizon is T = 10 years (left graph) and T = 20 years (right graph).
uC = uA and πOBPI is optimal. It holds lT,A (πCP ) = 0 < lT,A (πOBPI ) < lT,A (πCPPI ). The utility loss implied by the guarantee is higher for the CPPI than for of the OBPI. This is illustrated in Fig. 10.2. In addition, notice that the loss rates of the CM strategies are symmetric in the sense that for > 0, a strategy parameter m = m∗ + implies the same loss rate as the parameter m = m∗ − . In contrast, OBPI strategies yield a lower loss rates in the case of m = m∗ + than for m = m∗ − . Intuitively, this is clear since the protection feature implies that the portfolio weights of the portfolio insurance strategies are too low compared to the optimal investment proportion m∗ , cf. Inequalities (10.19) and (10.28). For the comparison of CPPI and OBPI, it is important to keep in mind that m∗ is the optimal OBPI parameter for u = uA . In contrast, m∗ is not the optimal CPPI parameter for u = uA , but for u = uB which includes a subsistence level. In order to compare the loss rates implied by CPPI and OBPI w.r.t. u = uA , it is thus necessary to consider the maximization problem (GT + φ(V0 − e−rT GT , m)STm )1−γ . max EP [uA (VT (π))] = max EP m π∈CPPI 1−γ (10.47) Figure 10.3 illustrates the loss rates for CPPI and OBPI strategies with varying parameter m. In addition, the minimal loss rates are summarized in Table 10.4, i.e., the loss rates for OBPI strategies with m = m∗ and CPPI for m = m∗∗ , where m∗∗ is the optimal solution for the maximization problem (10.47). Although the OBPI strategy with parameter m = m∗ is, per construction, the uA utility maximizing portfolio insurance
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
245
How Good are Portfolio Insurance Strategies?
0.06
0.06
Lossrate
0.08
Lossrate
0.08
0.04
0.04
0.02
0.02
0.00
0.00 2.04
4
6
8
10
12
14
16
18
1.63
4
6
8
m
10
12
14
16
18
m
Figure 10.3 Loss rates implied by CPPI (solid lines) and OBPI (dashed lines) for varying m, cf. Table 10.3. The risk aversion is γ = 1.2 (left figure) and γ = 1.5 (right figure). The other parameters are given in Table 10.2.
Table 10.4 Minimal Loss Rates (uA Optimal Strategy Parameter m) for Varying T and γ . The Other Parameters are Given in Table 10.2. Strategy
γ\T
1
2
5
10
20
CPPI
1.2
0.040 (11.32)
0.035 (7.83)
0.026 (4.91)
0.018 (3.57)
0.010 (2.73)
OBPI
1.2
0.037 (2.04)
0.031 (2.04)
0.022 (2.04)
0.014 (2.04)
0.007 (2.04)
CPPI
1.5
0.031 (10.60)
0.026 (7.25)
0.019 (4.45)
0.013 (3.16)
0.007 (2.36)
OBPI
1.5
0.028 (1.63)
0.023 (1.63)
0.015 (1.63)
0.009 (1.63)
0.005 (1.63)
CPPI
1.8
0.024 (10.03)
0.020 (6.80)
0.014 (4.10)
0.009 (2.86)
0.005 (2.08)
OBPI
1.8
0.021 (1.34)
0.017 (1.34)
0.011 (1.34)
0.007 (1.34)
0.003 (1.34)
strategy, the additional loss of the CPPI strategy with parameter m = m∗∗ which is measured by the difference of its loss rate and the one of the OBPI is rather low. Notice that the differences remain approximatively equal for varying time horizons and risk aversion parameters.
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
246
FA
Balder and Mahayni
10.4. UTILITY LOSS CAUSED BY TRADING RESTRICTIONS AND TRANSACTION COSTS It is important to notice that in practice the concept of portfolio insurance is impeded by market frictions. The protection mechanism of portfolio insurance implies that the asset exposure has to be reduced when the asset price decreases. A sudden drop in the asset price, where the investor is not able to adjust his portfolio adequately, causes a gap risk, i.e., the risk that the terminal guarantee is not achieved. One illustrative and meaningful approach to capture the gap risk is given by the introduction of trading restrictions in the sense of discretetime trading and transaction costs. Notice that discretetime trading is one possibility to introduce gap risk. In contrast to adding jumps in the dynamics of the risky asset, it also allows to take into account for transaction costs, which are of practical importance if one decides whether to use the CPPI or OBPI protection mechanism. In particular, we consider transaction costs which are proportional to a change in the position of the risky asset. The proportionality factor is denoted by θ. As before, we assume that the present value of the terminal guarantee GT prevailing at T is lower than the initial investment V0 , i.e., V0 > e−rT GT . Let τ n denote a sequence of equidistant refinements of the interval [0, T ], i.e., τ n = n n n < tnn = T }, where tk+1 − tkn = T/n for k = 0, . . . , n − 1. {t0 = 0 < t1n < · · · < tn−1 To simplify the notation, we drop the superscript n and denote the set of trading dates with τ instead of τ n . The restriction that trading is only possible immediately after tk ∈ τ implies that the number of shares held in the risky asset is constant over the intervals [ti , ti+1 ] for i = 0, . . . , n − 1. However, the fractions of wealth which are invested in the risky asset and the zero bond change as underlying prices fluctuate. Thus, it is necessary to consider the number ϕS of shares held in the risky asset and the number ϕB of zero bonds with maturity T , i.e., the tupel ϕ = (ϕS , ϕB ).
10.4.1. DiscreteTime CPPI Along the lines of [3], we consider a discretetime CPPI version ϕτ = (ϕSτ , ϕBτ ) which is for t ∈ ]tk , tk+1 ] and k = 0, . . . , n − 1 defined by mCtτk 1 τ τ τ ϕt,S := max , 0 , ϕt,B := (V τ − ϕt,S Stk ). (10.48) Stk Btk tk Notice that we do not allow for short positions in the risky asset, i.e., the asset exposure is bounded below by zero. However, similar as for the simple (continuoustime) CPPI, the discretetime CPPI does not include short sale restrictions on the riskless asset. The terminal value of the discretetime CPPI is given by
min{s,n}
VTτ
−rT
= GT + (Vt0 − e
GT )
i=1
Sti m − K(m) er(T −min{ts ,T }) , Sti−1
(10.49)
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
247
How Good are Portfolio Insurance Strategies?
where K(m) := (m − 1)er n /m and ts := min{tk ∈ τVtk < e−r(T −tk ) GT }. We set ts = ∞ if the minimum is not attained. In particular, it holds that ts := min{tk ∈ τStk /Stk−1 < K(m)}. For m > 1, the value of the simple CPPI can drop below the floor. Therefore, the discretetime CPPI version introduces a gap risk, i.e., the risk that the guarantee is violated. For a detailed analysis of the gap risk, we refer to [3]. Besides discretetime trading, we take also transaction costs into account. Along the lines of [13], we assume that the transaction costs are financed by a reduction of the asset exposure arising in the case without transaction costs. This can be justified by the argument that the protection feature of the CPPI is based on a prespecified riskfree investment such that the introduction of transaction costs must not change the number of riskfree bonds, which are prescribed by the CPPI method (without transaction costs). Thus, the discretetime CPPI version with transaction costs ϕτ,TA = (ϕSτ,TA , ϕBτ,TA ) is, for t ∈ ]tk , tk+1 ] and k = 0, . . . , n − 1, defined by ! mCtτ,TA 1 τ,TA τ,TA k+ τ ϕt,S := max , 0 , ϕt,B := (V τ,TA − ϕt,S Stk ), Stk Btk tk+ T
(Ctτ,TA := Vtτ,TA − e−r(T −tk ) GT ) denotes the portfolio (cushion) value where Vtτ,TA k+ k+ k+ immediately after tk , i.e., the value net of transaction costs, which are proportional to the asset price Stk . First, consider the portfolio value Vtτ,TA before transaction costs, k := V and for k = 1, . . . , n i.e., Vtτ,TA t0 0 := ϕtτ,TA Stk + ϕtτ,TA Btk Vtτ,TA k k ,S k ,B = max
mCtτ,TA k−1+ Stk−1
! , 0 Stk + (Vtτ,TA − ϕtτ,TA Stk−1 )er(tk −tk−1 ) k ,S k−1+
= m max{Ctτ,TA , 0} k−1+
Stk − er(tk −tk−1 ) + Vtτ,TA er(tk −tk−1 ) . k−1+ Stk−1
(10.50)
Consider now the adjustment to the proportional transaction costs, which are due immediately after the trading dates. Assuming that the transaction costs are also due at t0 and that the asset positions are transferred into a cash position immediately after tn is consistent to the following definitions τ,TA Vtτ,TA := Vt0 − ϕS,t θSt0 = Vt0 − mθCtτ,TA 0+ 0+ 0+
(10.51)
:= Vtτ,TA − ϕtτ,TA − ϕtτ,TA θStk Vtτ,TA k k ,S k+ k+ ,S =
Vtτ,TA k
St − mθ max{Ctτ,TA , 0} − max{Ctτ,TA , 0} k k+ k−1+ S
tk−1
(10.52)
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
248
FA
Balder and Mahayni
Vtτ,TA := Vtτ,TA − ϕtτ,TA θStn n n ,S n+ = Vtτ,TA − mθ max{Ctτ,TA , 0} n n−1+
Stn . Stn−1
(10.53)
With Eq. (10.51), it immediately follows Ct0 + = Ct0 /(1 + θm). Using Eqs. (10.52) and (10.50) implies that Ctk + > 0 (k = 0, . . . , n − 1) and θ < 1/m, it holds Ctk+1 + = Ctk+1
Stk+1 − mθ max{Ctk+1 + , 0} − Ctk + S
(10.54)
tk
St 1 + θ Stk+1 m − 1 rT T n for er n ≤ k+1 − m e Ctk + 1 + θm S 1 + θm Stk tk 1 − θ Stk+1 m − 1 rT otherwise − m e n = Ctk + 1 − θm Stk 1 − θm S St m − 1 rT T Ctk + (1 − θ)m tk+1 − (m − 1)er n e n > k+1 . for Stk m(1 − θ) Stk (10.55) T
For Ctk + ≤ 0, it follows Ctk+1 + = Ctk+1 = er n Ctk + . It is worth mentioning that the < GT } corresponds to the event that the adjusted cushion drops below event {Vtτ,TA n+ zero during the investment horizon. Notice that for Ctk + > 0 {Ctk+1 + < 0} ⇔
Stk+1 m − 1 rT < e n Stk m(1 − θ)
=: Ak+1 .
Since the complementary of the event {∪n−1 k=0 Ak+1 } is given by the event that all asset T price increments are above er n (m − 1)/(m(1 − θ)), it follows with the assumption that the asset price increments are independent and identically distributed that P(Vtτ,TA n+
n St1 m − 1 rT n e < GT ) = 1 − P > St0 m(1 − θ) = 1 − (N (d2TA (θ)))n
(10.56)
where d2TA (θ)
:=
ln
(1−θ)m m−1
+ (µ − r) Tn − 21 σ 2 Tn . σ Tn
(10.57)
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
249
How Good are Portfolio Insurance Strategies?
10.4.2. DiscreteTime OptionBased Strategy According to Eq. (10.26), the (continuoustime) selffinancing and duplicating strategy for the T payoff GT + φ(V˜ 0 , m)[STm − GT /φ(V˜ 0 , m)]+ is given by GT PO ˜ ϕt,S := φ(V0 , m) (10.58) t, St ; m, φ(V˜ 0 , m) ϕt,B := GT +
φ(V˜ 0 , m)PO t, St ; m, φ(VG˜ T,m) − ϕt,S St 0
B(t, T)
,
(10.59)
where PO is defined as in Eq. (10.27). We consider as a discretetime version of an arbitrary continuoustime trading strategy ϕτ = (ϕSτ , ϕBτ ) with respect to the trading dates τ ϕtτ := ϕtk
for t ∈ ]tk , tk+1 ]
and
for all t ∈ [0, T ].
Setting V0 (ϕ; τ) := V0 (ϕ), the value process V(ϕ; τ) which is associated with ϕτ is Vt (ϕ; τ) = ϕtk ,S St + ϕtk ,B e−r(T −t)
for t ∈ ]tk , tk+1 ]
and
0 ≤ k ≤ n − 1.
In general, the discretetime version of a continuoustime strategy is not selffinancing. In particular, there are in or outflows from the portfolio which occur immediately after a trading date tk+1 (k = 0, . . . , n − 1). Formally, the costs of discretization (ϕ; τ) which occur immediately after the trading date tk+1 are defined by ξtdis k+1 ξtdis (ϕ; τ) := Vtk+1 (ϕ) − Vtk+1 (ϕ; τ) k+1
(10.60)
= (ϕtk+1 ,S − ϕtk ,S )Stk+1 + (ϕtk+1 ,B − ϕtk ,B )e−r(T −tk+1 ) . Notice that negative costs refer to inflows while positive costs imply that further money is needed to continue the strategy. Taking proportional transaction costs into account also implies that there are trans(ϕ; τ), which occur immediately after a trading date tk+1 , i.e., action costs ξtTA k+1 ξtTA (ϕ; τ) := ϕt0 ,S θSt0 0 ξtTA (ϕ; τ) = ϕtk+1 ,S − ϕtk ,S θStk+1 k+1
for k = 0, . . . , n − 2
and ξtTA (ϕ; τ) := ϕtn−1 ,S θStn . n Defining the payoff V¯ T (ϕ; τ) according to the assumption that the inflows into the strategy are lent according to the interest rate r and outflows are saved according to r
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
250
FA
Balder and Mahayni
gives
er(T −t0 ) + V¯ T (ϕ; τ) := Vtn (ϕ) − ξtTA 0
n−1 "
(ξtdis (ϕ; τ) k+1
+
ξtTA (ϕ; τ))er(tn −tk+1 ) k+1
.
k=0
(10.61)
10.4.3. Comments on Utility Loss and Shortfall Probability The loss rates w.r.t. u = uA as well as the shortfall probabilities of the above discretetime versions are illustrated in Figs. 10.4 and 10.5. Notice that the introduction of a
Loss rate
Shortfall probability
0.010
0.008
0.006
0.004
0.002
0.000 3
4
m
5
6
7
m
Figure 10.4 Loss rates w.r.t. u = uA (shortfall probabilities) implied by continuoustime CPPI (solid line), monthly CPPI without transaction costs (dashed lines) and monthly CPPI with θ = 0.01 (dotted line) for varying m. The parameter setup is given in Table 10.2.
0.8
Shortfall probability
0.10
Loss rate
0.08
0.06
0.04
0.6
0.4
0.2
0.02
0.0
0.00 2.03704
4
m
6
2.04
4
6
m
Figure 10.5 Loss rates w.r.t. u = uA (shortfall probabilities) in the Merton setup implied by continuoustime OBPI (solid line), monthly OBPI without transaction costs (dashed lines) and monthly OBPI with θ = 0.01 (dotted line) for varying m. The parameter setup is given in Table 10.2.
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
251
How Good are Portfolio Insurance Strategies?
gap risk, i.e., a strictly positive shortfall probability, gives a loss of minus infinity in the case of u = uB and u = uC . Observe, that for both strategies, OBPI and CPPI, the loss which is in the first instance caused by timediscretizing the strategies is rather low. However, there is a huge impact caused by transaction costs. The effect is even more pronounced for the OBPI than for the CPPI method. Intuitively, it is clear that in the limit to continuoustime trading, the transaction costs eat up the cushion, which allows a risky investment. The convergence of the value (cushion) process of the CPPI is for example analyzed in [3]. Recall that loss rates above the one of a riskfree investment are prohibitive. In the case of the basic parameter constellation, this critical value w.r.t. u = uA is equal to 0.5γ(σm)2 = 0.056. In particular, if the strategy parameter m is not chosen in a cautious way, portfolio insurance can get prohibitively bad because of transaction costs. Consider now the shortfall probability. Recall that, in contrast to the continuoustime CPPI strategy, the continuoustime OBPI results in a strictly positive probability that the payoff is not above the terminal guarantee. Thus, it is to be expected that the shortfall probability of the OBPI is more sensitive to discretetime trading and transaction costs than the one of the CPPI, cf. the right graphs presented in Figs. 10.4 and 10.5. This effect is also illustrated in Fig. 10.6, where the distribution function of the discretetime versions of OBPI with m = m∗ and CPPI with m = m∗∗ is plotted. However, to some extend the sensitivity of the OBPI to the gap risk measured by the shortfall probability is also caused by the dicretization scheme. While the transaction costs are financed via the cushion for the CPPI, this is not true in the case of the discretetime OBPI.
1.0
Distribution function
Distribution function
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.0
0.2
1
2
4
Terminal payoff
6
8
0.0
1
2
4
6
8
Terminal payoff
Figure 10.6 Distribution functions of terminal values for OBPI (left) and CPPI (right) with transaction costs (dotted line) and without transaction costs (solid line).
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
252
FA
Balder and Mahayni
10.5. UTILITY LOSS CAUSED BY GUARANTEES AND BORROWING CONSTRAINTS To be practically relevant, it is necessary to take borrowing constraints into account. Notice that the basic CPPI approach and an OBPI strategy, which is based on synthesizing a power option, can result in arbitrarily high investment weights. Regarding the OBPI, it is not clear how borrowing constraints should be incorporated. One possibility is given by setting m = 1, i.e., referring to a standard option instead of a power option. Although this is often done in practice, it will not be considered in the following. Instead we focus on incorporating borrowing constraints into the classic CPPI strategy. This straightforwardly results in the capped CPPI, which we abbreviate with CCP. The capped CPPI strategy ϕCCP = (ϕSCCP , ϕBCCP ) is defined by CCP = ϕt,S
min(ωVtCCP , mCtCCP ) , St
CCP ϕt,B =
CCP VtCCP − ϕt,S St
Bt
,
(10.62)
where CtCCP := VtCCP − e−r(T −t) GT and w (w ≥ 1) denotes the restriction on the investment proportion. In the following, we refer to borrowing constraints in the strict sense, i.e., we assume that ω is set equal to one. It is worth mentioning that the borrowing constraints introduce a pathdependence and the payoff implied by the capped CPPI version cannot be stated as a function of the terminal asset price as it is the case without borrowing constraints. Intuitively, it is to be expected that the path dependence yields an additional utility loss in a Black/Scholes type model setup. In order to calculate the loss rate, we consider the distribution of the terminal value of the CCP strategy. Let f denote the density function of the terminal value of the capped CPPI. Then it holds # E[uA (πCCP )] =
∞
uA (v)f(v)dv GT
# CE T,A (π
CCP
)=
∞
v
1−γ
1 1−γ f(v)dv
GT
such that
r+m+ (µ−r)− 21 γ(m∗ σ)2 )T ( 1 V0 e lT,A (πCCP ) = ln . 1 1−γ $∞ T 1−γ f(v)dv v GT
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
FA
253
How Good are Portfolio Insurance Strategies?
For details on the distribution of the capped CPPI, we refer to [4]. Basically, the distribution can be obtained by considering the process (X)0≤t≤T , which is given by the dynamics dXt = (Xt )dt + dWt , where 1 µ−r σ − 2 mσ (x) = µ − r 1 − σ σ 2
x≤0 and x>0
1 (m − 1)V 0 σ ln mG 0 X0 = 1 (m − 1)C0 ln mσ G0
mC0 ≥ V0 . mC0 < V0
Along the lines of [4], one can show that the value process (VtCCP )0≤t≤T and cushion process CtCCP 0≤t≤T are, for ω = 1, given by
VtCCP
m σXt e = Gt m − 1 1 + 1 eσmXt m−1
Xt ≥ 0 (10.63) Xt < 0
and
CtCCP = Gt
m σXt − 1 Xt ≥ 0 e m−1
1 emσXt m−1
.
Xt < 0
In particular, it holds
P[VtCCP
(m−1)v ln mGt 1 p dv σv σ (m − 1)(v − Gt ) ∈ dv] = ln 1 Gt dv p σm σm(v − Gt )
v≥
m Gt m−1
v
1 (m < 1), additional asset are bought (sold) if the asset price increases. In particular, for m > 1 the resulting payoff is convex in the asset price so that a constant mix strategy can, at least technically, be classified as a portfolio insurance strategy. However, the payoff is floored by zero. In theory, it is straightforward to achieve optimal strategies yielding payoffs with a positive floor. Here, the expected utility is maximized under the additional constraint that the terminal portfolio value must be above a strictly positive terminal guarantee. Alternatively, the floor can be achieved if the utility is measured in terms of the difference of the portfolio value and the guarantee instead of the portfolio wealth itself, i.e., if a utility function with a subsistence level is used. The modified optimization problems help to understand the most prominent approaches of portfolio insurance strategies, i.e., CPPI and OBPI strategies. The modifications which are imposed on the unconstrained optimization problem give interesting modifications for the payoffs. Using the unrestricted optimization problem as a benchmark, the constant mix strategy and its associated payoff with floor zero defines
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
How Good are Portfolio Insurance Strategies?
FA
255
also a benchmark for OBPI and CPPI strategies. The CPPI results from a subsistence level, i.e., the utility is measured in terms of the difference of portfolio value and guarantee instead of the portfolio value itself. In contrast, the OBPI results from the additional constraint that the terminal payoff is above the guarantee or floor. Considering the associated payoffs, both approaches result in payoffs which consist of a fraction of the payoff of the constant mix strategy and an additional term stemming from the guarantee. Intuitively, it is clear that the fraction is linked to the price of the guarantee, i.e., the fraction is less than one. The main difference between OBPI and CPPI can easily be explained by the additional term. In the case of the CPPI approach, the additional term is simply the guarantee itself, i.e., the payoff of the adequate number of zero bonds. In contrast, the additional term implied by the OBPI is based on a put on the fraction of the constant mix payoff with strike equal to the guarantee. Obviously, the put is cheaper than the zero bond itself. Thus, the OBPI fraction which is held of the optimal unrestricted payoff is higher than in the case of the CPPI method. This is a major advantage in terms of the associated utility costs, i.e., the loss in expected utility which is caused by the introduction of a strictly positive guarantee. The utility costs are measured and illustrated in terms of a loss rate linking the certainty equivalents of the strict portfolio insurance strategies to the certainty equivalent of the optimal solution. One major drawback of the OBPI method is due to its kinked payoffprofile. The terminal value of the OBPI is equal to the guarantee if the put expires in the money. In contrast to the CPPI method, this implies a positive point mass that the terminal value is equal to the guarantee. This relevant probability is given by the real world probability that the terminal asset prices are below the strike of the put. Intuitively, it is clear that this can cause a high exposure to gap risk, i.e., the risk that the guarantee is violated, if market frictions are introduced. We illustrate this effect by taking trading restrictions and transaction costs into account. It turns out that the guarantee implied by the CPPI method is relatively robust. However, the probability that the guarantee is not reached under the corresponding synthetic discretetime OBPI strategy is rather high. Finally, we tackle the question of borrowing constraints. In a strict sense, borrowing constraints imply that the proportion of asset must not be above one. In the case of the CPPI method, the capped CPPI which simply states that the asset proportion is adequately capped according to the borrowing constraint is of high practical relevance. We give the distribution of the capped CPPI and illustrate the corresponding loss rate, i.e., the loss which is due to borrowing constraints.
References [1] Ahn, DH, J Boudoukh, M Richardson and RF Whitelaw (1999). Optimal risk management using options. Journal of Finance, 54(1), 359–375. [2] Avellaneda, M, A Levy and A Parás (1995). Pricing and hedging derivative securities in markets with uncertain volatilities. Applied Mathematical Finance, 2(2), 73–88.
May 12, 2010
256
17:47
WSPC/SPIB913
b913ch10
FA
Balder and Mahayni
[3] Balder, S, M Brandl andA Mahayni (2009). Effectiveness of CPPI strategies under discretetime trading. The Journal of Economic Dynamics and Control, 33, 204–220. [4] Balder, S and A Mahayni (2008). Cashlock comparison of portfolio insurance strategies. Technical Report. [5] Basak, S (1995). A general equilibrium model of portfolio insurance. Review of Financial Studies, 8(4), 1059–1090. [6] Basak, S (2002). A comparative study of portfolio insurance. The Journal of Economic Dynamics and Control, 26(7–8), 1217–1241. [7] Benninga, S and M Blume (1985). On the optimality of portfolio insurance. Journal of Finance, 40(5), 1341–1352. [8] Bergman, YZ, BD Grundy and Z Wiener (1996). General properties of option prices. Journal of Finance, 51, 1573–1610. [9] Bertrand, P and JL Prigent (2002). Portfolio insurance strategies: OBPI versus CPPI. Technical Report, GREQAM and Université Montpellier1, pp. 68–86. [10] Bertrand, P and JL Prigent (2002). Portfolio insurance: The extreme value approach to the CPPI. Finance, 23(2). [11] Bertrand, P and JL Prigent (2003). Portfolio insurance strategies: A comparison of standard methods when the volatility of the stock is stochastic. International Journal of Business, 8(4), 15–31. [12] Black, F and R Jones (1987). Simplifying portfolio insurance. The Journal of Portfolio Management, 14, 48–51. [13] Black, F and AR Perold (1992). Theory of constant proportion portfolio insurance. The Journal of Economic Dynamics and Control, 16(3–4), 403–426. [14] Bookstaber, R and JA Langsam (2000). Portfolio insurance trading rules. The Journal of Futures Markets, 8, 15–31. [15] Brennan, MJ and ES Schwartz (1976). The pricing of equitylinked life insurance policies with an asset value guarantee. Journal of Financial Economics, 3, 195–213. [16] Brennan, MJ and ES Schwartz (1989). Portfolio insurance and financial market equilibrium. Journal of Business, 62, 455–472. [17] Browne, S (1999). Beating a moving target: Optimal portfolio strategies for outperforming a stochastic benchmark. Finance and Stochastics, 3, 275–294. [18] Cesari, R and D Cremonini (2003). Benchmarking, portfolio insurance and technical analysis: A monte carlo comparison of dynamic strategies of asset allocation. Journal of Economic Dynamics and Control, 27, 987–1011. [19] Cont, R and P Tankov (2007). Constant proportion portfolio insurance in presence of jumps in asset prices. Financial Engigeering No. 2007–10, Columbia University Center for Financial Engigeering. [20] Cox, CJ and HE Leland (2000). On dynamic investment strategies. The Journal of Economic Dynamics and Control, 24, 1859–1880. [21] Cox, JC and CF Huang (1989). Optimal consumption and portfolio policies when the asset price follows a diffusion process. Journal of Economic Theory, 49, 33–83. [22] Døskeland, TM and HA Nordahl (2008). Optimal pension insurance design. Journal of Banking and Finance, 32, 382–392. [23] Dudenhausen, A, E Schlögl and L Schlögl (1998). Robustness of Gaussian hedges under parameter and model misspecification. Technical Report, University of Bonn, Department of Statistics. [24] El Karoui, N, M Jeanblanc and V Lacoste (2005). Optimal portfolio management with american capital guarantee. Journal of Economic Dynamics and Control, 29, 449–468.
May 12, 2010
17:47
WSPC/SPIB913
How Good are Portfolio Insurance Strategies?
b913ch10
FA
257
[25] El Karoui, N, M JeanblancPicqué and SE Shreve (1998). Robustness of the Black and Scholes formula. Mathematical Finance, 8(2), 93–126. [26] Grossman, SJ and J Villa (1989). Portfolio insurance in complete markets: A note. Journal of Business, 62, 473–476. [27] Grossman, SJ and J Zhou (1993). Optimal investment strategies for controlling drawdowns. Mathematical Finance, 3, 241–276. [28] Grossman, SJ and J Zhou (1996). Equilibrium analysis of portfolio insurance. Journal of Finance, 51, 1379–1403. [29] Hobson, D (1998). Volatility misspecification, option pricing and superreplication via coupling. Annals of Applied Probability, 8(1), 193–205. [30] Kahnemann, D and A Tversky (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47, 263–291. [31] Leland, HE (1980). Who should buy portfolio insurance. Journal of Finance, 35(2), 581–594. [32] Leland, HE (1988). Portfolio insurance and october 19th. California Management Review, Summer, 80–89. [33] Leland, HE and M Rubinstein (1976). The Evolution of Portfolio Insurance. In: Portfolio Insurance: A guide to Dynamic Hedging, DL Luskin (ed.), pp. 3–10, Wiley. [34] Lyons, TJ (1995). Uncertain volatility and the riskfree synthesis of derivatives. Applied Mathematical Finance, 2, 117–133. [35] Mahayni, A (2003). Effectiveness of hedging strategies under model misspecification and trading restrictions. International Journal of Theoretical and Applied Finance, 6(5), 521–552. [36] Mahayni, A and E Schlögl (2008). The risk management of minimum return guarantees. Business Research, 1, 55–76. [37] Merton, RC (1971). Optimal consumption and portfolio rules in a continuous time model. Journal of Economic Theory, 3, 373–413. [38] Munk, C (2006). Financial asset pricing theory. Lecture notes. [39] Prigent, JL and F Tahar (2006). Optimal portfolios with guarantee at maturity: Computation and comparison. International Journal of Business, 11(2), 171–185. [40] Tepla, L (2000). Optimal portfolio policies with borrowing and shortsale constraints. The Journal of Economic Dynamics and Control, 24, 1623–1639. [41] Tepla, L (2001). Optimal investment with minimum performance constraints. The Journal of Economic Dynamics and Control, 25(10), 1629–1645. [42] Tversky, A and D Kahnemann (1992). Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5, 297–323. [43] von Neumann, J and O Morgenstern (1944). Theory of Games and Economic Behavior, 1953 edn. Princeton, NJ: Princeton University Press. [44] Zagst, R and J Kraus (2008). Stochastic dominance of portfolio insurance strategies — OBPI versus CPPI. Technical Report. [45] Zhang, PG (1998). Exotic Options — A Guide to Second Generation Options. World Scientific.
May 12, 2010
17:47
WSPC/SPIB913
b913ch10
This page intentionally left blank
FA
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
FA
11
PORTFOLIO INSURANCES, CPPI AND CPDO, TRUTH OR ILLUSION?
ELISABETH JOOSSENS∗,‡ and WIM SCHOUTENS†,§ ∗
†
Joint Research Centre of the European Commission Katholieke Universiteit Leuven, Department of Mathematics, Celestijnenlaan 200B, 3001 Leuven, Belgium ‡ elisabeth.joossens@jrc.it § wim@schoutens.be
Constant proportion portfolio insurance (CPPI) and constant proportion debt obligations (CPDO) strategies have recently created derivative instruments, which try to protect a portfolio against failure events and have only been adopted in the credit market for the last couple of years. Since their introduction, CPPI strategies have been popular because they provide protection while at the same time they offer high yields. CPDOs were only introduced into the market in 2006 and can be considered as a variation of the CPPI with as main difference the fact that CPDOs do not provide principal protection. Both CPPI and CPDO strategies take investment positions in a riskfree bond and a risky portfolio (often one or more credit default swaps). At each step, the portfolio is rebalanced and the level of risk taken will depend on the distance between the current value of the portfolio and the necessary amount needed to fulfill all the future obligations. We first analyze in detail the dynamics of both investment strategies and afterwards test the safetyness of both products under a multivariate Lévy setting. More precise we first propose a quick way to calibrate a multivariate Variance Gamma (VG) process on correlated spreads, which can then be used to quantify the gap risk for CPPIs and CPDOs.
259
May 12, 2010
17:47
WSPC/SPIB913
260
b913ch11
FA
Joossens and Schoutens
11.1. INTRODUCTION Financial institutions try to protect their portfolios against failure events and derivative instruments are a possible solution. Derivative instruments are a fastgrowing market in which alternative investment strategies such as Constant Proportion Portfolio Insurance (CPPI) and Constant Proportion Debt Obligation (CPDO) are created. Although these two recently developed instruments function in different ways when deciding on their investment strategy, both investment funds attempt to provide a portfolio insurance. More precisely, their strategy is to invest only a part of the capital in a risky asset and to invest the remainder in a safe way. The total value of the portfolio at each time step will influence the position taken in the risky asset. The decisions taken on the risk position at each time step aim to allow the investor of a CPPI or CPDO to recover, at maturity, a given percentage of their initial capital, which allows them to benefit from a capital guarantee while participating in the upside of an underlying asset. It could happen that the promised return is not achieved. In this case, for the CPPI structure, the bank will have to cover the losses at maturity while for the CPDO structure, the CPDO will unwind and the investor will not receive the promised amount at maturity but only the remainder amount at the time of unwinding. CPPI structures clearly safeguard a given percentage of the invested capital for the investor, while for CPDO the investor appears to be taking a risk. In the past, this risk has always been seen as very small and CPDOs have been sold as very safe. Here, we will study those two products in depth in order to answer the question whether CPDO and CPPI are really as safe and attractive as they seem. First, the concept of credit risk will be introduced. This is the risk that, after agreeing on a certain contract, one of the involved parties will not fulfil its financial obligations (such as paying a premium). Often the quality and price of financial products will heavily depend on this risk. Different ways to model this risk are presented and can be used to price financial products. Next, a credit default swap (CDS) is introduced. This financial instrument tries to provide a protection for credit risk by transferring it. In exchange for a predefined cost, also called “spread”, a second party will cover the losses one might suffer due to credit risk. The costs for such an insurance will clearly depend on the size of the risk and the impact of the possible losses. Hence, the price can be determined using the abovementioned models for credit risk. There are no obligations to trade credit default swaps so their changes in credit spread can be used for speculations, for example: CDS spreads can be used as risky assets in investment structures such as CPPI and CPDO. The third section will be dedicated to a particular type of insurance portfolio, the CPPIs. They are a derivative security with capital guarantee using a dynamic trading strategy in order to incorporate the performance of a certain underlying product such
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
FA
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
261
as a simple stock or a CDS. They were introduced more than 10 years ago and are frequently used. The investment structure of CPPIs has been and continues to be a popular topic for research. First, a multivariate jumpdriven model, which can be applied for pricing credit derivatives such as CPDOs, is discussed in detail in Sec. 11.4. Next, in Sec. 11.5, a more general overview of the other recent developments regarding CPPIs is given. All papers could be classified under three different fields of interest. A first group of papers concentrates on proposing different models for pricing. The second group concentrates on the estimation of the leverage factor in order to fix an upper bound for the gaprisk, and the third group tries to extend the structure even more in order to include an extra safety factor. One of the recent new products based on the idea of a CPPI is the CPDO or constant proportion debt obligation, which is discussed in Sec. 11.6. CPDOs are used for credit portfolios comprising exposures to credit indices such as iTraxx and CDX. The CPDO structure borrows many features from the CPPI structure, such as the constant proportion. The main goal of a CPDO is to produce a highyielding product and this is achieved through a high degree of leverage. Contrary to a CPPI investment strategy, leverage will be increased when the net asset value of the portfolio decreases and descends below the target amount, but leverage will be decreased when the net asset value of the portfolio increases and approaches the target. Once a CPDO reaches this target amount, it will completely deleverage. In this section, we will not only focus on the dynamics of the structure but we will also give a short overview of the research concentrating on this topic and discuss the question of the “safeness” of this structure as has been highlighted recently in the news. We will conclude by discussing in more depth the differences and similarities of CPPI and CPDO. This should give an even better insight into the structure of both financial instruments.
11.2. CREDIT RISK AND CREDIT DEFAULT SWAPS This section is intended to give a short introduction to the main financial concepts, which will play a role in this work. First, the concept of credit risk is introduced and different ways to model this risk are discussed. Next, a short introduction on credit derivatives, i.e., more precise credit default swaps, is provided.
11.2.1. Credit Risk Credit risk refers to the risk that a specified reference identity does not meet its credit obligations within a specified time horizon T (called maturity). In other words, whenever two or more parties sign an agreement, there is a risk that one of them will not
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
262
FA
Joossens and Schoutens
meet its obligations. A simple example is the case, where a single person signs a loan with a bank. Here, it can happen that the person does not repay his debt according to the agreement. In such a case, we say that default will have occurred. The risk that such an event will happen is called credit risk and will always be spread over a certain time length. Taking a more global perspective, in finance we do not only deal with the situation of a person and a loan but it will always be possible to characterize the credit risk in terms of the following components: the obligor, the set of criteria defining the default, and a time interval over which the risk is spread. When, for instance, we talk about bonds, their default can be defined in several ways. It could, for example, be bankruptcy but it might also be a rating downgrade of the company or failure to pay an obligation (such as a coupon). But it could also concern the value of a firm — here a firm’s value is linked to the value of its financial assets. In general one will look at the firm’s asset value V = Vt , 0 ≤ t ≤ T and default will be defined as a boundary condition on the asset value. For example a default event will occur if the value falls below a certain fixed level L within the time horizon. Figure 11.1 presents two possible paths 250
200
V
t
150
100
50
0
0
1
2
3
4
τ
5 T
6
7
8
9
10
Figure 11.1 Two possible paths for a firm’s asset value over time with T = 10. When the black line occurs the obligor will default (when Vt < L) in the gray case the obligor will survive.
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
263
for the firm value over time [0, 10] modeled through Black–Scholes, where µ = 0.05, σ = 0.4, and S0 = 100 as explained below. In the case of the black line, the value of the firm will fall below the lower bound, which is fixed at 20 just before t = 5 and hence will default. In the case that the firm value follows the gray path, no default will occur before T = 10. As one will try to protect oneself from defaults, the size of credit risk will be one of the crucial factors when determining the prices and hence the techniques for estimating the probability of default of a reference entity within time T will be very important. Developing new models for the estimation of credit risk is an important topic in the field of finance. The strong interest in this topic is also linked to the fact that in the last couple of years the volume of instruments linked to credit risk traded on the market has increased exponentially. Besides the increase in investments, there has also been interest due to the Basel II Accord, which encourages financial institutions to develop methods for assessing their risk exposure. Credit risk models are usually classified into two categories: reducedform models, including intensitybased models, and structural models. Intensitybased models, also known as hazardrate models, focus directly on modeling the default probability. The main idea of these models lies in the fact that at any moment in time (as long as the contract is running) there is a probability that an obligor might default. Default is defined at the first jump of a counting process with a certain intensity. In practice, the models assume that the intensities of the default times follow a certain process (stochastic or deterministic) and under those conditions the underlying default model can be constructed. This intensity of the process depends heavily on the firm’s overall health and on the situation of the market. The structural models, also known as firmvalue models, link default events to the value of the financial assets of the firm, such as in the example presented above. Credit risk will hence depend on the model used for the value of the financial assets of the firm and the criteria used for a default. This approach will almost always be used in the remainder of this chapter when modeling the credit default. A common way to model the time evolution of assets uses the following diffusion process: dSt = St (µdt + σdWt ),
S0 > 0,
(11.1)
where Wt is a standard Brownian motion. Here, µ and σ > 0 are the socalled drift and volatility factors. The path of a Brownian motion is continuous and can easily be simulated when one discretizes the time using very small steps t. The value of a Brownian motion at time {nt, n = 1, 2, . . .} is obtained by sampling a series of independent standard
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
264
FA
Joossens and Schoutens
normal random numbers {νn , n = 1, 2, . . .} and setting: W0 = 0
and Wnt = W(n−1)t +
√ tνt .
The solution of the SDE given in Eq. (11.1) is called geometric Brownian motion and is given by St = S0 exp((µ − σ 2 /2)t + σWt ).
(11.2)
The above way to model price changes of financial products is also referred to as the Black–Scholes model. The main advantages of using the Black–Scholes model is that it is easy to understand. Calculation of prices of derivatives under the model is moreover typically not that timeconsuming. A drawback is that it assumes normality of the log returns of the financial assets, which is often not true in reality. Moreover, the Black–Scholes model does not capture the possibility of sudden jumps, which do occur in real live and often cause extra credit risk. Hence, a more flexible stochastic process is often required to model reality in a better way. It would be good to keep some properties of the Brownian motion such as independence and stationarity of the increments but to drop the constraints of normality and continuity of the paths. To create such a process, we must restrict ourselves to the group of infinitely divisible distributions. For each infinitely divisible distribution (with characteristic function φ(u)), a stochastic process can be defined which starts at zero and has independent and stationary increments such that the distribution of the increments over [s, s + t], s, t ≥ 0 has (φ(u))t as characteristic function. Such processes are called Lévy processes, in honour of Paul Lévy, a pioneer of the theory. Definition 11.1. Lévy process: A cadlag stochastic process X = {Xt , t ≥ 0} defined on a probability space (, F, P) is a Lévy process if the following conditions hold: (1) Xt is a continuous process Palmost surely: ∀ε > 0 :
lim P(Xt+h − Xt  ≥ ε) = 0.
h→0
(2) X0 = 0. (3) The process has stationary increments. (4) The process has independent increments. Lévy process became very popular and are still more and more used in practice. Examples of Lévy process are the (compound) Poisson process, the Gamma process, the inverse Gaussian process, and the Variance Gamma process (VG). An indepth study of Lévy process in finance and how they can be applied can be found in [1]. Next, we will define an important group of credit derivatives, which are often used in practice.
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
265
11.2.2. Credit Default Swaps (CDS) Credit Default Swaps (CDSs) are very simple credit derivatives and have a big share in the market of credit derivatives. Credit derivatives can be defined as the group of all derivatives, whose payoffs are affected by the default of a specified reference entity (or a basket of entities). They are often used to hedge, transfer, or manage the risk and can hence be considered as an insurance against default. The main idea of credit derivatives is that credit risk is transferred without reallocating the ownership of the underlying asset(s). This way they provide a certain protection against decreasing solvency or default of the underlying asset(s). CDSs in particular are designed to isolate the risk of default on a credit obligation. A CDS is a bilateral agreement, where the protection buyer transfers the credit risk of a reference entity to the protection seller for a determined amount of time T . In exchange for this shift of risk, the protection buyer will make predetermined payments to the protection seller. These payments will occur until the end of the contract (the time of maturity) T unless a default event occurs before the time to maturity. If default of the reference entity occurs, the protection seller will cover the losses (or part of the losses) of the protection buyer due to the default of the underlying entity and the contract will be terminated. The yearly rate paid by the protection buyer to enter a CDS contract against failure is called the CDS spread. Spreads are almost always quantified in bp where bp stands for “basis point” and is equal to 0.01%. The amount of the spread will reflect the riskiness of the underlying credit, if the probability of default increases also the cost of the CDS (and hence the spread) will increase. We note that CDSs are often also used to speculate on changes in credit spread. Figure 11.2 presents the cash flows for two possible scenarios (default at time t = 7 or no default) for an example. We consider the case where a person owns a zerocoupon defaultable bond of a company with a face value F = 10,000 Euro and maturity T = 10 years. Suppose that this person would like to cover himself against the possible default of the bond. He can buy this protection by entering into a CDS contract. A possible situation would be that the contract requests an annual payment of an amount of 400 bp from the protection buyer to be protected against the default. In return, the protection seller will cover the loss, which might result from defaulting. The amount of the loss will be equal to the difference between F and the recovery value after default. We hence take into account that when a default event occurs, the total amount will not automatically be lost completely, but part of the value might be recovered. The concept of recovery can be understood through the following example. In the case that a company goes bankrupt, there are creditors claiming against the assets of the company, and the owner of the bond is one of those creditors. The assets are sold by a liquidator and the profits are used to meet the claims as far as possible. Historically values of the recovery rate fall between 20%
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
266
FA
Joossens and Schoutens 7000
6000
5000
Euro
4000
3000
2000
1000
0
1000
1
2
3
4
5
6
7
8
9
10
Year
Figure 11.2 Cash flows from the protection buyer for a 10year CDS. Black: cash flows in case no default occurs. Gray: cash flows in case there is a default at time t = 7.
and 50%. For the current example, we assume that the recovery rate will be equal to R = 40%. The annual amount paid by the protection buyer in this example is hence equal to 400 bp · 10,000 = 400 Euro and the payment of the protection seller in case of default will be F(1 − R) = 6000 Euro. In Fig. 11.2, it is assumed that in the second scenario the bond defaulted at the beginning of the seventh year. Pricing models for CDSs based on Lévy processes can be found in [2] and [3]. In practice, CDS are not only used to reallocate the risk of a single asset (the socalled single name CDS), but also a basket of assets might be considered. A credit default swap index is a credit derivative used to hedge credit risk or to take a position on a basket of credit entities. There are currently two main families of CDS indices: CDX and iTraxx. CDX indices contain North American and Emerging Market companies, whereas iTraxx contains companies from the rest of the world. One of the most widely traded index is the iTraxx Europe index composed of the most liquid 125 CDS referencing European investment grade credits. There are also significant volumes, in nominal values, of trading in the HiVol index. HiVol is a subset of the main index consisting of what are seen as the most risky 30 constituents at the
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
FA
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
267
time the index is constructed. Typically every six months new series of CDS indices are issued. Indices are like CDSs not only used as insurance against default risk but are also traded on the market in a speculative way.
11.3. PORTFOLIO INSURANCES Portfolio Insurances are capital guarantee derivative securities that embed a dynamic trading strategy in order to make a contribution to the performance of a certain underlying product (e.g., an asset, a CDS or a CDS index, …). One particular type is considered here, the constant proportion portfolio insurance e.g., [4], investing partially in a riskfree way and combining this with a risky asset. The family of Constant Proportion Portfolio Insurance consists of investments for which the amount necessary for guaranteeing a repayment of a fixed amount N at maturity T is invested in a riskfree way, typically a bond, B, and only the exceeding amount will be invested in one or more risky assets, S (i) . This way an investor can limit its downside risks and maintain some upside potential. This type of portfolio insurance has first been introduced by [5] and [6]. The product manager will take larger risks when his strategy performed well. But if the market went against him, he will reduce the risk rapidly. The following factors play a key role in the risk strategies an investor will take: • Price: The current value of the CPPI. The value at time t ∈ [0, T ] will be denoted as Vt . • Floor: The reference level to which the CPPI is compared. This level will guarantee the possibility of repaying the fixed amount N at maturity T , hence it could be seen as the present value of N at maturity. Typically this is a zerocoupon bond and its price at time t will be denoted as Bt . • Cushion: The cushion is defined as the difference between the price and the floor: Cushion = Price − Floor. • Cushion%: It is defined as the ratio of the cushion over the price. • Multiplier: The multiplier is a fixed value, which represents the amount of leverage an investor is willing to take. • Investment level: It is the percentage invested in the risky asset portfolio; this also known as the exposure and is for each step fixed at: e = Multiplier × Cushion%. • “gap” risk: This is the probability that the CPPI value will fall under the floor.
May 12, 2010
268
17:47
WSPC/SPIB913
b913ch11
FA
Joossens and Schoutens
The level of risk an investor will take at each time t is equal to the investment level as long as the value of the CPPI exceeds the floor. For any time t, the future investment decision will be made according to the following rule: • If Vt ≤ Floor = Bt , we will invest the complete portfolio into the zerocoupon bond. • If Vt > Floor, we will invest a proportion equal to e in the risky asset portfolio. It can easily be shown that under the assumption that the underlying asset price follows a Black–Scholes model with continuous trading, there is no risk of going below the floor and that the expected return at maturity of the CPPI is equal to [7] E[VT ] = N + (V0 − Ne−rT ) exp(rT + m(µ − r)T ). In practice, however, it is known that the probability of going below the floor is nonzero. It might, for instance, happen that during a sudden downside move or due to overnight changes, the fund manager might not be able to adjust the portfolio in time. In the case of an event where the actual portfolio value falls under the floor, at maturity the issuer will have to cover the difference between the actual portfolio value and the guaranteed amount N. It is therefore of importance for the issuer of a CPPI note to be able to quantify this risk, also called gap risk. We will present an example of a possible cash flow for a CPPI with maturity T equal to 10 years. For the sake of simplicity, we will consider only one underlying asset with prices St and a riskfree asset, a zerocoupon bond Bt with a constant interest rate r = 5%. We also assume that the initial price of the asset is equal to S0 = 100 and the prices over time are simulated from a Variance Gamma model, which will be presented later (see Eq. (11.4)) with parameters σ = 0.5, ν = 0.25, and θ = 0.026. For the CPPI process, the leverage or multiplier is fixed at 2.5 and the starting capital is 100. We also consider that the CPPI at maturity repays the investor with at least the initial capital. Figure 11.3 presents two examples of possible scenarios for the simple CPPI. In the example on the left, the value of the CPPI will always stay above the floor, while in the second example (on the right) at time τ a sudden drop of the risky asset will result in a CPPI value below the floor, which is the gap risk. Note that different scales are applied for the presentation of the results. In this example, as the repayment at maturity of the initial value should be insured, the floor will be 100 exp(−r(10 − t)) at each time t. For each step, the value of the cushion is calculated and the portfolio is rebalanced according to the risk exposure. The rebalancing is such that the bigger the difference between the CPPI value and the floor, the higher the cushion value and the more risk one will take. The process will stop at maturity or once a drop of the asset value occurs of such a level that the CPPI value falls below or hits the floor. If such a drop happens the product manager will put the risk exposure to zero and only invest in a riskfree way until maturity.
Value
Value
Stockvalue
100
100
Stockvalue 0
1
2
3
4
5
6
7
8
9
0
10
120
200
100
150
Value
Value
0
80 60
CPPI Bondfloor 0
1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
10
CPPI Bondfloor
100 50
10
0
0
1
2
3
4
5
6
7
8
9
10
60
Value
20 0
Cushion
100
50
0 0
1
2
3
4
5 Time
6
7
8
9
10
0
1
2
3
4
5 Time
6
7
8
9
10
FA
Figure 11.3 Value of the risky asset (top), CPPI performance (central), and cushion (bottom) in case the CPPI value stays above the floor (left) and in case a sudden downwards jump occurs and the CPPI value drops bellow below the floor (right).
b913ch11
Value
Cushion 40
WSPC/SPIB913
200
17:47
200
May 12, 2010
CPPI value drops bellow floor 300
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
CPPI value stays above floor 300
269
May 12, 2010
17:47
WSPC/SPIB913
270
b913ch11
FA
Joossens and Schoutens
Recent papers discuss different aspects of this group of investment strategies. In the next section, an overview of the main findings is given.
11.4. MODELING OF CPPI DYNAMICS USING MULTIVARIATE JUMPDRIVEN PROCESSES This section is based on [8] and presents a possible way of modeling credit CPPI structures. More precise, a dynamic Lévy model (VG) is set up for a series of correlated spreads. A recent innovation is the possibility to trade swaptions, an option granting its owner the right but not the obligation to enter into an underlying swap (e.g., a CDS), on indices. We will make use of these instruments to calibrate the underlying dynamic spread models. The parameters of the model actually come from a twostep calibration procedure. First, a joint calibration of the correlated spreads on swaptions and second by a correlation matching procedure. For the joint calibration, we make use of equitylike pricing formulas for payers and receivers swaptions based on characteristic functions and Fast Fourier Transform methods. To obtain the required correlation, we set by a closedform matching procedure the models correlation exactly equal to a prescribed (e.g., historical) correlation. We then have a model in place that can generate very fast correlated spread dynamics under jump dynamics. The calibrated model can be used to price a whole range of exotic structures. We illustrate this by pricing credit CPPI structures. The CPPI structures considered in this section are such that the invested capital is put in a riskfree bond and a position is taken on credit derivatives indices (usually protection is sold). As discussed before, important in the handling of CPPIs is assessing the risk that spreads of the underlying credit indices jump and lead to socalled gap risk. Because of the built in jump dynamics, a better assessment of this gap risk is made possible.
11.4.1. Multivariate Variance Gamma Modeling Next, a dynamic Lévy model, more precise a Multivariate Variance Gamma (MVG) model, is set up for a series of correlated spreads. As this model can generate correlated spreads in fast way, it can be applied in order to price different exotic structures such as the CPPI. For the construction of the Multivariate Variance Gamma model, the authors start from the Univariate Variance Gamma process. The Variance Gamma process was introduced in the financial literature by [9]. The characteristic function of the Variance Gamma (VG(σ, ν, θ)) distribution is defined as follows: 1 2 2 −1/ν . φVG (u; σ, ν, θ) = 1 − iuθν + σ νu 2
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
FA
271
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
The distribution is known to be infinitely divisible and by standard Lévy theory one thus can build out of such a distribution a process. The VG process is defined as X(VG) = {Xt(VG) , t ≥ 0} with parameters σ, ν > 0, and θ. It is a process, which starts at zero has stationary and independent increments, and for which the increments √ (VG) − Xs(VG) follow a VG(σ t, ν/t, tθ) distribution over the time [s, t + s] (for more Xs+t details see [1]). Theoretical background on Lévy process can be found in [10] and [11]. For some historical background on the VG model see [12]. Another way of constructing the VG process is by the technique of time changing. Here, we will start from a Gamma process. Recall that the density function of a Gamma(a, b) distribution is given by ba a−1 x exp(−xb), x > 0, and a, b > 0, (a) where (.) is the Gamma function. It is known that the distribution is infinitely divisible and hence using the Gamma distribution one can build a process with independent and stationary Gamma increments. Defining G = {Gt , t ≥ 0} as a Gamma process with parameters a = b = 1/ν, the resulting Gt will follow a Gamma(at, b) distribution and E[Gt ] = t. A VG process can be constructed by timechanging a Brownian Motion with drift. More precisely, one can show (for more details see [1]) that the process fGamma (x; a, b) =
Xt(VG) = θGt + σWGt ,
t ≥ 0,
(11.3)
where W = {Wt , t ≥ 0} is a standard Brownian motion independent from the Gamma process, is indeed a VG process with parameters (σ, ν, θ). The construction can be interpreted as if we now look at a Black–Scholes world but now measured according to a new business clock (Gamma time). It has proven to be very successful in the univariate setting, as the underlying VG distribution can take into account, in contrast to the Normal distribution, skewness, and excesskurtosis. In [13], a VG model was used to accurately fit CDS curves. With this alternative definition, we have an easy way to simulate a sample path of the VG processes, which can be obtained by sampling a standard Brownian motion and a Gamma process. The Gamma process can easily, like the Brownian motion, be simulated at time points {nt, n = 1, 2, . . .} with t small. First generate independent Gamma(at, b) random numbers {gn , n = 1, 2, . . .}. Then, the Gamma process can be constructed by G0 = 0
and Gnt = G(n−1)t + gn ,
n ≥ 1.
Similarly to Eq. (11.2) for the Brownian motion, the value over time of an asset using the dynamics of a Variance Gamma process can now be modeled by St = S0 exp(ωt + θGt + σWGt ), with ω = ν−1 log(1 − σ 2 ν/2 − θν).
(11.4)
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
272
FA
Joossens and Schoutens
We work with multivariate extensions of the model along the technique described in [14] and [15]. Additionally, we want to note that in [9], a symmetric version of a multivariate VG is initiated. To build the multivariate VG process a Gamma process G = {Gt , t ≥ 0} is needed, where the parameters a and b are both fixed at 1/ν. Also an Ndimensional Brownian motion is introduced W = {W t , t ≥ 0}, where W t = (Wt(1) , Wt(2) , . . . , Wt(N) ). It is assumed that this process is independent of the Gamma process and that the Brownian motions have a correlation matrix given by ρW = (ρijW , i, j, = 1, . . . , N) (j)
(i) ρijW = E[W1 , W1 ].
A multivariate VG process X = {X t = (Xt(1) , . . . , Xt(N) ), t ≥ 0} is defined as Xt(i) = θi Gt + σi WG(i)t . Note that there is dependence between the Xt(i) ’s due to two causes: they are all constructed by a time change with a common Gamma time. This will mean that the processes will all jump together, but jumps’ sizes can be different. Moreover, there is dependency also built in via the Brownian motions. A straightforward calculation shows that the correlation between two components is given by ρij =
(j) (j) θi θj ν + σi σj ρijW E[X1(i) X1 ] − E[X1(i) ]E[X1 ] = . (j) σi2 + θi2 ν σj2 + θj2 ν Var[X1(i) ] Var[X1 ]
(11.5)
This clearly shows that, even if we assume the Brownian components to be independent of each other, one still obtains a correlation between the different components because of the common gamma time. Suppose that we need to model, as later on will be the case, the evolution of Ncorrelated spreads. Take for example the evolution of the iTraxx Europe main index and its overseas opponent the Dow Jones CDX.NA.IG main index and/or the spreads of their HiVol subsets. We assume the following correlated dynamics for the evolution of Ndependent spreads: St(i) = S0(i) exp(ωi t + θi Gt + σi WG(i)t ) = S0(i) exp(ωi t + Xt(i) ), where
i = 1, . . . , N,
1 ωi = ν−1 log 1 − σi2 ν − θi ν . 2
These meancorrection terms, ωi , are in place because one can then easily show that the spread processes are meanreverting in the sense that we have for every t ≥ 0 E[St(i) ] = S0(i) .
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
273
11.4.2. Swaptions on Credit Indices The two most liquidly traded index swaptions types are payers and receivers. These are typically European style. A payer/receiver option holder has on expiry the right but not the obligation to buy/sell protection on the underlying index at the strike level. If a default happens among the index constituents prior to option expiration, both the buyer of a payer or the seller of a receiver option can trigger on expiry a credit event (socalled nonknockout feature). The payoff of an index swaption at expiry has two components: payoff due to difference between expiry spread level and the strike and payoff due to any default losses. For shortmaturity options, the later is very unlikely to happen and is often ignored. Let us introduce some notation. Denote by T the (payer or receiver) swaption maturity (typical 3, 6, or 9 months) and by T ∗ the index maturity (typically 5, 7, or 10 years). Let us denote with At the risky annuity for maturity t (i.e., the present value of 1 bp of the fee leg). The forward annuity is denoted by A(T, T ∗ ) and is the forward annuity from swaption maturity to index maturity as of the trade day (t = 0). We have of course that At = A(0, t) and A(T, T ∗ ) = AT ∗ − AT . The forward spread as of the trade day (i.e., at time t = 0) with 0 ≤ T < T ∗ , F0 = F0 (T, T ∗ ) is the forward spread from swaption maturity T to index maturity T ∗ and is given by St At − Ss As . F0 (s, t) = At − A s
11.4.2.1. Black’s model The market standard for modeling credit spreads options is a modification of the Black’s formula for interest rate swaptions (see ( [16])). It models spread dynamics in the same way as in Eq. (11.2), where µ = 0. Black’s formula for the value of a payer/receiver swaption with maturity T and strike value K is given by Payer(T, K) = A(T, T ∗ )(F0 N(d1 ) − KN(d2 )), Receiver(T, K) = A(T, T ∗ )(KN(−d2 ) − F0 N(−d1 )), where √ log(F0 /K) + σ 2 T/2 and d2 = d1 − σ T . √ σ T If the payer swaption is nonknockout, as is typically the case for the index swaptions we are dealing with, we adjust the forward spread to account for the nonknockout feature of index options. We account for this additional protection by increasing the d1 =
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
274
FA
Joossens and Schoutens
forward spread by the cost of this protection. More precisely, the adjusted forward spread is given by (adj)
F0
(adj)
= F0
(T, T ∗ ) = F0 (T, T ∗ ) +
S T ∗ AT , AT ∗ − A T
and the price of nonknockout payers and receivers are respectively given by Payer(T, K) = A(T, T ∗ )(F0
(adj)
N(d1 ) − KN(d2 )),
Receiver(T, K) = A(T, T ∗ )(KN(−d2 ) − F0
(adj)
N(−d1 )),
where (adj)
d1 =
log(F0
/K) + σ 2 T/2 √ σ T
and
√ d2 = d1 − σ T .
Summarizing, we want to remark on the striking connection with vanilla option prices in equity. Basically, pricing comes down to pricing under a Black–Scholes regime with no interest rates and no dividends. However, the model has all the deficiencies of the Black–Scholes framework: nojumps, lighttails, symmetric underlying distribution, etcetera.
11.4.2.2. The variance gamma model Completely similar to the equity setting. The Black–Scholes dynamics are replaced with the better performing jump dynamics of a Variance Gamma process. We now model the spread dynamics as St = S0 exp(ωt + θGt + σWt ) = S0 exp(ωt + Xt ). Pricing of vanillas has already been worked out in full detail in equity settings by [17] and by a slight adaption to the credit setting we can very fast calculate payer and receiver swaptions using the Carr–Madan formula in combination with fast Fourier transform methods. More precisely, Payer(T, K) = A(T, T ∗ ) ×
exp(−α log(K)) π
+∞
exp(−iv log(K)) 0
φ(v − (α + 1)i; T ) dv, α2 + α − v2 + i(2α + 1)v
where the characteristic function of the log of the adjusted forward spread process at maturity T is given by (adj)
φ(u; T ) = E[exp(iu(log F0
+ ωT + XT ))],
which is known analytically in many Lévy settings, including VG.
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
275
11.4.3. Spread Modeling by Correlated VG Processes We now have a multivariate spread model available, and a fast pricers for the standard payer and receiver swaptions. Using these fast pricers on these index swaption, one can calibrate the model in a two steps procedure. (1) We will make sure that our model reproduces the swaption market data as best as possible by doing a joint calibration using our fast FFT pricer. This step will give parameter estimators for the σi , ν, and θi . (2) We put in place the exact correlation structure we want, by calculating (using a closedform formula) the correlation matrix of the underlying driving standard Brownian motions. Note, that the matrix calculated in the this fashion is not necessary positivedefinite. In such a situation, one can then as a kind of ad hoc method look for the closed positivedefinite matrix and work with that one. The model is illustrated by a worked out example, where we price a CPPI structure.
11.4.3.1. The pricing of CPPIs We work out the details of credit CPPI by an example, where positions will be taken in four highly correlated indexes and a predefined trading strategy is in place. In our example, we take positions in the following index products: • • • •
iTraxx Europe Main on the run (5 years) unfunded iTraxx Europe HiVol on the run (5 years) unfunded DJ CDX.NA.IG Main on the run (5 years) unfunded DJ CDX.NA.IG HiVol on the run (5 years) unfunded
The swap rates for the above products are highly correlated as can be seen from Fig. 11.4. The corresponding correlation matrix of spreads itself and the correlation matrix of the corresponding daily log returns based on observations from the 21st June 2004 until 13th March 2007 are given in Tables 11.1 and 11.2, respectively. We start with a portfolio of say 100M EUR and an investment horizon of six years. We want to have the principal of the initial investment protected. Therefore, we calculate the bond floor as the value of a riskfree bond that matures at the end of the investment horizon. Suppose we take r = 0.04 and use compound interest rates, then the bondfloor is initially at 78.6628M EUR. Suppose we set the leverage at m = 30. Initially, the cushion is thus at 21.3372M EUR. Multiplying the cushion with the constant leverage factor of 30 gives the risky exposure that we are going to take, namely, 640.12M EUR. We are taking the following positions: • Sell protection on iTraxx Europe Main on the run (5 years) for half of the risky exposure.
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
276
FA
Joossens and Schoutens Evolution of Main and HiVol indices : iTraxx and CDX.IG 180 iTraxx Main iTraxx HiVol CDX Main CDX HiVol
160 140
Spread
120 100 80 60 40 20 Jan04
Time
Mar07
Figure 11.4 Evolution of Main and HiVol indices: iTraxx and CDX.IG. Table 11.1 Correlation of Spreads: iTraxx (Main and HiVol) and CDX (Main and HiVol). Correlation
iTtraxx Main
iTraxx HiVol
CDX Main
CDX HiVol
iTtraxx Main iTraxx HiVol CDX Main CDX HiVol
1.0000 0.9162 0.9194 0.8469
0.9162 1.0000 0.7687 0.7580
0.9194 0.7687 1.0000 0.9446
0.8469 0.7580 0.9446 1.0000
Table 11.2 Correlation of LogReturns: iTraxx (Main and HiVol) and CDX (Main and HiVol). Correlation
iTtraxx Main
iTraxx HiVol
CDX Main
CDX HiVol
iTtraxx Main iTraxx HiVol CDX Main CDX HiVol
1.0000 0.9258 0.4719 0.3339
0.9258 1.0000 0.4398 0.3281
0.4719 0.4398 1.0000 0.8580
0.3339 0.3281 0.8580 1.0000
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
277
30 • Buy protection on iTraxx Europe HiVol on the run (5 years) for 12 125 of the risky exposure. • Sell protection on DJ CDX.NA.IG Main On the run (5 years) for half of the risky exposure. 30 • Buy protection on DJ CDX.NA.IG HiVol On the run (5 years) for 12 125 of the risky exposure.
The risky exposure is actually taken in position in iTraxx Europe Main and DJ CDX.NA.IG Main. By buying some protection on their HiVol components, the actual risky exposure is reduced and one is only exposed to the nonHiVol names in both indices. This is often done to eliminate as much as possible default risk. If a name defaults, it will most likely be a HiVol name. The initial 100M EUR are put on a riskfree bank account at a compound rate of four percent. Suppose the current quotes for the four components of our portfolio are these as given in Table 11.3. We rebalance after regular times, say daily and continue doing this until maturity or until we have a negative cushion at a rebalancing date. In that case, all positions are closed. We can however not pay back the principal amount since the portfolio value is below the bondfloor. This is called gap risk (see Fig. 11.5). One of the aims of the model is to calculate the gap risk or in other words the present value of these gaps. We assume the following correlated VG dynamics for the spreads: St(1) = S0(iTraxxMain) exp(ω1 t + θ1 Gt + σ1 WG(1)t ) St(2) = S0(iTraxxHiVol) exp(ω2 t + θ2 Gt + σ2 WG(2)t ) St(3) = S0(CDXMain) exp(ω3 t + θ3 Gt + σ3 WG(3)t ) St(4) = S0(CDXHiVol) exp(ω4 t + θ4 Gt + σ4 WG(4)t ), where Gt is a common Gamma Process such that Gt ∼ Gamma(t/ν, 1/ν) and Wt(i) are correlated standard Brownian motions with a given correlation matrix ρW = (ρijW ). Table 11.3 Current Quotes for the Four Components of Our Portfolio in bp. t=0 iTtraxx Main iTraxx HiVol CDX Main CDX HiVol
24.625 48.75 37.5 88.5
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
278
FA
Joossens and Schoutens CPPI 105 Bond floor Portfolio
Deleverage 100
Value
95
90 Gap risk
Leverage
85
80
75
0
1
2
3 Time
4
5
6
CPPI Risky exposure 600 Risky exposure 500
Deleverage
400
300
200
100 Leverage 0
0
1
2
3 Time
4
5
6
Figure 11.5 CPPI performance, leveraging, deleveraging, and gap risk in a multivariate VG driven model.
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
279
The logreturns in this model have a correlation structure as in Eq. (11.5). We will first perform a joint calibration on swaptions of the individual indices. This determines the parameters ν, θi , and σi , i = 1, . . . , 4. Next, we match the historical correlations with ρij by setting ρij σi2 + θi2 ν σj2 + θj2 ν − θi θj ν ρijW = . σi σj Hence, we are able to match quite accurately all the individual spread dynamics by correlated jump processes and moreover are able to impose a correlation structure completely matching the observed historical correlation. Indeed for our example, the result of the calibration can be found in Fig. 11.6. In order to match with the required correlation, which we have taken from the logreturn historical correlation from 21 June 2004–13 March 2007 as shown in Table 11.2, we need to set the Brownian correlation matrix equal to 1.0000 0.9265 0.4935 0.3352 0.9265 1.0000 0.4470 0.3247 ρW = 0.4935 0.4470 1.0000 0.8688 . 0.3352 0.3247 0.8688 1.0000 In Fig. 11.7, one sees a typical picture of the correlated moves under the multivariate VG jump dynamics.
11.4.3.2. Gap risk As already mentioned in the previous section, the gap risk under Black’s model under continuous rebalancing is zero. Of course, because the continuous paths of the Brownian Motion, the bond floor is never crossed but at maximum hit. One could artificially rebalance only periodically, say quarterly, in order to generate some gap risk. Much more natural and conform reality is to rebalance continuously (or daily), but to include jump dynamics in the model. Indeed, if jumps are present in the spread dynamics, then the portfolio value can suddenly jump below the bondfloor. Hence, the price of this gap risk is not zero anymore. The price to cover against the gap risk, estimated on 100,000 Monte Carlo simulations of the multivariate VG model in our example for instance is around 5 bp per year. Finally, we want to note that the model is not restricted to a credit setting, but one can pimp the model to a hybrid setting. One can set up multivariateVG dynamics, where for example equity indices and stock dynamics are combined with credit dynamics. This is possible in case fast vanilla pricers are available under a univariate VG model, like is the case for equity vanillas and credit swaptions.
May 12, 2010
280
17:47
WSPC/SPIB913
b913ch11
FA
Joossens and Schoutens
Calibration on swaptions. Figure 11.6
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
FA
281
Portfolio Insurances, CPPI and CPDO, Truth or Illusion? Correlated VC spread paths 120 iTraxx Europe iTraxx Hivol CDX IG CDX Hivol 100
Spreads (bp)
80
60
40
20
0
0
1
2
Figure 11.7
3 time
4
5
6
Correlated VG spread paths.
11.5. RECENT DEVELOPMENTS FOR CPPI CPPIs received and still receive a lot of attention, not only from banks but also from academia. Here, we try to give an overview of the results discussed in a list of papers. Three main topics of research can be identified: the behavior of CPPI, limiting the risk, and building insurances for this risk. The first group of papers study in detail the behavior of CPPI strategies under specific conditions for the underlying portfolio. One way of pricing was already presented in detail in the previous section. The next group of papers try to measure the risk factors involved. In the first paper discussed here, an upper bound for the multiplier m is searched for in such a way that the investment in the risky portfolio is maximized under the condition that the gap risk must stay under a certain limit. The second discussion presents an extended way to calculate the VaR and GVaR of the CPPI portfolio. The last group of papers concentrate on possible ways of extending the CPPI in such a way that an insurance against the small but existing gap risk is build in. The price and size of such an extra insurance will depend on the probability of hitting the floor and hence ways to quantify this risk are also discussed.
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
282
FA
Joossens and Schoutens
For each of the groups, we discuss one example in more detail in the following three subsections.
11.5.1. Portfolio Insurance: The Extreme Value Approach to the CPPI Method In this paper, published by Bertrand and Prigent [18] in 2002, extreme value theory (EVT) is applied to CPPIs. They aim to find a multiplier as high as possible while ensuring that the portfolio value will always be above the floor at a given probability level (typically 99%). If the obtained multiplier is applied, in practice one would maximize its risk under the constraint that the gap risk stayed below a certain level (here 1%). The methods applied in the paper focus on the opposite of the relative jump of the risky asset at time t + 1: Xt+1 =
St − St+1 . St
To determine an upper bound for the multiplier, we only need to concentrate on the positive values of X. First a general and very strong upper bound for m is constructed which is then relaxed considering different assumptions. The proposed methods are then applied to calculate empirical estimates for the upper bound for m using S&P 500 index during the period January 1969–September 1997 as risky assets under the different assumptions. As an initial step, it is shown that in order to keep the cushion positive at all times the upper bound on the multiple needs to be the following: m≤
1 maxk≤n (Xk )
,
where n indicates the size of the dataset. This strong condition is then relaxed by applying quantile hedging. In order to get a result for the multiplier, the possible values of Xk are first truncated to the interval [a, b] and Xk[a,b] denotes the truncated jumps (Xk 1a≤Xk ≤b ) with their corresponding arrival times Tk[a,b] (i.e., the sequence of times at which Xt takes values in the interval [a, b]). The values of a and b will be chosen in such a way that the values of Xk will all fall in the interval. It is assumed that interarrival times of the truncated values are exponentially distributed with parameter λ[a,b] . Under this condition, it is concluded that m≤
1 F [a,b]−1 1 +
ln(1−) λ[a,b] T
,
where T is the length of the time period covered and F [a,b] denotes the distribution function of the variations Xk .
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
283
If the distribution F [a,b] is not known, two different cases can be considered. Either one can assume that no prior information is available or, alternatively, one might be able to fix an upper bound due to anticipation. If no historical information is available, extreme value statistics will be applied on the maximum value of any time period T (0 < T < T ). EVT states that the distribution of the maximum of n observations will converge to a limiting distribution with two parameters. After applying maximum likelihood estimation the parameters of this limiting extreme value distribution, an approximation of the upper bound of m is obtained. In the other case, where some prior information is available, a portfolio manager might fix a maximal drop level b based on his believes. Under this assumption of a maximal drop level, the calculation of the upper bound for m becomes more complicated but the portfolio manager will be able to use a higher multiplier m. Finally, changes of the upper bound under the different assumptions are compared when using S&P 500 index as a risky asset.
11.5.2. VaR Approach for Credit CPPI In [19], the authors discuss the need to build an overnight profit and loss distribution of the portfolio in order to guarantee the capital of a CPPI. Based on this P&L distribution, one could measure the underlying risks and adjust the multiplier in order to limit gap risk. Credit CPPIs strategies invest in risky assets based on an index or single name CDSs, an index or bespoke tranches or spread options. The latest generation of credit CPPIs involve generic portfolios of credit derivatives and the credit index alone might not be a good proxy of the risk. As a solution, it is proposed to measure the impact of each risk factor separately. The different factors identified are spreads, jumptodefault and credit correlation, and the possibility of secondorder risk factors is also included. Those should then be combined to get an idea of the global risk. Next, it is argued that as historical data are often of low quality when looking at portfolios of credit derivatives (such as CDSs but also CDO tranches). It is not recommended to use them in order to compute the risks. It might be possible to find a distribution for each of the risks, using VaR, separately, but not for the sum, hence GVaR will be used for estimating the risks. First, the problem of estimating spreadVaR and defaultVaR for a portfolio of CDSs will be tackled. Here, a model for the spread dynamic is proposed. The correlation caused by the fact that spreads of CDSs are driven by global, regional, and sectoral factors is included in order to estimate the profit and loss coming from spread moves. For the default VaR of a CDS portfolio, the aim is to find the profit and loss coming from overnight defaults. It is assumed that defaults are correlated through a Gaussian copula but no detailed discussion is presented in [19].
May 12, 2010
17:47
WSPC/SPIB913
284
b913ch11
FA
Joossens and Schoutens
To finish the discussion, portfolios with CDOs are included. For the spread VaR and default VaR, the previous results can be reused. Besides those, also the VaR of the profit and loss coming from daily moves of the correlation smile is of importance. Here a modelfree approach is proposed. The above risk measures for each factor are finally combined using GVaR to give a global estimation of risk. This is supposed to be stable and to offer great netting.
11.5.3. CPPI with Cushion Insurance In [20], under the assumption that the dynamics of the value of the risky asset follow the classic diffusion process given in Eq. (11.1), the authors discuss the situation where an investor does not invest the total amount available but only a part of the initial amount, qV0 , into a CPPI. He will use the remainder amount, (1 − q)V0 , to create an extra insurance on the CPPI. This extra insurance will be build in such way that it pays out a value K in case the value of the CPPI hits an adjusted floor Fta = (F0 + a)ert with a > 0 and r the continuous interest rate. As the adjusted floor lies above the initial floor, this protection will avoid creating gap risk up to a certain level. This approach seems to be acceptable as the adjusted floor will, first of all, prevent the CPPI portfolio value coming too close to the floor and hence gap risk will only remain possible in the event of a large downward jump. Second, for the calculation of the risk exposure a manager will only consider an adjusted cushion (= priceadjusted floor), which is smaller than the original cushion. These two factors will make a jump creating gap risk less likely. The first time the CPPI hits the adjusted floor is denoted as Ta1 . If Ta1 > T , only qV0 is invested into the CPPI, otherwise at Ta1 the amount K is paid out and will from then on be invested in the CPPI portfolio together with the current value of the CPPI portfolio (VTa1 ). In order to calculate the costs linked to this system, the authors first calculate the possibility that the CPPI portfolio hits the modified floor under Black’s model, and hence that the investor will receive the amount K. Using this probability (for given values of a, m, and q) together with the principle that expected cost and outcome must be equal, the correct value of K can be computed. Now all the parameters have been obtained and properties of this investment strategy are studied using the “greeks”. Greeks are the quantities representing the market sensitivities, the name is used because the parameters are often denoted by Greek letters. The Delta, for example, measures the sensitivity to changes in the price of the underlying asset. The paper concludes that the Delta of the modified CPPI portfolio is similar to one of the simple CPPI portfolio. Next, the same exercise is performed assuming that the underlying assets dynamics are not given by Eq. (11.1) but by a Lévy process defined as follows dSt = St− [µ(t, St )dt + σ(t, St )dWt + δ(t, St )dlt ],
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
FA
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
285
where (Wt )t is a standard Brownian motion, independent from the Poisson process with measure of jumps (lt )t . It is also assumed that the expected value of jumps in every finite interval [0, t] is finite and equal to λt (λ ≥ 0). The paper concludes by extending the insurance in the sense that not only the first time but every time the CPPI value reaches the adjusted floor, the amount K will be invested in the CPPI portfolio.
11.6. A NEW FINANCIAL INSTRUMENT: CONSTANT PROPORTION DEBT OBLIGATIONS Constant Proportion Debt Obligations (CPDOs) first appeared in August 2006 and are a variation on the CPPI structure. They are used for credit portfolios comprising the exposures to credit indices such as CDX and iTraxx. The CPDO’s risk exposure, just as with the CPPI, is determined using a constant proportion approach and rebalances its portfolio between the credit portfolio and a safe asset. The CPDO structure does this with the aim of producing a highyielding “AAA”rated product. A CPDO funds itself through the issuance of longterm debt paying timely coupon and principal on the notes. The promised coupon is a spread above for example LIBOR. The combination of a high coupon payment and the high rating have made CPDOs very popular products. First, we will try to explain the structure in detail. CPDOs are currently also a popular topic in the news and a small overview is provided the next subsection. This section is concluded presenting an example on rating of CPDOs.
11.6.1. The Structure Constant Proportion Debt Obligations are structures which use, as suggested by their name, a constant proportion approach for their risk exposure and rebalance their portfolio at every time step between the credit portfolio and a safe asset. The CPDO structure takes leveraged exposure to a risky asset by selling protection on individual names or indices (CDS or indices on CDS). The risky exposure ensures that there is enough spread to meet the promised liabilities and also covers the costs and potential losses that the transaction will absorb. Risk positions will be taking in function of the value of the CPDO. If the structure is not performing well, it will increase its risk exposure (up to a predefined maximum leverage level) in order to allow for recovery from the negative performance by increasing the income from the risky asset to rebuild the portfolio’s value. The following factors play a key role in the risk strategies an investor will take: • Net Asset Value (NAV ): This is the current value of the CPDO. It will be the sum of the safe investment and the market value of the risky portfolio.
May 12, 2010
286
17:47
WSPC/SPIB913
b913ch11
FA
Joossens and Schoutens
• Riskfree: The amount of the total CPDO value invested in a riskfree way. • PV(liabilities): The present value of the current liabilities will be the sum of all discounted coupons still to be paid and the discounted value of the final principal amount. • Shortfall: The shortfall is the capital, which is missing to fulfill all the future obligations. It is defined as the difference between the present value of the liabilities and the net asset value, Shortfall = PV(liabilities) − NAV. • CDS premium: The premium is the amount to be paid to the protection seller (such as a CDS or CDS index) in exchange for the insurance. An increasing value of the spread refers to an increased default probability of the underlying asset and hence it will be negative as a default will lead to a decreasing NAV . Conversely, an increasing spread results in a higher income. • PV(CDS premium): The present value of all future premium payments up to maturity. • Leverage: The leverage refers to the degree of risk which will be taken at each time step. A maximum level is often fixed at 15. The leverage is hence defined as
Shortfall , max(Leverage) , (11.6) Leverage = min β PV(CDS premium) where β is a multiplier. Similar as in [21], page 18, in this document β will be fixed at 1/Riskfree. • Cashin: In case the NAV is equal to or exceeds the target value (PV(liabilities)), the necessary amount to cover all future liabilities is reached and hence the risky exposure and leverage will be put to zero. From this point onwards the NAV will be completely invested at the riskfree rate, with coupon and fees being paid until maturity. • Cashout: if there are substantial losses and the NAV falls below a certain threshold (often fixed at 10% of the initial investment), it will be said that cashout has occurred. In such a situation, the CPDO will unwind and the investor will receive the remaining value. The investment strategy follows the following steps. At every timestep, one should check the discounted value of the future obligations (coupon and principal payment), which is the amount which one tries to reach. Next, as the CPDO exposes itself to risk by selling protection (CDSs or CDS indices), the marktomarket of this risky investment is checked and compared to the price paid at the previous time step for this investment. An increase in the spread of the risky CDS means that the underlying insurance has become more expensive and linked to this it has become more likely that the seller of the protection will have to cover a
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
287
future loss. Compared to the previous timestep, this can be considered as a loss for the protection seller, as he appears to be underpaid for the protection it provides, and a gain for the protection buyer. Besides this gain or loss linked to the protection spread, the costs (or incomes) related to the protection also need to be taken into account. Those costs are referred to as the fee which is equal to the sum of all the CDS premium payments made by the protection buyer for the insurance within the time step, taking into consideration the possibility of a default event. The total value of the CPDO (NAV ) at time t + 1 will hence be equal at the accumulated value of the cash, invested risk free at time t, plus the value of the risky asset at time t + 1 augmented with the gains (or losses) made by investing and the fees collected (or paid) for the insurance in the time period [t, t + 1]. From this value, we should subtract the coupon payments which need to be made to the CPDO investor. Based on the above value, and before going to the next step, the new leverage is determined and according to this leverage new investment positions are taken in the risky asset(s). If the NAV increases, the shortfall decreases and hence the leverage will go down; whereas if the NAV decreases, the shortfall will increase and the leverage will go up in order to try to fix the previous negative performance. The above steps are repeated at each time step until cashin or cashout occurs, or until maturity. As described in the definition, cashin occurs when the total value of the CPDO reaches or exceeds the current value of the future obligations. In this case, the seller of a CPDO is sure to be able to fulfil all its future obligations and will from then on only invest in a riskfree way. The probability for a CPDO that such an event takes place will have a big impact on its rating. Cashout occurs when the value of the CPDO hits or shoots below a lower bound which is fixed at time t = 0. If this happens, the CPDO will unwind and the investor will receive the remaining proceed. Such an event could be called the gap risk and is comparable to the gap risk of a CPPI. The risk of a cashout event cannot be excluded, since in case of underperformance more risk will be taken, which increases even more the possibility of arriving below the lower bound when a downward jump occurs. A similar situation as in Sec. 11.3 is used in order to present the CPDO functioning in an example. For the risky asset, a CDSindex is considered with starting spread S0 = 100, and the spreads follow the same Variance Gamma model as before. The riskfree interest rate is fixed at r = 5%. For the CPDO structure, the coupon payment is fixed at r + 2% and the maximum leverage is equal to 15. In this example, a cashout level of 10% of the initial investment is considered. Figure 11.8 presents two examples of possible scenarios of a CPDO. In the left column, the target level is reached before maturity at time t around 8 from that point onwards all cash will be invested in a riskfree way. For the second example, in the right column, the CPDO value will drop
May 12, 2010
288
17:47
0.2
0.06
Spread
Spread
WSPC/SPIB913
0.04 0.1 0.02 0
0 0
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
CPDO Target Cashout
50
0 0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
150 Shortfall
Shortfall 40
100
20
50 0 0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
6
6
Leverage
Leverage
2 0 0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
Figure 11.8 Value of the risky asset (CDS index) (first graph); the corresponding CPDO performance (second); shortfall at each time step (third) and leverage taken at each time (final) in case of a cashin situation (left) and a cashout situation (right).
b913ch11
4
2
FA
Joossens and Schoutens
4
0
10
50
60
0
9 CPDO Target Cashout
100
100
0
0
150
150
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
289
below the cashout level just before time t = 5. This is similar to what, in the CPPI case, we refer to as a gap. If such an event occurs, the CPDO will unwind and the investor will receive all remaining cash. In practice, risk positions will almost always be taken in CDS indices (iTraxx and CDX). The benefits of this choice are summarized on Slide 10 of [22]. However, questions about safety in the sense of the correctness of the high tripleA rating do remain and this sensitive topic will be discussed in the next two subsections. Some literature on CPDOs can be found in [21, 23–28].
11.6.2. CPDOs in the Spotlight Recently, it has become clear that CPDOs are not as safe as is often thought. In real life, cashout events have occurred and these events have also received attention in the media. As a result, the safety of CPDOs has been put up for discussion. On 16th November 2007, for example Moody’s Investors Service downgraded its ratings on six CPDOs, one even to a “junk” rating of Ba2. The downgrading was done because of the continuing spread widening on the financial names underlying these CPDOs. On 28th November 2007, the first CPDO unwinding was announced. This unwinding shows the controversial credit product’s potential for volatility, and moreover it has raised the question of whether the probability they will pay off is as high as implied in a tripleA rating. And also more recently, on 25th January 2008, Moody’s Investors Service confirms that two more series of notes from structured deals backed by financial companies were liquidated after losing investors approximately 90% of their investment. Besides the two unwindings, it also discusses a list of downgrading which has occurred. Many more cases of downgrading and even unwinding have taken place in the last year.
11.6.3. Rating CPDOs Under VG Dynamics A short example has already been presented in this section to explain the dynamics of the CPDO structure. In this section, another example is presented in which positions are taken in two high correlated indices: the iTraxx (iTraxx Europe Main on the run (5 years) unfunded) and CDX (DJ CDX.NA.IG Main on the run (5 years) unfunded). As trading position, we will always invest half of the risky exposure in each index. As presented in [8] and discussed in detail in Sec. 11.4, a Multivariate Variance Gamma model is used to model a series of correlated spreads. First, the calibration is performed using swaption prices based on those indices in order to construct a model
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
290
FA
Joossens and Schoutens
100 iTraxx CDX 50
0
0
1
2
3
4
5
6
7
8
9
10
150 CPDO Target Cashout
100 50 0
0
1
2
3
4
5
6
7
8
9
10
30 Shortfall 20 10 0
0
1
2
3
4
5
6
7
8
9
10
15 Leverage 10 5 0
0
1
2
3
4
5
6
7
8
9
10
Figure 11.9 First graph: value of the risky asset (CDS index); second: the corresponding CPDO performance; third: shortfall at each time step; final: leverage taken at each time.
for their dynamics. The resulting model is then applied to simulate the evolution over time of a CPDO. Assume that the riskfree interest rate is fixed r = 0.04. For the CPDO structure the coupon payment is fixed at r + 2% and the maximum leverage is equal to 15. The initial quotes for our portfolio are iTraxx is quoted 24.625 bp and CDX is quoted 37.5 bp. The top graph of Fig. 11.9 presents a typical picture of the correlated spread evolution under the multivariate VG jump dynamics, which is obtained after calibration. The corresponding path of the CPDO is presented in the second graph and the third and fourth graphs present the corresponding shortfall and leverage positions at each time. In most of the current research documents, it is considered that the spreads follow a Brownian motion which do not allow sudden jumps and hence it is less likely to hit the barrier. If jumps are present in the spread dynamics, the value of the portfolio can suddenly jump under the floor. Based on 10,000 Monte Carlo simulations using the model as obtained from the calibration, we find that the likelihood of occurrence of a cashout event is around 1.94%, which is not as negligible as might be suggested in
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
FA
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
291
the literature. It looks that this percentage is closer to what is also observed in practice, as discussed before.
11.7. COMPARISON BETWEEN CPPI AND CPDO When CPDOs were created, they were considered as a variation of the CPPI. They borrow certain features such as a “constant proportion” approach to determine the leverage and the rebalancing of the portfolio between the credit portfolio and the safe asset. On the other hand, they are also very different. Initially, the CPDO value will be below the target value, while the CPPI manager tends to invest only the amount exceeding the floor, which is needed to make the principal payment at maturity in a risky way. Once decisions on the risky exposure need to be taken the idea is that a CPDO investor will increase its risk as it is performing negatively while the CPPI will decrease its risk as it is not performing well and approaches the floor level. In other words, at each time step, the CPPI investor takes risk exposure positions based on the amount of surplus the portfolio value has with respect to the floor value. The CPDO investor, on the other hand, will at each time take risk exposure proportional to the amount the CPDO portfolio is lacking in order to reach the target value. Once the CPDO value reaches the target value, the manager will stop investing in a risky way as there is sufficient capital to pay out all future liabilities and there is no more need to create capital in a risky way. A CPPI manager will try to optimize its profit but will stop taking risk at the moment that the CPPI value touches the floor, as he is afraid to fall below by taking more risk. Also the behavior at the final stage of the contract is different. A CPPI will, at maturity, irrespective of the performance of the risky asset, receive the principal, together with any positive return generated from the risky asset. In case of loss, when the CPPI portfolio falls below the floor, the losses are covered by the seller of a CPPI so that the investor will still receive the principal. An investor can hence always be sure of receiving the principal and, in the case of good performance, even more. For the CPDO, on the other hand, a target value is aimed for and in the case of a sufficiently well performing risky asset a cashin event will occur and the investor will receive all promised coupon and principal payments. But when the risky asset does not perform well and a cashout event occurs, the CPDO will unwind before maturity. In such a situation only the remaining amount will be paid out to the investor. Investors will want to know the size of this risk at the time of investing in a CPDO and they will use its rating as an indication. This way, the rating of a CPDO becomes important and will have an effect on the price of a CPDO. In practice, the risk of a cashout event has apparently been underestimated.
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
292
FA
Joossens and Schoutens
11.8. CONCLUSIONS As the market of structured credit products grew over the last years, also the request of protection mechanisms in structured credit transactions stays high and hence, also here, a continuous evolution can be observed. CPPI and CPDO are recently developed alternative investment strategies which aim to provide a protection. CPPIs first came into use around 10 years ago and promise a predefined principal payment at maturity. A constant proportion rule is applied to decide the investment strategy. At every time step, the investment in the underlying risky asset and the safe asset is rebalanced in order to optimize the profit. CPDOs were introduced in 2006 and are intended to be safe, highyielding instruments. A similar constant proportion rule is used for their investment strategies and they will invest in a risky asset by selling protection (such as CDX and iTraxx). Similarly to the CPPI, rebalancing will be done at every time step until the targeted value has been reached. The aim of this chapter was to create an in depth view of the dynamics and risks linked to both alternative investment strategies. Hence, first we have tried to explain step by step how they both function and how investment decisions are made. Understanding the dynamics well helps to identify remaining “safety gaps” and allows an investor to get an idea about the size and possibility of experiencing such a gap. For the CPPI, the possibility that the value of the total portfolio will fall below the floor exists and will create a loss, while for the CPDO a loss occurs when a cashout event occurs. In both cases, there is a strong interest in quantifying this risk. In a first group of papers, the researchers concentrate on quantifying the risk using specific conditions. Next, some propose ways to limit the multiplier factor in order to limit the risk, while other papers suggest the possibility of investing on an insurance to cover this risk. But also for those new developments a good and robust way of quantifying the risk is necessary. It could be concluded that the gap risk for CPPIs should not be neglected but safety nets can be used to avoid suffering from it. As CPDOs are still very new, the field of research is still limited, and as they have only been used on the market for a couple of years, their performance in the real world has only been observed over a short time. As recent experience has shown, in real life CPDOs do not seem to be as safe as they were expected to be. Clearly, there is still a strong need to quantify the risk of cashout events in a more realistic manner, and a great deal of research remains to be done in this field.
References [1] Schoutens, W (2003). Lévy Processes in Finance: Pricing Financial Derivatives. John Wiley & Sons.
May 12, 2010
17:47
WSPC/SPIB913
b913ch11
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
FA
293
[2] Cariboni, J (2007). Credit Derivatives Pricing Under Lévy Processes. PhD thesis, Katholieke Universiteit Leuven. [3] Madan, D and W Schoutens (2007). Break on through to the single side. Technical Report, Section of Statistics Technical Report 0707, (2007). URL http://perswww.kuleuven.be/ ∼u0009713/LevyCDS.pdf. [4] Overhaus, M, A Bermudez, H Buehler, A Ferraris, C Jordinson and A Lamnouar (2007). Equity Hybrid Derivatives. John Wiley & Sons. [5] Black, F and R Jones (1987). Simplifying portfolio insurance. Journal of Portfolio Management, 48–51. [6] Perold,A (1986). Constant portfolio insurance. Unpublished manuscript, Harvard Business School. [7] Cont, R and P Tankov (2007). Constant proportion portfolio insurance in presence of jumps in asset prices, SSRN eLibrary. URL http://ssrn.com/paper=1021084. [8] Garcia, J, S Goossens and W Schoutens (2008). Let’s Jump Together: Pricing of Credit Derivatives. Risk Magazine, September. [9] Madan, D and E Seneta (1990). The variance Gamma model for share market returns. Journal of Business, 63, 511–524. [10] Bertoin, J (1996). Lévy Processes. Vol. 121, Cambridge Tracts in Mathematics, Cambridge: Cambridge University Press. [11] Sato, K (2000). Lévy Processes and Infinitely Divisible Distributions. Cambridge University Press. [12] Seneta, E (2007). The early years of the variancegamma process. Advances in Mathematical Finance. Boston: Birkhäuser. [13] Cariboni, J and W Schoutens (2004). Pricing credit default swaps under Lévy models. UCS Report 200502, K. U. Leuven. [14] Luciano, E and W Schoutens (2006). A multivariate jumpdriven financial asset model. Quantitative Finance, 6(5), 385–402. [15] Leoni, P and W Schoutens (2008). Multivariate smiling, Wilmott Magazine, March Issue. [16] Pederson, C (2004). Introduction to Credit Default Swaptions, Lehman Brothers. [17] Carr, P and D Madan (1999). Option pricing and the fast fourier transform. Journal of Computational Finance, 2(4), 61–73. [18] Bertrand, P and J Prigent (2002). Portfolio insurance: The extreme value approach to the CPPI method. Finance, 23(01A13), 68–86. [19] Brun, J and L Prigneaux (2007). VaR approach for credit CPPI and counterparty risk. Quant Congress, Quant Congress. [20] Prigent, JL and F Tahar (2005) CPPI with cushion insurance. SSRN eLibrary. URL http://ssrn.com/paper=675824. [21] Standard&Poor’s (2007). Quantitative modeling apporach to rating index CPDO structures. Technical Report, Standard&Poor’s. [22] ABNAMRO (2007). Surf CPDO: A breakthrough in synthetic credit investments. Structured Products ForumTokyo. [23] Varloot, E, G Charpin and E Charalampidou (2006). CPSO an asset class on its own or a glorified bearish rated equity, UBS Investment Research — European Structured Credit. [24] Lucas, D and R Manning (2007). A CPDO primer. UBS Investment Research — CDO Insight. [25] Varloot, E, G Charpin and E Charalampidou (2007). CPDO insights on rating actions and manager impacts using a new formula. UBS Investment Research — European Structured Credit.
May 12, 2010
294
17:47
WSPC/SPIB913
b913ch11
FA
Joossens and Schoutens
[26] Cont, R and C Jessen (2008). Constant propoertion debt obligations. Financial Risks Inernational Forum. [27] Saltuk, Y and J Goulden (2007). CPDOs and the upcoming roll. Technical Report, JP Morgan. [28] Linden, A, M Neugebauer, S Bund, J Schiavetta, J Zelter and R Hardee (2007). First generation CPDO: Case Study on Performance and Rating. Structured credit global special report, Derivative Fitch.
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
12
ON THE BENEFITS OF ROBUST ASSET ALLOCATION FOR CPPI STRATEGIES
KATRIN SCHÖTTLE∗,‡ and RALF WERNER†,§ ∗
MEAG MUNICH ERGO AssetManagement GmbH, OskarvonMillerRing 18, 80333 München, Germany † Hypo Real Estate Holding AG, Unsöldstrasse 2, 80538 München, Germany ‡ katrin.schoettle@gmx.de § werner_ralf@gmx.net
In recent years, new ideas for the robustification of the traditional Markowitz frontier have appeared in the literature. Based on one of these ideas — the socalled robust counterparts — we introduce the concept of the robustified efficient frontier. As mean– variance efficient portfolios are frequently used as risky assets for CPPI strategies, we investigate the behavior of such strategies under estimation risk. Based on a toy example, we explain the main idea how the concept of a robustified frontier can be used to improve the performance of CPPI strategies. For this purpose, we compare the theoretical performance of CPPI strategies based on the original and the robust efficient frontier.
295
May 12, 2010
17:47
WSPC/SPIB913
296
b913ch12
FA
Schöttle and Werner
12.1. MOTIVATION During the last years, dynamic investment strategies have gained more and more importance in quantitative asset management. Until now, traditional asset management was mainly based on the choice of a strategic asset allocation (SAA), which has been fixed for one year in most cases. Nowadays, there is a move toward dynamic asset allocation (DAA), mainly driven by portfolio insurance concepts and the influence of option pricing theory on investment strategies, like the bestoftwo concept of [9]. Further, it has been observed that the utilization of a dynamically rebalanced allocation allows for a bettertailored risk/return profile for the investor, as described in [1]. In the following study, we focus on one of the most popular portfolio insurance concepts — a constant proportion portfolio insurance (CPPI) strategy. This concept was probably first introduced by Perold in 1986, see [24] (see also [25] as well as [5]), followed by a large variety of theoretical and empirical studies, comparison to alternative strategies, etc. For a recent and extensive overview of related work, we refer to the book [27], Chapter 9 and the references therein. Meanwhile, the properties as well as the pros and cons of CPPI like strategies are well known and discussed in detail, see also [27], Chapter 9. However, in our opinion, one rather relevant practical issue has not yet found much attention by researchers — the influence of estimation risk on the performance of CPPI strategies. In contrast to this observation, there is a good amount of research on the influence of estimation risk on static oneperiod mean– variance optimization, see [8] and the references contained therein. Hence, we focus on the investigation of the influence of this estimation risk on the performance of CPPI strategies as well as a mitigation of this risk by the robustification of mean–variance efficient portfolios. The remainder of this paper is organized as follows: first, we introduce the traditional CPPI strategy. A few numerical examples will show that the choice of the risky asset has a significant impact on the investment performance of the CPPI strategy and that especially the allocation according to mean–variance optimal portfolios adds value to the strategy. Thereafter, it is shown that this behavior may deteriorate in the presence of estimation risk. Therefore, robustified mean–variance efficient portfolios are introduced and their impact on the investment performance is investigated.
12.2. THE FINANCIAL MARKET In contrast to most academic research, our exposition is based on a discrete time framework instead of a continuous time setting. This setup was chosen for several reasons. First, in continuous time, CPPI strategies only exhibit a shortfall risk in the presence of discontinuous price paths (i.e., in jump models). Second, estimation risk is much
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
Robust Asset Allocation for CPPI Strategies
FA
297
more important in discrete time. Third and most important, practical implementations are always based on discrete time setups.
12.2.1. The Basic Financial Market In order to properly formulate the standard CPPI strategy and to highlight the aforementioned issues with this strategy, we introduce the following basic financial market. Definition 12.1. Let us assume that we are given a discrete time arbitragefree financial market (, F, F, P) with time steps t and final horizon T . Further, assume that on this market, there are I tradeable assets, e.g., stocks, bonds, or equity indices, with I prices {St }t=0,...,T , St ∈ R+ following some adapted (discrete time) stochastic process. This process will be specified in more detail in Assumption 12.3. Remark 12.2. Please note that the probability measure P represents the true real world measure and should not be confused with some martingale measure. Furthermore, we do not assume that the market is complete. Assumption 12.3. For the remainder of this paper, we will assume that the discrete oneperiod asset returns St are i.i.d. in time and follow a multivariate elliptically contoured distribution, i.e., Si,t − Si,t−t (12.1) St ∼ E(µ, , ψ) with Si,t := Si,t−t I independent of t, with parameters µ ∈ R+ , ∈ SI+ , and characteristic generator ψ with 2ψ (0) = −1. Further, we assume that St is bounded, i.e.,
St ∈ [µ − 1 , µ + 1 ],
especially St > −11.
Since St is bounded, it possesses finite moments, which means that E[St ] = µ,
Cov[St ] = −2ψ (0) =
and the characteristic function of St −µ is given by u → ψ(u u). For more details on elliptical distributions, we refer to [11] or [10]. Remark 12.4. In Assumption 12.3, the oneperiod mean µ and covariances are used to describe the distribution of St . However, if not explicitly stated otherwise, all tables and figures are scaled to annualized numbers through division by t. Remark 12.5. We need to point out that the usual assumption of multivariate norI mal (or student t) returns contradicts the assumption of St ∈ R+ . We have cured this by restricting the distribution of St to [µ − 1 , µ + 1 ]. For practical purposes, P[Si,t ≤ −1] is negligibly small anyway for sufficiently small t, even if we were in a multivariate normal or student t framework.
May 12, 2010
17:47
WSPC/SPIB913
298
b913ch12
FA
Schöttle and Werner
As we want to focus on estimation risk, we assume in the following that the characteristic generator ψ is known (i.e., no model risk is involved) but that the parameters µ and are not given a priori. Hence, one has to rely on estimators for these quantities. Therefore, we will only consider parameter uncertainty and no model uncertainty for the ease of this exploration. For the formulation of the standard CPPI strategy, we need to make some additional assumptions about the financial market. First, we introduce two riskless assets in the financial market; a cash account and a zero bond. Second, we define what in practice is known as a risky asset. Whereas the cash account and the zero bond enhance the financial market by two additional assets, all risky assets will be completely reproducible by the existing tradeable assets (see Formula (12.2) below) and do therefore not enlarge the market.
12.2.2. The Riskless Asset Assumption 12.6. As mentioned above, on top of the described assets, we need the following additional (locally) riskless assets: • The market includes a cash account, also called (locally) riskless asset. For brevity and notational convenience, we denote the process of the cash account with Ct instead of including it in the already defined process St . The main feature of the cash account is that Ct+t = exp(rt t)Ct ,
t = 0, . . . , T − t,
where the oneperiod rate rt is assumed to be known at time t, hence Ct+t is Ft measurable, i.e., predictable. In this sense, we may speak of the cash account as locally riskless asset. • The market is equipped with a zero bond Zt with maturity T , i.e., ZT = 1. Although the price evolution of this zero bond is stochastic in its very nature, zero bonds are still subsumed within the class of riskless assets as the payment at maturity is known in advance. We will see in the following that for the standard CPPI strategy, only the zero bond Z is actually necessary, the cash account C will only become important when dealing with the problem of intertemporal risk budgeting or in case of general liabilities, see Remark 12.20.
12.2.3. The Risky Asset As risky asset, we understand a basket consisting exclusively of the original I assets building the basic financial market. In general, the risky asset is assumed
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
299
Robust Asset Allocation for CPPI Strategies
to have a constant asset mix, i.e., it is described by a constant vector of weights I , w1 = 1 (i.e., no short sales are allowed). The risky asset is usually deterw ∈ R+ mined by the client of the asset manager or the asset manager herself and remains fixed throughout the whole investment period. In order to obtain fixed weights, a permanent rebalancing in this risky asset has to be assumed. In this sense — as already mentioned above — the risky asset can be seen as a mutual fund, which is constructed within the financial market. The risky asset is usually established at the inception of the CPPI strategy and remains fixed throughout the life time. Under the assumption of rebalancing at each time t, the price process {Xt }t=0,...,T of the risky asset is given by Xt+t = Xt · (1 + wSt+t ) t = 0, . . . , T − t.
(12.2)
In analogy to St in Eq. (12.1), we set Xt =
Xt − Xt−t , Xt−t
and
µX := E[Xt ] = w µ, σX2 := Cov[Xt ] = ww. I Please note that due to w ∈ R+ it holds that Xt is bounded as well, i.e., Xt ∈ [µX − 1, µX + 1], and especially Xt > −1.
Remark 12.7. In practice, there are a few different ways how the asset allocation for the risky asset can be determined: • The most simple way uses the naïvely diversified portfolio, i.e., w=
1 1 ,..., I I
1 = 1, I
i.e., wi =
1 . I
This portfolio is often used as a benchmark against which other methodologies are tested and compared. The advantage of this choice is that it does not rely on any estimate of future returns or volatilities (i.e., covariance matrix of asset returns). • More sophisticated allocations are based on the traditional Markowitz optimization. At the inception of the CPPI strategy (i.e., at t = 0), an asset allocation is chosen from the efficient frontier based on mean–variance optimization. For this approach, estimates for the expected returns µ and the corresponding covariances need to be available. Given these estimates, superior investment results should be obtained due to the usage of efficient portfolios. The most popular choices are — The global minimum variance portfolio — The maximum Sharpe ratio portfolio
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
300
FA
Schöttle and Werner
More details on the setup of the mean–variance optimization will be given in the next subsection. Besides these initially fixed asset allocations, dynamically updated allocations are also possible within the framework of CPPI strategies. As these are also subject to the same estimation issues as the above myopic portfolio choice, we later focus on this simple myopic setup.
12.2.4. Classical Mean–Variance Analysis The traditional mean–variance optimization was first introduced by Markowitz in 1952, see [21]. The basic idea is that a portfolio is solely characterized by the two figures risk (mostly measured in terms of the variance or volatility) and expected return. Since an investor is seeking for an allocation with little risk and high expected return, a tradeoff between these two conflicting aims has to be made. In many practical applications, the set of assets describing the financial market is supposed to stem from a multivariate normal distribution, where the expectation and covariance matrix thereof are then used to express the risk and return of the portfolios in the optimization problem. The Markowitz approach is naturally applicable as well in the more general case of multivariate elliptical distributions, see [15]. In the following, we use the Markowitz optimization framework to determine particular portfolios from the efficient frontier, which are then used as underlying risky asset in CPPI strategies. The asset universe under consideration consists of the initial I tradeable assets only. As we will see, the move to efficient portfolios of tradeable assets has a beneficial impact on the performance of the CPPI strategy. To determine all efficient portfolios, i.e., portfolios lying on the efficient frontier, we consider the following family of portfolio optimization problems: Definition 12.8. The classical mean–variance portfolio optimization problem is given by √ min (1 − λ) ww − λw µ, (Pλ ) w∈W
where W ⊂ {w ∈ RI  w1 = 1} is assumed to be nonempty, convex, and comI  w1 = 1}, which will be pact. This is, for example, fulfilled by W = {w ∈ R+ used in the following. The parameter λ ∈ [0, 1] expresses the tradeoff between risk (i.e., volatility) and return of the portfolio. The optimal solutions w∗cl (λ) of (Pλ ) for given tradeoff parameter λ span the efficient frontier w∗cl (λ)w∗cl (λ), w∗cl (λ) µ 0≤λ≤1
in the (σ, µ)diagram, see Fig. 12.1.
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
301
Robust Asset Allocation for CPPI Strategies 0.11
Excess return (p.a.)
0.105
0.1
0.095
0.09
Efficient frontier Min. variance portfolio Naive portfolio Max. sharpe ratio portfolio
0.085
0.08 0.13
0.14
0.15
0.16
0.17 0.18 Volatility (p.a.)
0.19
0.2
0.21
0.22
Figure 12.1 Markowitz efficient frontier and selected portfolios.
Remark 12.9. Based on the assumption that W ⊂ {w ∈ RI  w1 = 1} and since is positive definite, it is easy to show that there exists at least one solution for 0 ≤ λ < 1 and that the solution w∗cl (λ) is indeed unique, see [29]. Remark 12.10. Two portfolios have been of particular interest thus far: (1) The minimum variance portfolio w∗cl (0), i.e., the portfolio at the left end of the frontier where λ = 0 (2) The maximum Sharpe ratio portfolio — in this context — given by w∗cl (λS ) where λS = arg max λ∈[0,1]
w∗cl (λ) µ w∗cl (λ)w∗cl (λ)
.
We have illustrated both portfolios together with the naïve portfolio in Fig. 12.1. Remark 12.11. Under the above assumptions, problem (Pλ ) is equivalent to the classical Markowitz formulation max w∈W
w µ
(MVσ )
s.t. ww ≤ σ 2 , where σ and λ are in a monotonous onetoone relationship. For an exact definition of the equivalence, we refer to [32].
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
302
FA
Schöttle and Werner
12.2.5. The Trading Strategy For the mathematical definition of a selffinancing trading strategy in the above market setup, let us refer to [12], Sec. 5.1. Our setup differs only in notation from their framework, as we prefer that at time t the portfolio composition (holdings or lot sizes) ϕt ∈ RI+3 shall denote the holdings after trading at time t. This means that the portfolio allocation remains constant throughout the period [t, t +t[ and equals ϕt . A particular trading strategy is thus determined by the way the portfolio composition ϕt ∈ RI+3 at time t is determined based on known quantities at time t. Seen as a stochastic process, this means that ϕt is an Ft measurable process, i.e., adapted. For notational convenience, we split the trading strategy ϕt ∈ RI+3 into several components, based on the underlying assets: ϕt = ϕtS , ϕtC , ϕtZ , ϕtX ∈ RI × R × R × R for the first I tradeable assets S, the cash acount C, the zero bond Z, and the risky asset X, resp. Remark 12.12. In the following, the CPPI strategy will only invest in the zero bond and in the risky asset, i.e., we have ϕtS = 0 as well as ϕtC = 0. Nevertheless, the tradeable assets S are required to construct the risky asset X. Definition 12.13. Let us further introduce a few notations: • The corresponding portfolio value at time t is given as Vt := ϕtS St + ϕtC · Ct + ϕtZ · Zt + ϕtX · Xt . • Let us denote the difference in the holdings before and after trading, the traded lots, as ϕt := ϕt − ϕt−t . • The corresponding trading volume at time t can be computed as vt :=
I
S ϕi,t  · Si,t + ϕtC  · Ct + ϕtZ  · Zt + ϕtX  · Xt .
i=1
12.3. THE STANDARD CPPI STRATEGY Based on the riskless and the risky asset, we can define the main features of a CPPI strategy. Its core idea lies in the dynamic allocation of capital between the riskless and the risky asset. The proportion of the riskless asset in the portfolio is chosen in such a way that a prespecified minimum level of wealth is guaranteed at time T . This final
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
303
Robust Asset Allocation for CPPI Strategies
minimum level of wealth Vmin is called floor. In general, it is expressed in absolute terms or relative to the initial wealth V0 : Vmin := f · V0 ,
with 0 < f
E0 = V0 − L0 > 0 and therefore ϕ0X =
E0 , X0
ϕ0Z =
V0 − E0 L0 = = Vmin . Z0 Z0
A closer investigation of the simple CPPI strategy yields that — independent of the returns of the risky asset or the zero bond — the lot sizes ϕtZ remain constant over time: ϕtZ = ϕ0Z = Vmin . Therefore, since the strategy is selffinancing, it must also hold that ϕtX = ϕ0X . This means that in its simple version the CPPI strategy does not need to rebalance the portfolio — here we mean rebalancing between riskless and risky asset, the risky asset itself may of course need some rebalancing. Remark 12.15. In general, the above simple CPPI strategy is not very well suited to achieve the returns exceeding the risk free rate. Table 12.1 illustrates that although the underlying risky asset has a rather broad annual return distribution ranging at least from −0.20 to 0.50, the annual return of the CPPI remains within tight bounds from 0.01 to 0.06. This is immediately clear from the property of the simple CPPI strategy that it is just a buyandhold strategy between the zero bond and the risky asset, i.e., at each time t the fraction fZ0 is invested in the zero bond, whereas only 1 − fZ0 is
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
304
FA
Schöttle and Werner Table 12.1
Performance of Selected Rebalancing and CPPI Strategies.a
Strategy
E
Std
Skewness
Excess kurtosis
1% pctl.
99% pctl.
Reb. naïve Reb. minvar Reb. maxsharpe CPPI naïve CPPI minvar CPPI maxsharpe Zero bond
0.1192 0.1196 0.1290 0.0374 0.0374 0.0382 0.0305
0.1710 0.1505 0.1586 0.0134 0.0117 0.0124 —
0.4524 0.4009 0.4198 0.4524 0.4009 0.4198 —
0.3603 0.3275 0.3217 0.3603 0.3275 0.3217 —
−0.2255 −0.1903 −0.1953 0.0105 0.0132 0.0128 —
0.5771 0.5163 0.5502 0.0731 0.0684 0.0710 —
a This
table gives the average performance of the three risky assets and the zero bond in comparison to those of the corresponding three simple CPPI strategies (with f = 0.95) based on a Monte Carlo simulation with 25,000 random paths. Besides the expected performance (E), we have added the standard deviation (std), the skewness, and the excess kurtosis as well as the 1% and 99% percentile of VT /V0 − 1 for a better comparison.
invested in the risky asset: L0 V0 − L0 V0 − Z0 · Vmin Vt = · Zt + · Xt = Vmin · Zt + · Xt Z0 X0 X0 Zt Xt . = V0 · fZ0 + (1 − fZ0 ) Z0 X0 Nevertheless, using more information on the distribution of the return of the risky asset, the CPPI strategy can be refined to obtain better investment returns. Remark 12.16. The simulation and optimization setup for the results in Table 12.1 and all subsequent calculations is as follows: • The continuous risk free interest rate is assumed to be 3% p.a., i.e., the zero bond has a discrete return of r = 3.05% p.a. • The excess returns of six individual tradeable assets over the riskfree return of the zero bond are given by µann,exc = 4.68%, 10.56%, 6.36%, 10.56%, 9.48%, 8.52% , expressed in annualized terms. • The (annualized) covariance matrix of the financial assets is 0.0359 0.0188 0.0122 0.0115 0.0228 0.0199 0.0188 0.0588 0.0322 0.0247 0.0272 0.0158 0.0122 0.0322 0.0460 0.0182 0.0214 0.0108 ann = . 0.0115 0.0247 0.0182 0.0590 0.0202 0.0079 0.0228 0.0272 0.0214 0.0202 0.0431 0.0172 0.0199 0.0158 0.0108 0.0079 0.0172 0.0220
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
Robust Asset Allocation for CPPI Strategies
305
• The simulations of the discrete oneperiod asset returns are based on a normal distribution N (µt, t) with µ = µexc + r. The normal distribution has been cut off to guarantee Assumption 12.3 in the sense of Remark 12.5 simply by neglecting any violating scenario.
12.3.2. The General Case The main idea of the standard CPPI strategy is to leverage the investment in the risky asset and to actively balance between the risky asset and the zero bond. The leverage factor m, by which the investment in the risky asset is increased in comparison to the simple version, is usually called multiplier. The introduction of this multiplier m leads to the CPPI strategy in its standard form: Algorithm 12.17. Given all the necessary information at time t, the standard CPPI strategy can be formulated as (1) (2) (3) (4)
Compute the present value of the floor Lt = Zt · Vmin = f · Zt · V0 . Compute the cushion or risk budget Bt := (Vt − Lt )+ . Compute the leveraged exposure as Et := min(mBt , Vt ). Invest Vt − Et in the zero bond and Et in the risky asset, i.e., Et V t − Et ϕtX = , ϕtZ = . Xt Zt
Remark 12.18. In the above version, the exposure is bounded by the current portfolio wealth, however, it is possible to introduce an upper boundary level 0 < b < 1 and to restrict Et by Et = min(mBt , bVt ). Remark 12.19. The cap of Et by Vt in the standard CPPI strategy restricts an investment in the risky asset beyond the wealth of the strategy, i.e., short positions in the zero bond are prohibited. Note that this cap was not necessary in the simple version of the CPPI strategy. Remark 12.20. In more general settings, Lt is not linked to a deterministic zero bond investment any more, but given by a more general stochastic liability (hence the abbreviation Lt ). As long as this liability Lt is investable, e.g., by socalled replicating portfolios, CPPI strategies can still be formulated. For example, if a money market fund is to be beaten, f · Ct may be chosen as liability. As we see from Fig. 12.2 and Table 12.2, the multiplier m has a significant influence on the performance of the investment strategy. We can observe that the average return E[VT /V0 − 1] increases with increasing m up to a certain level. From there onwards, no significant increase in average performance is possible. We can also observe that the standard deviation behaves accordingly.
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
306
FA
Schöttle and Werner 0.18 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0
Expected Return Volatility 5
10
15 Multiplier m
20
25
30
Figure 12.2 Average return and standard deviation of the standard CPPI strategy on the naïve risky asset with f = 0.95 for different multipliers m.
Table 12.2 Descriptive Statistics of the Standard CPPI Strategy on the Naïve Risky Asset with f = 0.95 for Different Multipliers m. Multiplier m=1 m=3 m=5 m=7 m=9 m = 11 m = 13 m = 15 m = 17 m = 19 m = 21 m = 23 m = 25 m = 30
E
Std
Skewness
Excess kurtosis
0.0374 0.0531 0.0707 0.0837 0.0909 0.0944 0.0961 0.0969 0.0973 0.0976 0.0978 0.0979 0.0980 0.0981
0.0134 0.0493 0.0972 0.1288 0.1451 0.1537 0.1585 0.1613 0.1632 0.1645 0.1655 0.1662 0.1667 0.1677
0.4524 1.4819 1.9288 1.5671 1.3436 1.2221 1.1504 1.1012 1.0673 1.0417 1.0227 1.0076 0.9965 0.9761
0.3603 3.8084 4.8104 2.5131 1.5518 1.1190 0.8890 0.7365 0.6338 0.5610 0.5093 0.4701 0.4412 0.3899
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
307
Robust Asset Allocation for CPPI Strategies
Remark 12.21. This behavior can easily be explained by a deeper investigation of Algorithm 12.17, Step (3). If the multiplier m is increased step by step, the leveraged exposure Et will equal Vt as long as there is at least a cushion of Vt /m — which diminishes to 0 if m → ∞. This means that the CPPI strategy is fully invested in the risky asset until the risk budget is lost. As soon as Bt ≤ 0, the investment is completely shifted from the risky asset to the riskless asset. In other words, for m → ∞, the CPPI strategy converges to a simple stoploss strategy, which invests into the risky asset until the wealth is below the floor Lt . As soon as the cushion is lost, the CPPI strategy cannot recover as it is fully invested in the zero bond for the rest of the time and thus VT ≤ LT = Vmin . In contrast to the simple CPPI strategy, for m > 1, it might happen that VT ≤ LT , i.e., the floor is no longer guaranteed. This probability of a shortfall is investigated in more detail in the next subsection. Please note that this cannot happen in continuous time as long as the price paths are continuous, see [27]. In addition to the multiplier m, also the floor f has a significant impact on the performance and the risk of the strategy. In Tables 12.3 and 12.4, an overview of the dependence is given. It can be observed that with increasing multiplier the expected return of the CPPI strategy increases, since more capital can be invested in the risky asset which accounts for higher returns. The same line of argumentation holds for the floor: the lower the floor, the more budget for riskier investments and thus the higher the expected return. Table 12.3 Expected Return of the Standard CPPI Strategy on the Naïve Risky Asset for Different Multipliers m and Different Floors f .
m=1 m=3 m=5 m=7 m=9 m = 11 m = 13 m = 15 m = 17 m = 19 m = 21 m = 23 m = 25 m = 30
f = 0.90
f = 0.925
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.0417 0.0672 0.0913 0.1029 0.1072 0.1091 0.1099 0.1103 0.1105 0.1107 0.1107 0.1108 0.1109 0.1110
0.0395 0.0602 0.0816 0.0946 0.1005 0.1031 0.1043 0.1050 0.1053 0.1055 0.1057 0.1057 0.1057 0.1058
0.0374 0.0531 0.0707 0.0837 0.0909 0.0944 0.0961 0.0969 0.0973 0.0976 0.0978 0.0979 0.0980 0.0981
0.0352 0.0461 0.0587 0.0701 0.0776 0.0819 0.0844 0.0854 0.0858 0.0858 0.0858 0.0859 0.0859 0.0861
0.0339 0.0419 0.0512 0.0606 0.0677 0.0721 0.0747 0.0761 0.0768 0.0769 0.0770 0.0767 0.0765 0.0764
0.0331 0.0390 0.0461 0.0537 0.0600 0.0644 0.0670 0.0684 0.0692 0.0696 0.0696 0.0694 0.0693 0.0694
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
308
FA
Schöttle and Werner
Table 12.4 Standard Deviation of the Standard CPPI Strategy on the Naïve Risky Asset for Different Multipliers m and Different Floors f .
m=1 m=3 m=5 m=7 m=9 m = 11 m = 13 m = 15 m = 17 m = 19 m = 21 m = 23 m = 25 m = 30
f = 0.90
f = 0.925
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.0216 0.0799 0.1366 0.1588 0.1660 0.1687 0.1701 0.1710 0.1715 0.1720 0.1723 0.1726 0.1728 0.1731
0.0175 0.0646 0.1191 0.1468 0.1587 0.1638 0.1664 0.1680 0.1691 0.1698 0.1704 0.1708 0.1712 0.1718
0.0134 0.0493 0.0972 0.1288 0.1451 0.1537 0.1585 0.1613 0.1632 0.1645 0.1655 0.1662 0.1667 0.1677
0.0092 0.0340 0.0706 0.1032 0.1229 0.1347 0.1422 0.1470 0.1503 0.1526 0.1541 0.1553 0.1563 0.1579
0.0067 0.0248 0.0527 0.0827 0.1038 0.1172 0.1260 0.1321 0.1364 0.1394 0.1417 0.1434 0.1448 0.1469
0.0051 0.0187 0.0400 0.0665 0.0875 0.1017 0.1114 0.1180 0.1229 0.1265 0.1292 0.1313 0.1330 0.1361
Analogously, with larger investments in the risky asset (i.e., higher multiplier and/or lower floor), the volatility also increases.
12.3.3. Shortfall Probability of CPPI Strategies Definition 12.22. For a given CPPI strategy according to Algorithm 12.17 with parameters f and m, the shortfall probability SP(f, m) and the expected shortfall ES(f, m) are defined as SP(f, m) := P[VT ≤ LT ] = P[VT ≤ Vmin ], ES(f, m) := E[VT − LT  VT ≤ LT ]. The corresponding stopping time τ := inf{t  Vt ≤ Lt } is called shortfall time. In case no shortfall is experienced by the CPPI strategy, we have τ = ∞. We say that the CPPI strategy has reached the absorbing state if Vt ≤ Lt , as from then onwards the strategy is fully invested in the zero bond. Tables 12.5 and 12.6 illustrate the interrelation of the shortfall probability and the expected shortfall with the multiplier m and the floor f for a fixed risky asset (the naïve one). As expected, it can be observed that the shortfall probability increases with increasing multiplier m. The floor influences the shortfall probability in a similar way: the
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
309
Robust Asset Allocation for CPPI Strategies
Table 12.5 Shortfall Probability of the Standard CPPI Strategy on the Naïve Risky Asset for Different Multipliers m and Different Floors f .
m=1 m=9 m = 11 m = 13 m = 15 m = 17 m = 19 m = 21 m = 23 m = 25 m = 30
f = 0.90
f = 0.925
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0 0 0 0.08% 0.48% 1.82% 3.80% 6.68% 9.16% 11.64% 15.86%
0 0 0.02% 0.16% 0.70% 2.62% 5.78% 10.10% 13.74% 17.56% 24.04%
0 0 0.02% 0.26% 1.14% 3.88% 8.30% 14.22% 19.62% 24.94% 33.42%
0 0 0.02% 0.28% 1.78% 5.62% 12.00% 20.12% 28.22% 35.44% 46.94%
0 0 0.02% 0.32% 2.08% 6.60% 14.22% 24.08% 34.06% 43.26% 56.54%
0 0 0.02% 0.38% 2.32% 7.32% 16.00% 27.20% 38.52% 48.54% 63.02%
Table 12.6 Expected Shortfall of the Standard CPPI Strategy on the Naïve Risky Asset for Different Multipliers m and Different Floors f .
m=1 m=9 m = 11 m = 13 m = 15 m = 17 m = 19 m = 21 m = 23 m = 25 m = 30
f = 0.90
f = 0.925
f = 0.95
f = 0.975
f = 0.99
f = 1.00
— — — −0.0008 −0.0030 −0.0025 −0.0026 −0.0026 −0.0027 −0.0028 −0.0034
— — −0.0015 −0.0030 −0.0026 −0.0022 −0.0023 −0.0023 −0.0025 −0.0027 −0.0033
— — −0.0011 −0.0022 −0.0029 −0.0024 −0.0023 −0.0023 −0.0025 −0.0027 −0.0034
— — −0.0008 −0.0016 −0.0022 −0.0022 −0.0021 −0.0021 −0.0022 −0.0024 −0.0031
— — −0.0006 −0.0012 −0.0018 −0.0018 −0.0019 −0.0020 −0.0022 −0.0025 −0.0031
— — −0.0004 −0.0022 −0.0018 −0.0015 −0.0017 −0.0019 −0.0021 −0.0023 −0.0031
closer the floor to the initial wealth, the higher the shortfall probability. Further, the expected shortfall remains approximately constant. Remark 12.23. In practice, the multiplier m is deduced from the consideration of the shortfall risk by a Monte Carlo simulation. The multiplier is chosen as large as possible to obtain a sufficient average investment return, while keeping a predetermined level of shortfall risk, i.e., low shortfall probability and low expected shortfall. For this purpose, practitioners usually run simulation studies like the above to obtain a suitable multiplier. It should be noted that there are analytical approximations available
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
310
FA
Schöttle and Werner
(see [27]), which link the multiplier to the volatility or percentiles of the oneperiod return distribution. However, these approximations are only valid if Et = mBt , i.e., for sufficiently small multipliers, and do not longer hold if Et is capped at Vt . In the latter case, the approximation largely overstates the potential shortfall risk. Therefore, one has to rely on Monte Carlo simulations.
12.3.4. Improving CPPI Strategies Thus far, we have only used the naïve portfolio as risky asset. We already pointed out that it should be possible to improve the performance of the CPPI strategy in the same manner as the oneperiod portfolio performance can be improved by mean–variance optimization. If we replace the naïve portfolio (characterized by the weekly parameters µX = 0.16%, σX = 2.1%) by a mean–variance efficient portfolio with the same level of risk (volatility) and higher average return (µY = 0.19%, σY = 2.1%), we expect an improved average performance of the CPPI strategy and along with this a decreased shortfall risk. This choice of the mean–variance efficient portfolio is illustrated in Fig. 12.3. As we see in Fig. 12.4, the average return of the CPPI strategy is indeed significantly increasing, while the risk stays nearly at the same level. Therefore, we can
0.11
Excess return (p.a.)
0.105
0.1
0.095
0.09 Efficient frontier Min. variance portfolio Naive portfolio Max. sharpe ratio portfolio Naive portfolio efficient
0.085
0.08 0.13
0.14
0.15
0.16
0.17 0.18 Volatility (p.a.)
0.19
0.2
0.21
0.22
Figure 12.3 Markowitz efficient frontier including mean–variance efficient portfolio.
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
311
Robust Asset Allocation for CPPI Strategies 0.35
0.3
0.25
Exp. return naive Exp. return efficient Vol naive Vol efficient Shortfall prob. naive Shortfall prob. efficient
0.2
0.15
0.1
0.05
0
5
10
15 Multiplier (m)
20
25
30
Figure 12.4 Improvement of the standard CPPI (f = 0.95) by replacing the naïve portfolio with an efficient one.
conclude that mean–variance optimization can actually improve the performance of CPPI strategies. Since mean–variance optimization offers a complete efficient frontier, it is not clear which portfolio thereon should be chosen as risky asset. It is further not obvious how the performance of CPPI strategies on these different risky assets should be compared. We have therefore picked a rather practical approach toward this question. For each CPPI strategy, we choose the largest multiplier — cf. Table 12.7 — such that the shortfall risk is still smaller than 1% and compare their average performance and standard deviation, see Tables 12.8 and 12.9. As risky assets, we haven chosen the Table 12.7 Multipliers for Selected CPPI Strategies for Different Floors f , Such that the Shortfall Probability Is Below 1%.
CPPI naïve CPPI naïve efficient CPPI minvar CPPI maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
14.50 14.50 16.75 16.00
14.00 14.00 16.00 15.50
13.75 14.00 16.00 15.25
13.75 13.75 15.75 15.25
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
312
FA
Schöttle and Werner Table 12.8 Expected Return of Selected CPPI Strategies for Different Floors f and multiplier m, Such that the Shortfall Probability is Below 1%.
CPPI naïve CPPI naïve efficient CPPI minvar CPPI maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.0968 0.1092 0.1016 0.1085
0.0850 0.0958 0.0900 0.0960
0.0754 0.0848 0.0805 0.0854
0.0676 0.0754 0.0724 0.0768
Table 12.9 Standard Deviation of Selected CPPI Strategies for Different Floors f and Multiplier m, Such that the Shortfall Probability is Below 1%.
CPPI naïve CPPI naïve efficient CPPI minvar CPPI maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.1607 0.1661 0.1473 0.1555
0.1448 0.1511 0.1359 0.1437
0.1285 0.1360 0.1234 0.1301
0.1141 0.1209 0.1107 0.1172
already introduced canonical candidates from the previous section — the minimum variance portfolio, the maximum Sharpe ratio portfolio, the naïve portfolio, and its mean–variance efficient replacement. It can be observed that the three mean–variance optimal portfolios clearly dominate the inefficient naïve portfolio, both in terms of expected return and Sharpe ratio. Comparing the efficient portfolios with each other is not as easy. Taking the point of view of the asset manager, the risk of the asset manager is clearly the cost of shortfall. These costs are roughly given by the product of shortfall probability and expected shortfall. With shortfall probability below 1% and expected shortfall roughly equal to 30 basis points, these costs are below one basis point. On the other hand, as the asset manager wants to offer a strategy with a high expected return, the maximum Sharpe ratio portfolio is clearly the best choice. This portfolio offers a similar return as the efficient version of the naïve portfolio, but possesses smaller standard deviation, and should hence be preferred. Remark 12.24. Although the above analysis looks very promising, a few details should not be overlooked. • The CPPI strategy needs to be implemented in the simulation framework as close as possible to the realworld setup, as distinct features of such strategies may influence the results of simulation studies. Among these are for example trading filters (mainly directly linked to the traded lots ϕt ), fees (usually depending on Vt ), slippage effects (depending on the traded volume vt ), ratchets, allowance for short positions, additional risks from overlay management, planned cash injections, etc.
May 12, 2010
17:47
WSPC/SPIB913
Robust Asset Allocation for CPPI Strategies
b913ch12
FA
313
• The joint distribution of the zero bond and the risky asset has to be modeled rather exactly. Especially the tail behavior plays a major role in the quantification of the shortfall risk, whereas the estimation of the expected return is important for the communication of the investment target return. • The choice of the composition of the risky asset and thus the multiplier m are subject to the same estimation risk. If the parameters of the return distribution are uncertain, it is well known that the mean–variance efficient portfolios have a rather poor performance (see next section). In this case, the originally valid idea of using optimized portfolios may revert into the contrary. Unfortunately, while the first point can be covered with sufficient accuracy, the estimation of both the distribution and the distribution parameters are always subject to uncertainty. As shown in [4], [7] or [16] mean–variance efficient portfolios strongly depend on the input data. Due to the inherent uncertainty in the estimates, the resulting portfolios as well as their estimated returns and volatilities are thus not reliable, see for example [23]. Thus, a natural question in this context is: how much can the results differ in the presence of estimation risk? As already mentioned in the introduction, we focus on the estimation risk for the parameters and assume for simplicity that the distribution family itself is known. Remark 12.25. Bertrand and Prigent have considered a possible estimation of the multiplier by extreme value theory in [3]. Although this research is probably the closest one to this study, its line of thinking is solely concerned with shortfall risk. They concentrate on the choice of the multiplier by looking at the corresponding percentiles of the oneperiod return distributions by extreme value theory. Our study goes into a different direction, as we simultaneously cover shortfall risk and expected return, but do so via parameter uncertainty.
12.3.5. CPPI Strategies Under Estimation Risk For the illustration of estimation risk, we consider a very simple example. We still assume that the asset returns follow the originally specified distribution, however, the asset manager has to estimate these parameters, for example from an historical ˆ instead sample. Therefore, the asset manager has to work with estimates µ ˆ and of the original parameters µ and . We have chosen a typical example of µ ˆ and ˆ which represents all stylized facts known about estimation risk. In Fig. 12.5, the , solid line represents the efficient frontier based on the original parameters µ and , which are unknown to the investor. Results (see Table 12.10) corresponding to portfolios based on this efficient frontier will be tagged optimal or original. The dashed line in Fig. 12.5 represents the expectation of the asset manager, i.e., the efficient
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
314
FA
Schöttle and Werner
0.18
Excess return (p.a.)
0.16
efficient frontier based on original parameters efficient frontier based on random parameters random parameters only for optimization minimum variance portfolio maximum Sharpe ratio portfolio
0.14
0.12
0.1
0.08
0.06 0.1
0.12
0.14
0.16 Volatility (p.a.)
0.18
0.2
0.22
Figure 12.5 Original and estimated efficient frontiers.
ˆ This situation will lead to the frontier based on the random parameters µ ˆ and . numbers tagged perceived or random. Finally, the dotted line is the efficient frontier the investor will actually obtain, i.e., optimizing the portfolios according to the ˆ but realizing the risk–return profile with the true market expected parameters µ ˆ and , parameters µ and . Table 12.10 summarizes the return and volatility characteristics of the minimum variance and the maximum Sharpe ratio portfolios on the three frontiers. Therefore, in Fig. 12.5, two kinds of estimation risk can be noticed. First, the ˆ and thus return µ asset allocation is calculated based on µ ˆ and ˆ X and volatility σˆ X are obtained based on the optimized allocation. As shown in [17], these are Table 12.10 Illustration of the Estimation Risk of the Minimum Variance Portfolio and the Maximum Sharpe Ratio Portfolio.
Return minimum variance Return maximum sharpe ratio Volatility minimum variance Volatility sharpe ratio
Actual
Perceived
Optimal
0.0800 0.0733 0.1355 0.1539
0.0823 0.1147 0.1175 0.1323
0.0836 0.0924 0.1336 0.1393
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
315
Robust Asset Allocation for CPPI Strategies
biased estimators compared to the optimal values µX and σX and on average it holds that µ ˆ X > µX and σˆ X < σX . This also gets obvious from Fig. 12.5. Further, as the wrong asset allocation is implemented, an inferior portfolio performance can be achieved in the market, see also [17] for more details. This can again be observed in Fig. 12.5, where the true original frontier dominates the portfolios calculated by the asset manager. Although we have picked only one representative toy example, in [17], it is shown that this represents the average situation. Taking the estimates µ ˆ ˆ together with the corresponding risky asset compositions, the simulation of and the standard CPPI can be performed as usual Table 12.11. From this, optimal multipliers are obtained for both portfolios. The multipliers have been chosen in such a way that the corresponding shortfall probabilities are roughly 1%. This choice of multipliers yields the following perceived performance and standard deviation, see Tables 12.12 and 12.13. Based on these multipliers and the wrong allocation, we run the simulation again, this time with the true market parameters. Then, instead of the above figures, we obtain the corresponding actual figures, see Tables 12.14–12.16. Table 12.11
CPPI minvar CPPI maxsharpe
Table 12.12
f = 0.95
f = 0.975
f = 0.99
f = 1.00
19.50 17.50
18.75 16.50
18.25 16.25
18.00 16.00
Perceived Expected Return of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe
Table 12.13
Multipliers m of Selected CPPI Strategies.
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.1048 0.1358
0.0941 0.1205
0.0847 0.1075
0.0766 0.0960
Perceived Standard Deviation of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.1331 0.1574
0.1265 0.1496
0.1163 0.1384
0.1057 0.1264
Table 12.14 Actual Shortfall Probabilities of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.0409 0.0631
0.0446 0.0584
0.0447 0.0595
0.0449 0.0568
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
316
FA
Schöttle and Werner Table 12.15 Actual Expected Return of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.0978 0.0880
0.0868 0.0774
0.0778 0.0691
0.0704 0.0625
Table 12.16 Actual Standard Deviation of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.1493 0.1608
0.1391 0.1456
0.1267 0.1304
0.1145 0.1163
The most noteworthy differences are both in the expected performance and in the shortfall probability. Although the shortfall probability was believed to be below 1%, it significantly increased to about 5%. This is not only worrying from the point of view of the asset manager, who actually faces higher costs to cover shortfall, but as well from the perspective of the client, who has a much higher probability of ending up in the absorbing state than intended. Even more troublesome is the expected return of the formerly superior maximum Sharpe ratio portfolio. Due to the error maximization, the expected return of this portfolio turned out to be unreliable and the gap to the perceived performance is significant. In contrast to this, the minimum variance portfolio shows a rather stable behavior in terms of expected return. Based on this toy example, it is already obvious that a more prudent choice of multiplier and a consideration of the estimation risk in the composition of the risky asset is absolutely necessary. Therefore, in the next section, we will introduce a framework which not only provides more robust asset allocations, but at the same time gives prudent estimates of risk and return which can be used in the composition of the CPPI strategy.
12.4. ROBUST MEAN–VARIANCE OPTIMIZATION AND IMPROVED CPPI STRATEGIES In the last years, several attempts have been made to improve the estimation of mean–variance efficient portfolios. A meanwhile pretty wellknown way to robustify portfolios is the resampling technique introduced and made popular by Michaud, see [23], based on concepts as, e.g., given in [16]. Another rather popular approach is the usage of robust estimators, as, e.g., investigated in [18], [19] or [26]. The third line of thinking is based on the robust counterpart idea introduced by BenTal and Nemirovski,
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
317
Robust Asset Allocation for CPPI Strategies
see [2]. Several authors have considered instances varying from theoretical to practical settings, see e.g. [6], [13], [20], [31] and [33].
12.4.1. Robust Mean–Variance Analysis Depending on the availability of historical data (length of sample period, frequency of sampling, etc.) and the utilized estimation routine for µ and , the resulting estimates, ˆ can vary, i.e., they are uncertain. Now, instead of only considering denoted by µ ˆ and , these uncertain point estimates, an entire range of possible parameter realizations — a (convex and compact) uncertainty set U — is taken into account and the portfolio is optimized with respect to the worst outcome within U. Definition 12.26. The robust counterparts to problems (Pλ ) are given by √ min max (1 − λ) w Cw − λw r w∈W (r,C)∈U
(RPλ )
with U being the (joint) uncertainty set for the unknown parameters (µ, ). In analogy to the classical setting, the optimal solution of (RPλ ) will be denoted by w∗rob (λ). Remark 12.27. It can be shown that the resulting robust portfolios are unique for 0 ≤ λ < 1, see [29], Proposition 5.2, for general convex and compact uncertainty sets U. Most applications of the robust counterpart approach are based on statistical confidence sets around the corresponding point estimates, see [2], [6], [13], [20], [22], or [32]. In general, the most intuitive confidence sets coming from statistics are confidence ellipsoids, as for example nicely motivated in [22], Sec. 2.4.3. Typically, the center of the ellipsoid is determined as the estimated value of the point estimate and its shape is described using an according estimate for the estimator’s covariance matrix. In addition, the size of the uncertainty set is calculated from an appropriate percentile to achieve the desired level of confidence. Alternatively, several different expert opinions (or likewise different estimators) can be used to form an uncertainty set, see [20]. In most practical instances, the robustification is focused on the uncertainty in the mean, while the covariance matrix is assumed to be known with certainty, which is well supported by a number of theoretical results on the impact of both uncertainties.
12.4.2. Uncertainty Sets Via Expert Opinions or Related Estimators In the following, we assume that we have several different estimators available for the uncertain parameter, for example from experts’ forecasts. Alternatively, different estimators may be derived by different estimation routines. The standard maximum
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
318
FA
Schöttle and Werner
likelihood estimator, although very popular, is not the only unbiased consistent estimator, which may be used for elliptical distributions. In fact, there exist several other common estimators for the mean of an elliptical distribution used by practitioners. For example one could consider the following four consistent estimators for the return besides the maximum likelihood estimator: • The median, i.e., the 50% quantile of the data sample • The average of the 25% and the 75% quantile as well as two robust estimators, • The Huber estimator, see [14] • The trimmed mean, which is defined as the maximum likelihood estimator of the sample reduced by the αpercent smallest and largest values (i.e., the outliers) As shown in [30], ellipsoidal uncertainty sets have nicer theoretical properties — in terms of smoothing and robustness properties — than polyhedral ones. Thus, instead of using the (polyhedral) convex hull ˆ U = conv(r1 , . . . , r5 ) × {} 5 5 ˆ = r r = αi ri , with αi ≥ 0, αi = 1 × {} i=1
i=1
of these five estimators, denoted by r1 , . . . , r5 , to define the uncertainty set U, we prefer to use the ellipsoid ˆ U = {r  (r − r¯ ) C¯ −1 (r − r¯ ) ≤ δ2 } × {}, where r¯ is the average of the five different estimates and C¯ = D2 with D being the diagonal matrix with the standard deviations of the individual estimates from their common average. In this case, the size δ of the ellipsoid is determined such that all estimates are contained within. In other words, the uncertainty set for the return is given by the smallest ellipsoid centered at the common average of the various estimators and shaped by their standard deviations. Combining everything and solving the inner maximization problem in (RPλ ) analytically, the robust portfolios w∗rob (λ) can be obtained by solving ˆ + λ[δ w Cw ¯ − w r¯ ]. min (1 − λ) w w w∈W
This means that on top of the investment risk, i.e., risk from the stochastic asset returns, an additional penalty for estimation risk is considered in the optimization. This penalty grows linearly with the uncertainty in the estimators, i.e., with δ, as well as linearly in the importance λ of the return term. Although this is a very appealing approach
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
319
Robust Asset Allocation for CPPI Strategies
in practical settings, for the aim of the study, it is easier to work with an alternative specification of uncertainty sets.
12.4.3. Uncertainty Sets Via Confidence Sets As already mentioned, the most prominent choice for the uncertainty set lies in the traditional confidence ellipsoid around the maximum likelihood estimator µ, ˆ i.e., ˆ ˆ −1 (r − µ) ˆ ˆ ≤ δ2 } × {}, U1 := {r  (r − µ)
(12.6)
where δ is chosen appropriately to the target confidence level of the uncertainty set. Again, no uncertainty of the covariance matrix is considered. A more advanced version of the confidence set also considers the uncertainty of the covariance matrix. In the latter case, applying similar ideas as in [22], it is possible to show (see [32]) that the natural confidence ellipsoid around the joint maximum ˆ approximately has the following form likelihood estimator for both (µ, ˆ ) −1 ˆ 2 (C − ) ˆ ˆ − 12 2 ≤ δ2 . ˆ −1 (r − µ) ˆ ˆ + 2 · U2 = (r, C)  (r − µ) (12.7) tr Using any one of the above uncertainty sets a surprising result can be proved by a close investigation of the necessary and sufficient firstorder optimality conditions (i.e., the KKT conditions), see [28] or [32]: the robust efficient frontier is exactly the classical efficient frontier up to a certain risk (i.e., volatility) level depending on the choice of the parameter δ. In other words, the original classical efficient frontier is robust in itself, as long as we do not move too far to the right. This can be interpreted as if robustification is able to identify the unreliable upper part of the efficient frontier which is cut off. The following theorems, taken from [32] formally state the described result. Let δ H(θ) = 1 − θ + θ √ , S K(θ) = max (1 − θ) 1 + δ
κ∈[0,1]
2 (1 − κ) + θδ S−1
κ , S
where in this context S denotes the number of observations used for the point estimation. Although K is given in terms of an optimization problem, it is computed rather easily by any onedimensional optimization routine. Theorem 12.28. Let U = U1 as in (12.6) and let θ ∈ [0, 1]. Then the optimal solution w∗rob (θ) of (RPθ ) equals the optimal solution w∗cl (λ) of the original problem (Pλ ) for λ = H (θ) = θ/(θ + H(θ)).
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
320
FA
Schöttle and Werner
Theorem 12.29. Let U = U2 as in (12.7) and let θ ∈ [0, 1]. Then the optimal solution w∗rob (θ) of (RPθ ) equals the optimal solution w∗cl (λ) of the original problem (Pλ ) for λ = K (θ) = θ/(θ + K(θ)). Remark 12.30. The most important consequences of these theorems are • The cut off risk level for the robust frontier can be determined by solving the classical formulation with 1 √ < 1. λmax = H (1) = K (1) = 1 + δ/ S The optimal portfolio w∗cl (λmax ) then determines the cut off risk level as ˆ ∗cl (λmax ). σmax = w∗cl (λmax )w • Considering the special point θ = 0, i.e., the (robust) minimum variance portfolio, the corresponding λ is also zero, i.e., the minimum variance portfolio is already robust in itself. For U1 the robust counterpart reduces to ˆ min ww, w∈W
which is obvious as the minimum variance portfolio does not depend on the uncertain parameter µ. If uncertainty of the covariance matrix is considered, then the robust counterpart reduces to 2 ˆ ˆ w w. min K(0) w w = min 1 + δ w∈W w∈W S−1 for the uncertainty set U2 . This means that the consideration of the uncertainty of the covariance matrix is reflected in the factor K(0). Since this factor is larger than 1, a higher volatility (i.e., a larger objective value) is expected for the minimum variance portfolio in the robust setting. This increase in the risk can be interpreted as a measure for estimation risk and can be incorporated to obtain an improved estimation of the multiplier m. • The results of both theorems hold true due to the fact that the matrices for measuring the portfolio’s risk and for describing the shape of the uncertainty set U have the same structure and hence the expressions for the risk and the penalty for estimation risk can be combined. Otherwise we would end up as in the previous subsection, where different penalty terms for estimation risk appear in the objective function. The effect of such a robustification for an investor is that her position on the efficient frontier moves to the left, toward the minimum variance portfolio. Hence, robustification leads to more conservative portfolio choices that are nevertheless efficient — simply with respect to a different tradeoff parameter.
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
321
Robust Asset Allocation for CPPI Strategies
12.4.4. Usage and Implications for CPPI Strategies The above robust optimization framework can now be used to improve the performance of CPPI strategies under estimation risk. For this purpose, we suggest the following methodology: Algorithm 12.31. For a given robustification parameter δ and estimated parameters ˆ based on S historical observations, we proceed in the following steps to µ ˆ and determine an appropriate allocation of the risky asset w together with a corresponding multiplier m: (1) Calculate the robust efficient frontier according to the previous section. Choose an arbitrary portfolio w from the robust frontier, preferably the robust minimum variance or the robust maximum Sharpe ratio portfolio. (2) Determine the robustified parameters 1 ˆ w, µ ˜ =µ ˆ − δ√ ˆ S ww
ˆ ˜ = K(0)2 · .
(3) Run the CPPI strategy with the robustified parameters and fix the multiplier m as usual, i.e., at an acceptable level of shortfall risk. The choice K(0) is deduced from the fact that this factor appears in the robustified version of the minimum variance problem and represents the penalty for the uncertainty in . For all other portfolios, K(λ) is smaller and could be calculated, but for simplicity, we use K(0) instead. The correction of the expected returns is motivated by the fact that the worstcase return in the confidence set is given by µ, ˜ although this is a rather pessimistic approach which may need some refinement. If we apply Algorithm 12.31 to the previous setting for realistic choices of robustification parameters, δ = 1 and S = 52, we obtain µ ˜ and K(0) as in Table 12.17. It can be observed that the robustified return µ ˜ is overly pessimistic and cannot be used to estimate the expected performance of the portfolios or the CPPI strategy. Nevertheless, it is still well suited for the determination of the multiplier m, as for this purpose, an adverse market setting leads to a conservative choice. Further, the volatility plays the major role in the setting of the multiplier anyway. Table 12.17 µ µ ˆ µ ˜ K(0)
0.15% 0.16% −0.13% 1.0945
Period Returns for the Tradeable Assets.
0.26% 0.17% −0.17%
0.18% 0.33% 0.05%
0.26% 0.16% −0.13%
0.24% 0.24% −0.09%
0.22% 0.21% −0.07%
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
322
FA
Schöttle and Werner Table 12.18 Multipliers m of Robust CPPI Strategies Based on Robustified Parameters.
CPPI robust minvar CPPI robust maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
16.25 15.75
15.75 15.25
15.75 15.25
15.50 15.00
Using the robustified allocations for the risky assets and running the CPPI simulation with the robustified parameters, we get the following multipliers, illustrated in Table 12.18. The multiplier was again chosen in such a way that the corresponding shortfall probability is below 1%. Comparing these multipliers to those in Table 12.11, we see that the above multipliers are significantly smaller than those derived from the estimated, but nonrobustified, parameters. Remark 12.32. From the illustration of the robust frontier, see Fig. 12.6, it can be noticed that the robust frontier is usually rather short. This means that large deviations from the minimum variance portfolios are usually not possible within the robust framework. Therefore, we do not expect large differences between the (robust) minimum variance and the robust maximum Sharpe ratio portfolio.
0.11
Excess return (p.a.)
0.105
0.1
0.095
0.09
0.085 Efficient frontier Robust efficient frontier 0.08 0.13
0.14
0.15
0.16
0.17 0.18 Volatility (p.a.)
0.19
0.2
0.21
Figure 12.6 Illustration of the classical and the robust efficient frontier.
0.22
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
FA
323
Robust Asset Allocation for CPPI Strategies
12.4.5. CPPIs with Robust Asset Allocations After we have calculated both the (robustified) risky allocations and the corresponding multipliers, the CPPI strategy can be simulated as usual with the true market parameters. From this simulation, we obtain the actual shortfall probabilities and the actual average performance of the CPPI strategy. According to our expectation, the robustification, which resulted in a decrease of the corresponding multiplier as well as a more prudent asset allocation, yields very promising results on shortfall probabilities. In contrast to the significantly increased probabilities for the nonrobustified allocations we now have comparable shortfall probabilities as intended, see Table 12.19. Before comparing actual expected returns of selected (robust) CPPI strategies, we recall in the following Tables 12.20 and 12.21 the optimal and perceived expected returns of nonrobustified CPPI strategies as already given in Tables 12.8 and 12.12. Now finally, the actual expected returns of the considered nonrobust and robust CPPI strategies are summarized in Table 12.22. It can be observed that the two CPPI strategies based on minimum variance portfolios lead to almost identical expected returns. Eventually, the almost identical returns are not surprising since the allocations for the risky asset used in the respective CPPI strategies are identical, only the multipliers differ. Nevertheless, this difference in the multiplier is responsible for a rather improved Table 12.19 Actual Shortfall Probabilities of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe CPPI robust minvar CPPI robust maxsharpe
Table 12.20
f = 0.975
f = 0.99
f = 1.00
0.0409 0.0631 0.0077 0.0093
0.0446 0.0584 0.0081 0.0099
0.0447 0.0595 0.0098 0.0110
0.0449 0.0568 0.0094 0.0096
Optimal Expected Return of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe
Table 12.21
f = 0.95
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.1016 0.1085
0.0900 0.0960
0.0805 0.0854
0.0724 0.0768
Perceived Expected Return of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.1048 0.1358
0.0941 0.1205
0.0847 0.1075
0.0766 0.0960
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
324
FA
Schöttle and Werner Table 12.22 Actual Expected Return of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe CPPI robust minvar CPPI robust maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.0978 0.0880 0.0973 0.0945
0.0868 0.0774 0.0861 0.0833
0.0778 0.0691 0.0771 0.0744
0.0704 0.0625 0.0695 0.0670
Table 12.23 Actual Standard Deviation of Selected CPPI Strategies.
CPPI minvar CPPI maxsharpe CPPI robust minvar CPPI robust maxsharpe
f = 0.95
f = 0.975
f = 0.99
f = 1.00
0.1493 0.1608 0.1473 0.1503
0.1391 0.1456 0.1351 0.1368
0.1267 0.1304 0.1222 0.1229
0.1145 0.1163 0.1091 0.1093
shortfall probability for the robust minimum variance portfolio compared to the nonrobust one. In case of the CPPI strategies based on maximum Sharpe ratio portfolios, the expected returns of the robust versions are significantly higher compared to the nonrobustified risky asset allocations and much closer to the optimal ones. However, even with robustification, CPPI on maximum Sharpe ratio portfolios actually cannot beat those on the minimum variance portfolio, which in turn still beats the naïve strategy. This means that portfolio optimization still adds value to the investment process, but that estimation risks have to be taken very seriously. For completeness, Table 12.23 contains the actual standard deviations of selected nonrobust and robust CPPI strategies. As expected, all robust versions have a lower standard deviation than their corresponding nonrobust strategy.
12.5. CONCLUSION In this study, we have shown that under full information CPPI strategies should rely on mean–variance efficient portfolios as risky assets. Based on Monte Carlo simulations, suitable multipliers can be found. We have demonstrated that estimation risk may have a significant impact on the performance of the CPPI strategy. A potential remedy by robust mean–variance optimal portfolios seems to be beneficial for the overall performance of the strategy, especially for the shortfall risk of the asset manager and the average return. Nevertheless, our results also show that optimization beyond the minimum variance portfolio is heavily influenced by estimation risk. Based on these
May 12, 2010
17:47
WSPC/SPIB913
b913ch12
Robust Asset Allocation for CPPI Strategies
FA
325
preliminary results, it can be concluded that a more detailed and thorough analysis needs to be carried out to investigate estimation risk and its cures in a broader context.
Acknowledgments The authors want to express their gratitude to the editors for the opportunity to make this contribution. Further, the second author appreciates the hospitality at the HVB Institute for Mathematical Finance, TU München, especially the inspiring annual research seminars which fostered part of the research on robust allocations.
Disclaimer The views expressed in this article are the authors’ personal opinions and should not be construed as being endorsed by MEAG MUNICH ERGO AssetManagement GmbH or Hypo Real Estate Holding AG.
References [1] Allianz Global Investors (2005). Privatizing pensions. [2] BenTal, A and A Nemirovski (1998). Robust convex optimization. Mathematics of Operations Research, 23(4), 769–805. [3] Bertrand, P and JI Prigent (2002). Portfolio insurance: The extreme value theory of the CPPI method. Finance, 23(2), 69–86. [4] Best, MJ and RR Grauer (1991). On the sensitivity of mean–varianceefficient portfolios to changes in asset means: Some analytical and computational results. The Review of Financial Studies, 4(2), 315–342. [5] Black, F and R Jones (1987). Simplifying portfolio insurance. Journal of Portfolio Management, 14(1), 48–51. [6] Ceria, S and RA Stubbs (2006). Incorporating estimation errors into portfolio selection: Robust portfolio construction. Journal of Asset Management, 7(2), 109–127. [7] Chopra, VK and WT Ziemba (1993). The effect of errors in means, variances and covariances on optimal portfolio choice. Journal of Portfolio Management, 19(2), 6–11. [8] DeMiguel, V and F Nogales (2006). Portfolio selection with robust estimates of risk. SSRN Working Paper. [9] Dichtl, H and C Schlenger (2004). Aktien oder Renten? — Das Langfristpotenzial der Best of TwoStrategie. Die Bank, 12, 809–813. [10] Fang, KT, S Kotz and KW Ng (1990). Symmetric Multivariate and Related Distributions. London: Chapman and Hall. [11] Fang, KT and YT Zhang (1990). Generalized Multivariate Analysis. Beijing: Science Press. [12] Föllmer, H and A Schied (2004). Stochastic Finance — An Introduction in Discrete Time. de Gruyter Studies in Mathematics.
May 12, 2010
326
17:47
WSPC/SPIB913
b913ch12
FA
Schöttle and Werner
[13] Goldfarb, D and G Iyengar (2003). Robust portfolio selection problems. Mathematics of Operations Research, 28(1), 1–38. [14] Huber, PJ (1981). Robust Statistics. New York: John Wiley & Sons. [15] Ingersoll, JE (1987). Theory of Financial Decision Making. Maryland: Rowman & Littlefield. [16] Jorion, P (1992). Portfolio optimization in practice. Financial Analysts Journal, 48(1), 68–74. [17] Kan, R and DR Smith (2008). The distribution of the sample minimum–variance frontier. Mangement Science, 54(7), 1364–1380. [18] Kempf, A and C Memmel (2003). On the estimation of the global minimum variance portfolio. SSRN Working Paper. [19] Lauprête, GJ,AM Samarov and RE Welsch (2002). Robust portfolio optimization. Metrika, 55(2), 139–149. [20] Lutgens, F (2004). Robust Portfolio Optimization. PhD Thesis, Maastricht University. [21] Markowitz, H (1952). Portfolio selection. Journal of Finance, 7(1), 77–91. [22] Meucci, A (2005). Risk and Asset Allocation. Berlin: Springer. [23] Michaud, RO (1998). Efficient Asset Management. Boston: Harvard Business School Press. [24] Perold, A (1986). Constant portfolio insurance. Unpublished manuscript. [25] Perold, A and W Sharpe (1988). Dynamic strategies for asset allocation. Financial Analysts Journal, 44, 16–27. [26] PerretGentil, C and MP VictoriaFeser (2004). Robust mean–variance portfolio selection. FAME Research Paper no. 140. [27] Prigent, JL (2007). Portfolio Optimization and Performance Analysis. Chapman & Hall. [28] Scherer, B (2007). Can robust portfolio optimisation help to build better portfolios? Journal of Asset Management, 7(6), 374–387. [29] Schöttle, K (2007). Robust Optimization with Application in Asset Management. Dissertation Technische Universität München. [30] Schöttle, K and R Werner (2005). Costs and benefits of robust optimization. Working Paper. [31] Schöttle, K and R Werner (2006). Towards reliable efficient frontiers. Journal of Asset Management, 7(2), 128–141. [32] Schöttle, K and R Werner (2009). Robustness properties of mean–variance portfolios (to appear in: Optimization). [33] Tütüncü, RH and M Koenig (2004). Robust asset allocation. Annals of Operations Research, 132(1–4), 157–187.
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
13
ROBUST ASSET ALLOCATION UNDER MODEL RISK
PAULINE BARRIEU∗ and SANDRINE TOBELEM† Department of Statistics, London School of Economics, Houghton Street, London WC2A 2AE England, UK ∗ p.m.barrieu@lse.ac.uk † s.e.tobelem@lse.ac.uk
In this chapter, we propose a robust asset allocation methodology, when there is some ambiguity concerning the distribution of asset returns. The investor considers several prior models for the assets distribution and displays an ambiguity aversion against them. We have developed a twostep ambiguity robust methodology that offers the advantage to be more tractable and easier to implement than the various approaches proposed in the literature. This methodology decomposes the ambiguity aversion into a modelspecific ambiguity aversion as well as relative ambiguity aversion for each model across the set of different priors. The optimal solutions inferred by each prior are transformed through a generic absolute ambiguity function ψ. Then, the transformed solutions are mixed together through a measure π that reflects the relative ambiguity aversion of the investor for the different priors considered. This methodology is then illustrated through the study of an empirical example on European data.
327
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
328
FA
Barrieu and Tobelem
13.1. BACKGROUND In this chapter, we aim at characterizing and constructing a methodology for robust portfolio allocation under model risk, i.e., when the investors consider different models for asset returns distribution to take their allocation decision. More precisely, let us consider a financial market with N risky assets and a riskfree asset. An investor wants to allocate her wealth among these assets, by choosing φ ≡ (φ0 , . . . , φN ), the vector of weights for the riskfree asset and the N risky assets. The standard framework for investment decision making has been developed by Markowitz in [1], where there is no model uncertainty. The optimal portfolio allocation is obtained as the solution of the following optimization program: φ∗ ≡ argmax EP [u(Xφ , λ)],
(13.1)
φ
where u is a VonNeumann Morgenstern utility function characterizing the investor’s preferences and parametrized by the risk aversion parameter λ, and Xφ stands for the terminal value of the portfolio at a given time horizon. In this setting, P stands for the only prior (or model for the distribution of the assets returns) the investor has, and is known without ambiguity. Hence, the risk of the investor is perfectly quantifiable through the knowledge of the distribution P. The agent may also consider different models Q in a finite set of possible models Q. In this case, the investment problem is modified according to the subjective view π(Q) of the investor on each model Q. More precisely, π(Q) represents the subjective likelihood of the model Q for the investor. The investor operates a linear blending of the different models, weighted by their subjective probability π(Q) to be the “real” model. Under each model, the investor considers the objectiveexpected utility of her future wealth. Across all priors, the investor considers the subjectiveexpected value of the expected utilities under the different models. Such a framework is referred to as Subjective Expected Utility (SEU) and was first introduced by Savage in [2]. The optimal portfolio allocation is then obtained as EQ [u(Xφ , λ)]π(Q). (13.2) φ∗ ≡ argmax φ
Q∈Q
Note that in this framework, even if the agent has several priors as reference, he/she does not show any aversion toward the coexistence of different models to take her decision. He or she is neutral toward model uncertainty as there exists no ambiguity about the set of models considered and their likelihood to occur. However, as shown by Ellsberg in [3], the decision makers show a more averse behavior when betting on events for which outcomes are ambiguous (i.e., when there is also some uncertainty regarding the underlying model) than when betting on events for which the outcomes are only risky (i.e., the underlying model is well known).
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
Robust Asset Allocation Under Model Risk
329
A financial illustration of the Ellsberg paradox is the risk premium paradox: investors tend to invest more in their local market even though the expected return is lower than for foreign markets, the reason being that the investors add an ambiguity premium to risky assets (investors prefer investing in assets located in their geographical zone, because they can better apprehend their return distribution). For this reason, the SEU framework fails to take into account this additional source of aversion of investors. Various approaches have been developed in the literature to take into account this aversion toward model uncertainty in the investment decision process. Among them, Gilboa and Schmeidler in [4] proposed a min–max approach leading to a very conservative decision rule based upon the worst case model. More recently, Klibanoff et al. in [5] introduced a generalized model. They consider an increasing, concave transformation function characterizing the investor ambiguity aversion through a parameter γ. The optimal weights vector is then determined as {EQ [u(Xφ , λ)], γ}π(Q). (13.3) φ∗ ≡ argmax φ
Q∈Q
The main feature of this model is that it unifies all the previous approaches accounting for model ambiguity. However, this theoretical approach can be very challenging to implement in practice for different reasons, including the calibration of the various parameters. Indeed, no distinction is made between specific ambiguity aversion for a given model (“How good is this specific model to represent the reality?”) and general ambiguity aversion for the whole class of models (“How much can all the models explain the reality?”). Moreover, solving explicitly Program (13.3) can be extremely difficult, even numerically, especially in the multidimensional case or when adding some constraints on the portfolio allocation. Note that Klibanoff et al. only give a simple numerical example for a portfolio with three assets in their paper [5], whereas practitioners often consider portfolios with hundreds of assets. We have compared their example to our methodology in [6] and we also provide a more complex theoretical example that can be solved in close form if using our methodology. Finally, such an approach lacks some flexibility in the sense that if the investor considers a new model, he or she has to recompute the program entirely. To overcome those limitations, we propose a robust, general framework for decision making under uncertainty. We are not aiming at finding the optimal solution for a given criterion but more at finding a robust solution.
13.2. A ROBUST APPROACH TO MODEL RISK We propose a new approach to model ambiguity that is altogether more flexible, easier to compute, more robust and tractable than the methods proposed in the literature. This
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
330
FA
Barrieu and Tobelem
Ambiguity Robust Adjustment (ARA) approach is independent of the set of models considered, as well as of the choice criterion (i.e., the value function they consider to determine their preferred asset allocation) to define the optimal portfolio under each model Q ∈ Q, for Q a finite set of models. More precisely, we proceed in two steps and introduce a distinction between two types of ambiguity: • Absolute ambiguity: this refers to the ambiguity the investor has for a given model. We first solve the optimization problem assuming that the model considered is the true model. We thus compute a distorted expected value of the deduced optimal weights, transformed by an Absolute Ambiguity Robust Adjustment (AARA) function denoted ψ. Note that the absolute adjustment is made on the solution and not on the choice criterion. This allows us for some additional flexibility in the use of each model. • Relative ambiguity: this expresses the relative ambiguity the investor has among his or her different models: in a second step, we aggregate the adjusted optimal weights computed for each model through a RelativeAmbiguity RobustAdjustment (RARA) function, denoted π. Let us denote φQ ≡ argmaxφ EQ [u(Xφ , λ)]. Then the ARA portfolio allocation φARA ≡ (φiARA )i∈{1,...,N} is obtained as φiARA ≡ ψ{φQi , γ}π(Q), i ∈ {1, . . . , N} (13.4) N
Q∈Q
and φ0ARA ≡ 1 − i=1 φiARA . Due to its specific nature, the riskfree asset has no model risk associated with it (its future value is known with certainty). Therefore, it plays a specific role in the ambiguityadjusted optimal asset allocation. It can be assimilated to a refuge value in the following sense: the more the investor is averse to ambiguity, the more he or she will invest in the riskfree asset. In this sense, as the “disinvested” part of the wealth from the risky assets is transferred to the riskfree asset, the adjusted weight of the riskfree asset corresponds to the amount of money the investor is reluctant to invest in risky assets because of her aversion toward model risk. We describe below the characteristics of the functions ψ and π.
13.2.1. The Absolute Ambiguity Robust Adjustment The idea behind the AARA adjustment is that the investor, because she has doubts about the optimal weights generated by a given model, wishes to scale down those weights and especially the biggest absolute weights that could entail the biggest risks in her portfolio.
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
Robust Asset Allocation Under Model Risk
331
The function ψ must satisfy some key properties to be consistent with the rationality of the investor: • Definition of the function ψ: The investor treats the absolute ambiguity aversion with the same type of transformation across all the different models (ψ is the same for all the models). What distinguishes the absolute ambiguity aversion transformation across the models is the specific ambiguity aversion parameter γ Q the investor attributes to each model. As the optimal weights obtained for each model φQ are bounded by 1, ψ(1, γ Q ) represents the maximum weight the investor will assign to any asset after the AARA transformation. We note aQ ≡ ψ(1, γ Q ) and a ≡ maxQ∈Q aQ . Therefore, ψ is defined on the set of optimum model dependent weights [−1; 1] × Q onto a set [−a; a] of transformed weights. Note that the investor can set the value of aQ and deduce the value of γ Q depending on the explicit form he or she chooses for ψ. ψ : [−1; 1] × Q → [−a; a] (13.5) (φ, Q): → ψ(φQ , γ Q ) The following properties apply for the risky assets only, i.e., for i ∈ {1, . . . , N}. • Monotonicity: One of the key characteristics of ψ is its monotonicity property. ψ preserves the relative order of the optimal weights (φiQ )1≤i≤N deduced by a given model Q, so that the relative preference of the investor toward the different risky assets given a model Q is preserved through the transformation ψ. • Convexity: The function ψ is concave on [0; 1] and convex on [−1; 0], so that the function ψ reduces more the absolute largest weights given by the optimized portfolios under each model considered. The convexity scale is parametrized through the aversion coefficient γ Q : the bigger the aversion coefficient γ Q the more averse the investor is to large weights inferred by Q. Also the convexity of ψ will depend upon the investor ambiguity aversion toward the model Q ∈ Q considered. The function ψ has an Sshape, as it penalizes more the largest positive and negative weights (this is a particular case of the convexity property of ψ). For all assets i and all models Q, we have ψ(φiQ ) ≤ φiQ . The absolute ambiguity adjusted weights are smaller than the optimal weights computed under a given model Q in absolute terms. Some additional properties can be considered depending on the assumption made on the investor preferences and trading constraints. • Symmetry: In a context where short selling is possible, there is no reason to differentiate the long or short weights of the same magnitude in terms of ambiguity aversion.
May 12, 2010
332
17:47
WSPC/SPIB913
b913ch13
FA
Barrieu and Tobelem
The function ψ is then an odd function symmetric around zero. We can assume that a longshort investor has the same aversion to positive or negative weights of the same absolute value: ∀φi ∈ [−1; 1], ψ(−φi ) = −ψ(φi ). The AARA function penalizes the scale of the optimal weights of a given model without discriminating between negative and positive weights. • Invariant point: There is no ambiguity aversion for a zero weight: ψ(0) = 0. If the model Q assigns no weight on a given asset, the transformation ψ should not modify the “neutrality” of the model Q toward this asset. • Limit behavior: When the investor is infinitely averse to ambiguity, it will prevent her from trading as he or she trusts none of his or her models, and therefore all the portfolio weights should be defaulted to zero. On the contrary, if the investor is neutral to ambiguity, the function ψ should leave the modeldependent weights invariant. We used a very similar function than the one applied by Klibanoff et al. to account for ambiguity. The function ψ can be any classical SShape function, which has the nice property of being concave (convex for negative values), symmetric, and monotonic (similar attributes as for classical utility functions). What really characterizes the function ψ is its ambiguity parameter γ that accounts for the concavity of the function ψ and therefore the ambiguity aversion of the investor (Fig. 13.1).
Figure 13.1 ψ for different values of the ambiguity aversion parameter γ.
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
333
Robust Asset Allocation Under Model Risk
An example for the function ψ is 1 − exp−γx , γ ψ(x, γ) ≡ expγx −1 , γ
0 ≤ x ≤ 1, (13.6) −1 ≤ x ≤ 0.
13.2.2. Relative Ambiguity Robust Adjustment Once the optimal solutions have been computed for each prior Q and have been independently adjusted for ambiguity aversion through the AARA function ψ, we need to aggregate them across all priors in the set Q. The RARA function takes into account the ambiguity aversion of each prior relative to the whole class of priors Q. Such an adjustment is made through a mixture measure π. The RARA function π(Q) represents the likelihood or degree of confidence the decision maker has for the adjusted result given under Q when knowing all the adjusted results for all the other priors. Therefore, π(Q) can be seen as a subjective weight given by the decision maker to the adjusted solution for the model Q. π(Q) will therefore always be nonnegative. If the decision maker does not trust at all the prior Q relatively to the other priors, he or she will simply set the weight to zero. If on the contrary he or she fully trusts the prior Q, relatively to the other priors, then the weight for all the other priors will be zero. The weight π(Q) is not necessarily one, since the agent may still believe that he or she does not have the full understanding of the situation. More formally we have Definition 13.1 (RARA). The measure π : Q → [0; 1] is a Relative Ambiguity Robust Adjustment (RARA) measure if: π(Q) ≤ 1. ∀ Q ∈ Q, 0 ≤ π(Q) ≤ 1 and Q∈Q
After the transformation through the AARA function ψ and the RARA function π, the ARA weight of any risky assets is defined as ψ(φiQ , γ Q )π(Q) ∀i ∈ [1; N], φiARA = Q∈Q
and for the riskfree asset, the optimal weight is φ0ARA = 1 −
N i=1
φiARA .
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
334
FA
Barrieu and Tobelem
13.2.3. ARA Parametrization The investor’s aversion to ambiguity is dynamic in the sense that, depending on the period considered, he or she will be more or less confident about his or her models and the overall set of models he or she considers. Therefore, we allow the function π and the ambiguity aversion parameter γ to dynamically adapt and expand or contract the total investment size whether the total ambiguity aversion decreases or increases over time (the ambiguity parameter γ and the measure π can be reparametrized at every decision time). As pointed out by Epstein and Schneider [7], the ambiguity aversion of an investor is not monotonically decreasing over time. Our RARA function allows the investor to adjust her portfolio weights dynamically, depending on his or her overall belief of how much his or her models can explain the true distribution P. The approach is rather different from a classical Baysian updating approach, where the investor learns more about the underlying model with any new information flowing in the stock price returns. In a Baysian framework, the investor believes that once he or she gets enough information, he or she will ultimately converge toward the true model and therefore is gradually and monotonically more and more confident about her model. Under model ambiguity however, this is not the case. The investor can become more or less confident over time, in a nonmonotonic way. He or she does not assume that more information can systematically give his or her more confidence about his or her model. Many methods could be used in order to calibrate the different measures π(Q), as well as the ambiguity aversion parameter γ Q for a given model Q. We propose a simple empirical methodology that takes into account the relative historical performance of the different models: First, we compute a number of performance measures (the Sharpe, Sortino, Gain Loss or Win Lose ratios, as described in Sec. 13.4.2) on the different models considered, evaluated over a given time window. The measure π can then be computed as a weighted average of the performance measures, whereas the ambiguity aversion parameter γ can be parametrized as the inverse of those performance measures. More formally, we will consider the following parametrization for our empirical example in Sec. 13.4: if we denote by PMpQ a given performance measure p for a given model Q, we have γQ ≡
1
(13.7)
PMpQ
and π(Q) ≡
PMpQ
P∈Q
PMpP
.
(13.8)
Here we see that the measure π is relative, as it takes into account all the different models performances, whereas γ is absolute, as it only considers the performance of a given
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
335
Robust Asset Allocation Under Model Risk
model. Note that the absolute ambiguity parameter is −γ rather than γ, according to the shape of the function ψ given in Sec. 13.2.1. Both the relative ambiguity parameter π and the absolute ambiguity parameter −γ are proportional to the performance measure considered: the higher the performance measure, the higher the parameters. Note that it is for the sake of simplicity that we chose to parametrize similarly π and γ, what matters is that the absolute and relative ambiguity aversion be positively correlated with the performance measure considered.
13.3. SOME DEFINITIONS RELATIVE TO THE AMBIGUITYADJUSTED ASSET ALLOCATION To compare different asset allocations for different models, and also the impact of the ambiguity aversion to the different weights assigned to each asset, we present in the following section some properties and give some definitions. We also propose a measure of distance between two different asset allocations that we will use to compare different asset allocations in our empirical example developed in the last section. The Ambiguity Robust Adjustment refers to the combine adjustment by the AARA to the optimal weights independently computed for each prior and by the RARA performed to combine those adjusted weights. These measures can be used to describe how to measure the ambiguity aversion of an investor toward a model in general, and more specifically toward a single asset. For the following definitions, let φQ be the asset allocation conditional on model Q ∈ Q and φARA be the ARA asset allocation. Definition 13.2 (Portfolio Distance). Let us consider two models Q1 and Q2 in the set of priors Q. We define the distance measure δ between the two models as δ(φQ1 , φQ2 ) =
N
φiQ1 − φiQ2 
i=0
δ(φQ1 , φQ2 ) represents the turnover value to rebalance the investor’s portfolio from the asset allocation φQ1 to the asset allocation φQ2 . Definition 13.3 (Weighted Ambiguity Adjustment). We denote by WAAi the ARAweighted ambiguity adjustment of an investor toward the asset i: φiARA − φiQ π(Q) ∀i ∈ [1, N], WAAi ≡ Q∈Q
the ambiguity aversion of an investor toward an ambiguous asset i. WAAi represents the average difference between the ARA weight for the asset i and the optimal weights for the asset i under the different priors. For two ambiguous
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
336
FA
Barrieu and Tobelem
assets Si and Sj ,j, i ∈ [1, N], if WAAi < WAAj , we say that the investor shows more ambiguity aversion toward the asset j than toward the asset i. Definition 13.4 (Absolute Ambiguity Adjustment). We define the value of the Absolute Ambiguity Adjustment (AAA) of an investor toward the model Q as AAA(Q) ≡
N
φiQ − ψ(φiQ , γ Q ).
i=1
AAA(Q) represents the theoretical turnover to rebalance the investor’s asset allocation on the risky assets i = 1, . . . , N from the optimal weights obtained if the prior Q is assumed to be equal to the unknown real model P to the Absolute Ambiguity Adjusted portfolio, taking into account the investor absolute aversion against his or her prior Q. We deduce that the total value of the investor’s Weighted Ambiguity Adjustment toward all her priors is AAA(Q) ≡
N
φiQ − ψ(φiQ , γ Q )π(Q).
Q∈Q i=1
Definition 13.5 (Relative Ambiguity Adjustment). We define the value of the Relative Ambiguity Adjustment (RAA) of an investor as N ARA Q − φi π(Q) , RAA(Q) ≡ φi i=1 Q∈Q which represents effectively the turnover between the Robust Ambiguity Portfolio and the Subjective Expected Utility Portfolio (the SEU portfolio allocation is defined as: φiSEU ≡ Q∈Q φiQ π(Q)).
13.4. EMPIRICAL TESTS In this section, we detail the results of some tests we run to evaluate the performance of ambiguity robust portfolios, compared to other classical optimized portfolios (we consider optimized portfolios as well as other type of factor model portfolios, and therefore we mainly focus on the ambiguity transformation and not on the initial optimization under each prior). The tests were run on European assets. We collected daily close prices from January 2000 to April 2008 (equivalent to 2167 business days). Our set of securities is the set of Eurostoxx 600 constituents that were trading across the whole period. For this empirical study, we consider a null riskfree rate (the cash is not remunerated), as we consider a portfolio total return, as would be done for strategies without benchmarks such as hedge funds or proprietary groups strategies. We assume that our transactions costs, fees, and slippage correspond to three basis point of the daily
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
337
Robust Asset Allocation Under Model Risk
turnover of the strategy (i.e., the total costs and fees are assumed to be in percentage three basis point of the daily turnover, for instance if the turnover of a strategy at a given date is 50% and its return is 10 basis points, the costs and fees amount to 1.5 basis points return and therefore the startegy return for this day after cost is 8.5 basis points). We run a back test on historical data when the investor rebalances daily her portfolio by reestimating the different models and resetting her optimal investment weights over an estimation window of 120 days (we choose a daily rebalancing, as we consider daily returns for the asset time series). The performances of the different strategies were then evaluated. We find that our Ambiguity Robust approach allows the investor to achieve superior performance compared to that achieved by classically optimized portfolios.
13.4.1. Portfolios Tested We consider several models and their outcomes. More precisely, for each of them, we computed portfolio weights using an estimation window, as if each one was the true one. For the minimumvariance portfolio and the mean–variance portfolio, we compute these weights according to the classical Markowitz framework. For the other portfolios, we give some details below on how we compute the different weights. We focus here on how the investor deals with her model ambiguity after the different model portfolio weights have been computed. We use an estimation window to estimate the following portfolio weights: • The equally weighted portfolio (EW): gives an equal weight to all the risky assets. We define the EW portfolio asset allocation as ∀i ∈ {1, . . . N} : φtEW,i =
1 . N
• The minimum variance portfolio (MN): is the fully invested Markowitz efficient portfolio with minimum variance, obtained when the investor minimizes the expected variance of the portfolio (i.e., minφ φ φ, where represents the covariance matrix of the stock returns). The MN portfolio allocation is defined as φtMN =
−1 t−w,t−1 1 1 −1 t−w,t−1 1
where t−w,t−1 is the empirical covariance matrix estimated over the window [t − w, t − 1] and 1 is the Nvector of ones. • The mean variance portfolio (MV): is the fully invested, maximum Sharpe, mean– variance Markowitz efficient portfolio, obtained when the investor maximizes the empirical quadratic expected utility (i.e., maxφ µφ −φ φ, where µ is the empirical
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
338
FA
Barrieu and Tobelem
mean and the empirical covariance matrix of the stock returns, note that we consider a risk aversion equal to 1). The MV allocation is defined as φtMV =
−1 t−w,t−1 µt−w,t−1 1 −1 t−w,t−1 µt−w,t−1
,
where t−w,t−1 is the empirical covariance matrix estimated over the window [t − w, t − 1] and µt−w,t−1 is the empirical vector of mean returns estimated over the same window. • The CAPM portfolio weights (CA): we base our CAPM portfolio on the Jensen alphas. We estimate the CAPM betas over the estimation window. ConsiderM ing rt−w:t−1 as the vector of the Eurostoxx 600 market returns over the period [t − w, t − 1], the beta of the risky asset i is therefore estimated at time t as βti ≡
i M cov(rt−w:t−1 , rt−w:t−1 ) . M var(rt−w:t−1 )
We then compute the Jensen alpha as the difference between the observed return at time t of the asset i and the betaadjusted market return: αit ≡ rti − βti rtM . We then define the CAPM weights as the weighted average alphas across all the risky assets considered: αi φtCA,i ≡ N t j=1
j
αt
.
• The CAPM uncertain portfolio (UC): we define the UCAPM weights as the CAPM weights adjusted by the variance of the CAPM residuals: φtUC,i
αit var(αi. )
≡ N
j
αt j=1 var(αj. )
.
We also consider three fundamental portfolios based on stock specific financial ratios: • The price earning portfolio (PE): we compute the relative price earning ratio (PER) return of a stock among its sector peers and we give positive weights to the lower PER stocks and negative weights to the higher ones. • The cash flow portfolio (CF): we compute the relative cash flow ratio (CFR) return of a stock among its sector peers and we give positive weights to the lower CFR stocks and negative weights to the higher ones. • The price to book portfolio (PB): we compute the relative price to book ratio (PBR) return of a stock among its sector peers and we give positive weights to the lower PBR stocks and negative weights to the higher ones.
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
339
Robust Asset Allocation Under Model Risk
More formally, we compute the relative financial ratio FR of the asset i among its peers assets j in the sector S: i ≡ µ(FR.,S αFR,i,S t t ) − FRt ,
where µ(FR.,S t ) is the average of the given financial ratio among the assets in the sector S. And therefore we get the following asset allocation, by scaling the alphas per sector, so that the investor invests the same amount of money in each sector: φtFR,i ≡ N S
αFR,i t FR,j,S
j=1
αt

,
where FR stands either for PE, CF, or PB and NS is the number of assets belonging to sector S. • The subjective expected utility portfolio (SEU): we define by π the vector of Relative Ambiguity Aversion Weights given to the models considered. In the present empirical example, we assume that Q∈Q π(Q) = 1, therefore the SEU portfolio is defined as the different models weights weighted by π. ∀i ∈ {1, . . . , N} : φtSEU,i ≡ φQ,i π(Q). Q∈Q
• The ambiguity robust portfolio (RA): finally, the optimal ambiguous portfolio is defined as the different models weights adjusted by the Absolute Ambiguity Adjustment ψ and weighted by the vector of Relative Ambiguity Aversion Weights π: ∀i ∈ {1, . . . , N} : φtARA,i ≡ ψ(φQ,i , γ Q )π(Q). Q∈Q
13.4.2. Performance Measures In order to parametrize the absolute ambiguity parameter γ of the function ψ and the relative ambiguity function π, to compute the SEU and RA portfolios, we use different portfolio performance measures: φ Given a portfolio allocation φ, we denote by rt the return of this portfolio at time t. • Sharpe ratio (SHR): that represents at time t the ratio of the empirical mean return of a portfolio over its empirical standard deviation over the period [t − w, t − 1]. We denote: φ
φ
SHRt ≡
µ(rt−w,t−1 ) φ
σ(rt−w,t−1 )
.
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
340
FA
Barrieu and Tobelem
• Sortino Price ratio (SOR): that represents at time t the ratio of the empirical mean return of a portfolio over the empirical standard deviation of its negative returns over the period [t − w, t − 1]: φ µ(rt−w,t−1 )
φ
SORt ≡
φ
φ
σ(rt−w,t−1 {rt−w,t−1 < 0})
.
• Gain Loss ratio (GLR): which is the ratio of total positive returns over total negative returns: φ {rt−w,t−1 > 0} φ GLRt ≡ φ . {rt−w,t−1 < 0} • Winner Loser ratio (WLR): similar to the Gain Loss ratio, it is the ratio of the number of total positive returns over the number of total negative returns: φ φ rt−w,t−1 {rt−w,t−1 > 0} φ . WLRt ≡ φ φ rt−w,t−1 {rt−w,t−1 < 0} The following two measures are used to compare different portfolio performances. • Certain equivalent (CER): corresponds to the equivalent riskfree return of the portfolio return (the portfolio return adjusted for its risk aversionadjusted standard deviation), as used by DeMiguel et al. in their comparative study of portfolios performance [8]: φ
φ
φ
CERt ≡ µ(rt−w,t−1 ) − σ 2 (rt−w,t−1 ). • Turnover (T/O): corresponds to the change in portfolio weights from one rebalance period to the next. The investor aims to reduce the turnover as trading implies costs (exchange fees, price impact, etc.): φ
T/Ot ≡
N
i φti − φt−1 .
i=0
As mentioned in Sec. 13.2.3, we consider as an estimate for π the relative performance measure of a model within the class of models. Similarly, γ is estimated by the inverse absolute performance measure of a model (the worse the performance, the bigger the ambiguity aversion). Note that if the performance measure of a model is zero, then π and γ are defaulted to zero.
13.4.3. Results First we describe the individual performances of the eight models considered. Then, we show how the Subjective Expected Utility portfolios of Savage (as defined in Eq. (13.2) and where the single strategies are linearly weighted by the weighted average of their
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
341
Robust Asset Allocation Under Model Risk
different performance measures) outperform the individual strategies. We use the SEU portfolios as benchmarks for our ARA methodology, as the Klibanoff et al. model is almost impossible to compute for a large universe of assets. Finally, we display the performances of the ambiguity robust portfolios, parametrized by the four different performance measures considered. We conclude that the ambiguity robust portfolios outperform by far the nonambiguous SEU portfolios.
13.4.3.1. Performances of the different models We display in Table 13.1 the statistics of the different strategies, computed over the over the period January 2000 to April 2008 and annualized. We use the four performance measures described previously to parametrize π and γ and construct our ambiguity robust portfolios. At each date t, we compute the
Table 13.1 UC µ (%) µ ¯ (Bps) σ (%) max(µ) (Bps) min(µ) (Bps) SHR SOR GLR (%) WLR (%) CER (Bps) T/O (%)
8.71 0.43 3.95 210.82 −139.12 0.27 0.40 104.94 100.60 0.39 135.01 EW
µ (%) µ ¯ (Bps) σ (%) max(µ) (Bps) min(µ) (Bps) SHR SOR GLR (%) WLR (%) CER (Bps) T/O (%)
69.74 3.41 15.26 470.80 −540.82 0.56 0.72 110.48 119.16 2.94 2.89
Strategies Performances. PE 11.04 0.54 4.38 193.83 −217.76 0.31 0.44 105.88 102.52 0.50 141.51 MN 35.03 1.71 6.77 199.81 −317.57 0.63 0.75 112.05 123.96 1.62 21.85
CF 5.42 0.26 5.06 225.39 −205.43 0.13 0.19 102.48 99.11 0.21 140.83 MV 10.68 0.52 4.75 133.63 −219.94 0.27 0.32 104.98 122.98 0.48 28.60
PB 1.89 0.09 4.93 234.74 −234.40 0.05 0.07 100.88 98.72 0.04 141.59 CA 13.95 0.68 4.71 182.75 −162.10 0.36 0.52 106.77 101.00 0.64 135.31
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
342
FA
Barrieu and Tobelem
performance measure of the different strategies over the historical returns between date t − 120 and t − 1. Even though the strategies UC and CA returns underperform the EW strategy return over the whole period considered (the equally weighted portfolio has the biggest total return of almost 70%), there exist some periods where the reverse is true (in 2002 for instance, the CA and UC portfolios perform better). This confirms that the relative weight of each models should be dynamic, as proposed in our ambiguity robust methodology. In order to have numbers of similar magnitude, we scale the inverse Sharpe and Sortino measures by a reference ratio of three annualized: in practice, it means that an investor considers a portfolio for which the return is three times as big as its risk as benchmark. We have also capped the performance measures by three to prevent a strongly performing model dominating the others. Formally, we have: 3 3 ), 3) and SOR = min(max(0, SOR ), 3). SHR = min(max(0, SHR
13.4.3.2. SEU portfolio In Table 13.2 are displayed the performances and statistics for the SEU portfolios that weight the different models linearly with the measure π (respectively computed with the four different performance measures Sharpe, Sortino, Gain Loss and Win Lose Ratios), without considering the absolute ambiguity adjustment from the function ψ. The SEU portfolios outperform almost all the single strategy portfolios. The SEU portfolios outperform the individual strategies, however, we improve greatly these performances with our Ambiguity Robust approach as we show below.
13.4.3.3. Ambiguity robust portfolios We have computed the performance of the four different Ambiguity Robust portfolios (where the ambiguity parameters γ and π are estimated using the four different Table 13.2
SEU Strategies Performances. SHR
Total return (%) Mean daily return (Bps) Volatility (%) Max return (Bps) Min return (Bps) SHR SOR GLR (%) WLR (%) CER (Bps) T/O (%)
77.59 3.79 7.06 326.58 −411.38 1.34 1.52 128.85 135.45 3.69 76.33
SOR 75.26 3.68 6.92 322.47 −412.40 1.33 1.50 128.59 135.09 3.58 78.26
GLR 55.72 2.72 6.89 310.73 −275.82 0.99 1.25 119.60 121.06 2.63 112.96
WLR 59.58 2.91 6.95 321.65 −270.02 1.05 1.33 120.95 123.23 2.81 112.65
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
FA
343
Robust Asset Allocation Under Model Risk Table 13.3 Ambiguity Robust Strategies Performances. SHR µ (%) µ ¯ (Bps) σ (%) max(µ) (Bps) min(µ) (Bps) SHR SOR GLR (%) WLR (%) CER (Bps) T/O (%) RAA AAA
158.37 7.74 5.22 271.44 −206.80 3.71 5.59 196.40 158.16 7.68 137.35 0.25 0.24
SOR 157.04 7.67 5.27 279.05 −209.66 3.64 5.44 193.90 155.50 7.62 138.88 0.23 0.14
GLR 151.82 7.42 5.70 262.80 −189.94 3.25 4.66 181.26 155.36 7.35 136.28 0.34 0.20
WLR 149.43 7.30 6.11 264.90 −186.81 2.98 4.22 172.80 155.49 7.23 134.73 0.37 0.35
performance measures: Sharpe, Sortino, Gain Loss and Win Lose ratios). In Table 13.3, we display the statistics of the Ambiguity Robust portfolios estimated with the γ and π measures. The RA portfolios outperformed the best CA strategy portfolio in terms of all four different performance measures considered. The Absolute Ambiguity Aversion (AAA) and the Relative Ambiguity Aversion (RAA) measures allow us to quantify our ambiguity adjustment. The biggest ambiguity adjustments are made for the less performing strategies (based on Win Lose and Gain Loss ratios). As we can see, the Ambiguity Robust Portfolios outperform all the SEU portfolios, meaning that the ψ adjustments enhance the performance of ambiguity averse investors portfolios.
13.5. CONCLUSION Our aim was to provide the reader with a simple and easy to implement methodology for practitioners who want to allocate in a robust way their portfolio of assets when they have several models for the asset returns distributions but they are ambiguous about each of them (i.e., they do not fully trust them). We propose a simple, robust, and systematic method of combining the different weights computed conditionally per model. The concrete methodology we develop enables to account for ambiguity via practical empirical measures, such as the performance measures of each conditional portfolio. Our method is very different from the approach proposed in the literature and in particular that of Klibanoff et al. as we do not proceed to a unique complex optimization that is extremely challenging to solve in practice. Our approach is more practical and
May 12, 2010
17:47
WSPC/SPIB913
b913ch13
344
FA
Barrieu and Tobelem
very industryoriented: practitioners can easily define the optimal weights for each of the models they consider, then they can mix those prior conditional weights, taking into account the absolute and relative ambiguity they have against each model, as reflected by the concavity of ψ and the weight π(Q) attributed to each model Q. The parametrization we propose for the Absolute Ambiguity Aversion and the Relative Ambiguity Aversion in our empirical example is by no means optimal. However, we have shown that it can greatly enhance the performance of some classical portfolio strategies. The proposed robust ambiguity portfolio is more stable, with a lower turnover, and less risky, with a higher certainty equivalent ratio, than unadjusted portfolios as well as SEU portfolios. The ARA methodology smoothens and reduces the risk of classical strategies. One of the main features of the ARA methodology is to provide a dynamic adjustment for model uncertainty. As the parametrization depends on the past performance of the strategy, it would be interesting to study how the ARA model would behave in major downturns, and adapt to different market regimes.
References [1] Markowitz, H (1952). Portfolio selection. Journal of Finance, 7, 77–91. [2] Savage, L (1954). The Foundations of Statistics. New York: Wiley. [3] Ellsberg, D (1961). Risk, ambiguity and the savage axiom. Quarterly Journal of Economics, 75, 643–669. [4] Gilboa, I and D Schmeidler (1989). Maxmin expected utility with a nonunique prior. Journal of Mathematical Economics, 18, 141–153. [5] Klibanoff, P, M Marinacci and S Mukerji (2005). A smooth model of decision making under ambiguity. Econometrica, 73, 1849–1892. [6] Barrieu, P and S Tobelem (2009). Robust asset allocation under model risk. Risk Magazine. [7] Epstein, L and M Schneider (2007). Learning under ambiguity. Review of Economic Studies, 74(4), 1275–1303. [8] DeMiguel, V, L Garlappi and R Uppal (2009). Optimal versus naive diversification: How inefficient is the 1/n portfolio strategy? Review of Financial Studies, forthcoming.
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
14
SEMISTATIC HEDGING STRATEGIES FOR EXOTIC OPTIONS∗
HANSJÖRG ALBRECHER†,§ and PHILIPP MAYER‡,¶ †
Institute of Actuarial Science, University of Lausanne, Quartier UNILDorigny, Bâtiment Extranef, 1015 Lausanne, Switzerland ‡ Department of Mathematics, Graz University of Technology, Steyrergasse 30, 8010 Graz, Austria § hansjoerg.albrecher@unil.ch ¶ mayer@finanz.math.tugraz.at
In this chapter, we give a survey of results for semistatic hedging strategies for exotic options under different model assumptions and also in a modelindependent framework. Semistatic hedging strategies consist of rebalancing the underlying portfolio only at certain prespecified timepoints during the lifetime of the hedged derivative, as opposed to classical dynamic hedging, where adjustments have to be made continuously in time. In many market situations (and in particular in times of limited liquidity), this alternative approach to the hedging problem is quite useful and has become an increasingly popular research topic over the last years. We summarize the results on barrier options as well as strongly pathdependent options such as Asian or
∗ Supported
by the Austrian Science Fund Project P18392. 345
May 19, 2010
16:20
WSPC/SPIB913
346
b913ch14
FA
Albrecher and Mayer lookback options. Finally, it is shown how perfect semistatic hedging strategies for discretely observed options can be developed in quite general Markovtype models.
14.1. INTRODUCTION In the famous work [10] of 1973, it was shown how a standard European option can be replicated by dynamically trading the underlying asset and investing in the riskless bond. The socalled delta hedging strategy is (with some modifications) still among the most widely used methods to manage and reduce the risk inherent in writing an option. However, already in [60] static “hedging” methods were applied to relate different option prices, and also the Put–Call–Parity, which at least goes back to [76], can be interpreted as a consequence of a particular static hedging strategy. Recently, in [39], some pitfalls of the dynamic replication strategies were outlined and it was argued in favor of the more robust static replication methods, which in the meantime attained quite some attention in the practical and academic literature, starting in the mid1990’s with [14, 37]. The semistatic hedging approach structurally differs from the dynamic counterpart in terms of hedging instruments and trading times. While dynamic hedges base on the assumption of being able to trade continuously in time, semistatic portfolios do not need to be adjusted dynamically but only at prespecified (stopping) times. Furthermore, in contrast to classic dynamic hedging, they typically use plain vanilla (or other liquidly traded) options as hedging instruments. Studies comparing the performance of semistatic and dynamic hedging strategies (as for example undertaken in [43, 64, 79]) indicate that the semistatic ones are in many situations more robust and able to outperform their dynamic counterparts. Here, we aim to give an overview of the most prominent semistatic hedging strategies for options written on some asset (e.g., a stock or an exchange rate), for which some standard European options are liquid. In this chapter, we focus on fully discrete strategies, i.e., we exclude the possibility of dynamic hedging (for combined dynamicstatic hedging strategies see [52, 53]). In principle, the outlined methods directly translate to the case where the options are written on baskets of assets, but one has to be aware that plain vanilla options with the basket as the underlying are needed to apply the strategies in practice. Strategies, for which the composition of the portfolio is not changed after the initial composition except investing gains in the riskless asset are called static, and those with a finite number of transactions are termed semistatic (semistatic strategies can be further subclassified as done in [55]). We also distinguish between modeldependent and modelindependent strategies. The former depend on assumptions on the asset price process (like continuity) or need
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
SemiStatic Hedging Strategies for Exotic Options
347
a fully specified model, while the latter are correct in any market, which is frictionless in the following sense: Definition 14.1 (Frictionless market). A market is called frictionless, if • • • • • • •
The noarbitrage assumption holds. All investors are pricetakers. All parties have the same access to relevant information. There are no transaction costs, taxes, or commissions. All assets are perfectly divisible and can be traded at any time. There are no restrictions on shortselling. Interest rates for borrowing and lending are identical.
Throughout the text, we assume frictionless markets. Although the methods work for timedependent interest rates as well, we (mainly for notational convenience) further assume that the riskless interest rate is a constant r. Definition 14.2 (Weak and strong pathdependence). An option is said to be weakly pathdependent if the dynamics of the option price depend only on asset price and time, i.e., the stochastic differential equation (SDE) for the option price differs from the SDE of a vanilla option only in the boundary conditions. Otherwise the option is said to be strongly pathdependent. While for weakly pathdependent options, quite general methods are available, the theory is somewhat more involved for strongly pathdependent options. Nevertheless, we present hedging strategies for lookback options, as well as for Asian options, and discretely sampled options in general. The rest of the chapter is structured as follows. Section 14.2 introduces a method to synthesize arbitrary (sufficiently regular) payoffs that depend solely on the asset price at some later time T . In Sec. 14.3 techniques for weakly pathdependent options are outlined. The focus lies in particular on barrier options, which are the most liquid and best understood exotic options of this kind. Section 14.4 deals with two examples of strongly pathdependent options (the lookback and the Asian option) and shows some robust hedging strategies. In Sec. 14.5, discretely sampled options, which may be weakly pathdependent (e.g., discretely monitored barrier options) or strongly pathdependent (e.g., Asian options) are considered. Finally, Sec. 14.6 concludes and points out some potential future research topics.
14.2. HEDGING PATHINDEPENDENT OPTIONS Now consider options with a pathindependent payoff. To fix ideas let us denote the price process of some asset S up to time T by ST = {Ft }0≤t≤T and the payoff function
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
348
FA
Albrecher and Mayer
by p, so that p(ST ) = p(ST ). Moreover we assume that there is a riskless bond available which pays off 1 at time T .
14.2.1. Plain Vanilla Options with Arbitrary Strikes are Liquid First we suppose that standard European options with maturity T are liquid for all strike levels K ≥ 0, and the payoff function p is assumed to be twice differentiable. Then the Taylor expansion implies S p(S ) = p(K∗ ) + p (K∗ )(S − K∗ ) + p (x)(S − x)dx. K∗
Since (S − x) = (S − x)+ − (x − S )+ , we can rewrite this to obtain p(S ) = p(K ∗ ) + p (K∗ )(S − K∗ ) ∞ + + p (x)(S − x) dx + K∗
K∗
p (x)(x − S )+ dx.
(14.1)
0
Note that (14.1) holds for any K ∗ and even extends to cases where p is not twice differentiable (e.g., for a convex payoff function, p should then be understood as left derivative and p as a positive measure). For hedging purposes, we can rewrite (14.1) as p(S ) = p(K∗ ) + p (K∗ )(F − K ∗ ) + p (K∗ )(S − F ) K∗ ∞ p (x)(S − x)+ dx + p (x)(x − S )+ dx, + K∗
0
where F denotes the forward price of the asset. In this way, the payoff of the contingent claim p is decomposed into four parts, two of which are hedgeable by static positions in the riskless bond and a forward on the asset. The other two are synthesized by “infinitesimal” positions in standard European options. In total we get the following static hedging strategy (cf. [22]): Payoff
Hedged by positions in
p(K ∗ ) + p (K∗ )(F − K ∗ )
(p(K ∗ ) + (F − K ∗ )p (K ∗ )) bonds
p (K∗ )(ST − F ) ∞ + K ∗ p (x)(ST − x) dx K∗ + 0 p (x)(x − ST ) dx
p (K∗ ) forwards p (x)dx calls struck at x
(∀x ∈ [K∗ , ∞))
p (x)dx puts struck at x
(∀x ∈ [0, K∗ ))
This means that if European calls and puts with maturity T are liquid for all strikes, any (sufficiently regular) contingent claim on the timeT price of S can be replicated
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
SemiStatic Hedging Strategies for Exotic Options
349
perfectly — regardless of the model assumptions and in particular also in incomplete market models (this fact was already noticed e.g., by [11, 47, 62, 71]). A nice application of Formula (14.1) was given in [22], where it was shown that a variance swap written on some forward F can be replicated by a contingent claim with payoff ln(FT ) and a dynamic trading strategy involving the forward (see also e.g., [18, 19, 38, 42] for further details). It is worth noting that (14.1) also yields a pricing formula for the contingent claim. Denote by C(Ft , t; K, T ) and P(Ft , t; K, T ) the call and put price at current time t, respectively. Since due to the first fundamental theorem of asset pricing (see [35]) the price of the claim in an arbitragefree setup is given as the riskneutral expectation (conditioned on Ft ) of the discounted payoff, we have (the forward contract has value 0 by definition) EQ [e−r(T −t) p(ST )Ft ] = e−r(T −t) p(K∗ ) + e−r(T −t) (Ft − K∗ )p (K∗ ) K∗ ∞ p (x)C(Ft , t; x, T )dx + p (x)P(Ft , t; x, T )dx, (14.2) + K∗
0
where interchanging the order of integration is justified whenever the integrands in this last formula are absolutely integrable. A straightforward corollary to the above formula is the familiar PutCallParity C(Ft , t; K, T ) = e−r(T −t) (Ft − K) + P(Ft , t; K, T ),
(14.3)
which is obtained by setting p(S ) = (S − K)+ and choosing K∗ > K (Ft denotes the time tforward price of the asset). Note that (14.1) furthermore implies the following static hedging portfolio for the call payoff: • A long position in a put with the same strike K and maturity T as the call • A long position in a forward contract • Investing exp(−rT )(Ft − K) in the bond It is easy to see that the payoff of this portfolio at time T is the same as provided by the vanilla put, and the hedging strategy is therefore correct.
14.2.2. Finitely Many Liquid Strikes The derivation of the perfect hedging strategies in the last section relied on the assumption that standard European options are liquid for arbitrary strikes. In reality, of course, there are only a finite number of maturity/strike combinations available on the market. The first question in this more realistic context is whether the option prices themselves are consistent with the noarbitrage assumption. This question is, for instance. addressed in [21] and in a general setting in [32] (see also [28]).
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
350
FA
Albrecher and Mayer
Now, given a set of traded option prices, one can derive noarbitrage bounds for the price of contingent claims by solving a semidefinite optimization problem (see [9,51], where also more general options are considered). The dual of this problem is then to find the most expensive sub and the cheapest superreplicating strategy using static positions in liquid options. The latter question is obviously highly relevant in practical applications, as it shows an (in some sense) optimal way to cover the payoff in any case. Alternatively the right hand side of (14.1) may be approximated by a sum, where the sampling points correspond to the liquid strikes. However, then one cannot ensure that the payoff of the hedging portfolio is sufficient to cover the payoff of the contingent claim, unless the weights of the liquid options are chosen with great care.
14.3. HEDGING BARRIER AND OTHER WEAKLY PATH DEPENDENT OPTIONS The prototype of a weakly pathdependent option is the barrier option, which, starting from [14], has attracted by far the most attention in the literature of semistatic hedging. The payoff of a barrier option generally depends on whether the asset price has reached some prespecified region up to maturity or not. For example, the DownandOut call (DOC) with parameters strike K, barrier B < S0 , and maturity T has the payoff p(ST ) = (ST − K)+ 11{inf 0≤t≤T Ft >B} ,
(14.4)
where 11A denotes the indicator function of the set A. Analogously the payoff of a DownandIn call (DIC) is defined by p(ST ) = (ST − K)+ 11{inf 0≤t≤T Ft ≤B} . In a similar manner the payoff of UpandOut and UpandIn calls (UOC and UIC) are defined by p(ST ) = (ST − K)+ 11{sup0≤t≤T Ft t,
where Ws is a standard Brownian motion and σ(Fs , s) = σ(Ft2 /Fs , s).
(14.7)
Condition (14.7) implies that the local volatility at a fixed time s plotted as a function of the logreturns is symmetric around 0, which explains also the symmetry of the implied volatility surface. Starting from Theorem 14.1, in [14] it is investigated how the Put–Call–Symmetry can be used to hedge barrier options (a similar relation between different call and put prices, termed Put–Call–Reversal, was used in [6] to hedge longterm call options). Consider a DOC with strike K, maturity T , and barrier B < K. By definition, the payoff of this option is (ST − K)+ , if the barrier was never hit and 0 otherwise. In order to hedge the final payoff, a vanilla call with the same strike and maturity as the DOC should be bought. Clearly this is a superhedge and another vanilla option can be sold such that one obtains a portfolio with value 0 at the barrier. Assuming the conditions for the Put–Call–Symmetry at the first hitting time τ, we know that a put with the same maturity and strike B2 /K fulfills the geometric mean condition of the theorem. Therefore C(B, τ; K, T ) =
K P(B, τ; B2 /K, T ). B
Hence a portfolio consisting of a long position in the corresponding vanilla call and a short position in K/B puts with strike B2 /K has value 0, if the forward price is exactly at the barrier. Consequently, a replicating portfolio for a DOC is to take a long position in a standard European call with strike K and maturity T and a short position in K/B puts with strike B2 /K and the same maturity. The corresponding trading strategy is to hold the portfolio until expiry, if the barrier is never hit, or to close the positions (at 0
May 19, 2010
16:20
WSPC/SPIB913
354
b913ch14
FA
Albrecher and Mayer
cost) at τ ≤ T . Note that for the correctness of the replication strategy it is necessary that the forward price is exactly B at the first hitting time (which explains the continuity assumption). Summing up, under the noarbitrage principle this leads to the equation: K (14.8) P(F, 0; B2 /K, T ). B This hedging procedure can be extended to other types of barrier options but then typically becomes more complicated. For example, for an UOC one would also need to take a position in a binary call (i.e., an option, which pays 1 if the asset price at maturity exceeds the strike, and 0 otherwise) and the latter one can be hedged by the strategies outlined in the previous section. If the assumption of a constant barrier on the forward price is relaxed, super and subreplication strategies can be derived. For the details we refer to [17] (see also the nice summary [69]). In [56], the case when only a finite number of options with prespecified strikes are available is considered and conditions under which a strategy based on the Put–Call–Symmetry is a semistatic superhedge are identified. Another hedging strategy for barrier options has its roots in papers of Derman et al. [36,37]. In order to point out the main idea of the Derman–Ergener–Kani (D–E–K) algorithm, a binomial model is considered first, i.e., the asset price process is assumed to follow a binomial tree. In this setting the D–E–Kalgorithm for a DOC with maturity T , strike K, and barrier B is as follows: DOC(F0 , 0; K, T, B) = C(F0 , 0; K, T ) −
• Identify the boundary nodes, i.e., the maturity and the barrier nodes and denote the resulting time grid by t1 < · · · < tn ; • Buy a standard European option replicating the payoff of the barrier option if the barrier boundary is not hit during the lifetime of the barrier option. In the case of a DOC with B < K this would be a standard European call with the same strike and maturity as the barrier option; • Choose n options with different maturities Ti and strikes Ki having payoff 0 if the barrier was not hit. In the example of a DOC, one could for instance choose puts with strikes Ki ≤ B and maturities Ti ≤ T . Then solve the linear equation system Vi · w + C(B, ti ; K, T ) = 0,
i = 1, . . . , n
(14.9)
for w, where Vi := (P(B, ti ; K1 , T1 ), P(B, ti ; K2 , T2 ), . . . , P(B, ti ; Kn , Tn )) denotes the vector of the timeti prices of the n additional options and · is the scalar product. Note that due to (14.9) the portfolio consisting of the option replicating the final payoff and wi options with strike Ki and maturity Ti can be sold at zero cost at every barrier node and thus the semistatic trading strategy of holding the portfolio until expiry in
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
355
SemiStatic Hedging Strategies for Exotic Options
case the barrier has not been hit, or to close the positions at the first hitting time, is a perfect replication of the barrier option. In the binomial framework the D–E–Kalgorithm thus gives a perfect hedge for all kinds of barrier options — in particular also for doublebarrier options and options with a timedependent barrier — as long as enough different standard European options are liquid (i.e., at least as many as there are barrier nodes). However, since the prices in (14.9) are modeldependent (as they have to be calculated using a model — in the above case the binomial model), the entire portfolio is modeldependent. For continuoustime models, the approach outlined in [37] is to discretize (in time) the asset price process and hedge the values at the discrete monitoring times. In [78], the effectiveness of such a discretization for stochastic volatility is examined and it is found that it works well for a small volatility of volatility, but not satisfactory in the opposite case combined with discontinuous final payoff of the barrier option (as for an UOC). We come back to better alternatives in the next section. A possible limit strategy for continuoustime price models can be achieved by assuming standard options with all strikes and maturities to be liquid. This strategy was developed by Andersen et al. [5] (see also [61]), who observed that the key assumption on the price model is that the price of European options, besides their parameters, only depends on the current time t and asset price Ft (this assumption for instance rules out stochastic volatility models). The approach is, as the D–E–Kalgorithm, very powerful in terms of the barrier structure and allows for arbitrary (sufficiently regular) functions of time. For the ease of exposition we assume that the forward price again follows a local volatility model in the riskneutral world (the method can be extended to more general Markovtype models), i.e., dFt = σ(Ft , t)dWt , Ft
(14.10)
where σ is now just assumed to be positive and sufficiently regular to admit a unique solution of Eq. (14.10). Consider again the standard example of a DOC on the forward price of an asset with dynamics as specified in (14.10). Then the following PDE specifies the price function G(Ft , t) of a DOC with strike K, barrier B, and maturity T started at time t (and has in particular not reached the barrier up to the time t): 1 Gt (F, t) + σ 2 (F, t)F 2 GFF (F, t) = 0, 2 G(F, t) = 0,
t < T, F > B t < T, F ≤ B
G(F, T ) = (F − K) where we assumed again for simplicity, that B < K.
+
∀F,
(14.11)
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
356
FA
Albrecher and Mayer
Then G(F, t) = DOC(F, t; K, T, B), in case the asset price has not crossed the barrier before t. If the function σ in (14.11) is sufficiently regular, G(F, t) is twice differentiable with respect to F , except at F = B, where the (formal) second derivative is given by δ(F − B)G+ F (B, t) (δ denoting the + Dirac Delta measure and GF the right derivative). Using (14.11) and a generalized Meyer–Itôformula (see [68]), one finds: T 11{Ft >B} GF (Ft , t)Ft σ(Ft , t)dWt (FT − K)+ = G(F0 , 0) + 0
1 + 2
T
0
B G+ F (B, t)dLt ,
where LBt is the local time of F at B and hence fulfills 1 →0
LBt = lim
T
0
11{B≤Fs ≤B+} σ 2 (Fs , s)Fs2 ds
a.s.
Rearranging the terms gives: T 1 T + + G(F0 , 0) + 11{Ft >B} GF (Ft , t)dFt = (FT − K) − G (B, t)dLBt . 2 0 F 0
(14.12)
(14.13)
Using (14.12) and (14.13), one might deduce the following semistatic hedge for the DOC: take a long position in a call with strike K and maturity T and short positions in G+ F (B(t), t)dt options with maturity t (∀ 0 ≤ t ≤ T ) and payoff 11{B≤Ft ≤B+} σ 2 (Ft , t)Ft2 / ( very small). Note that the latter payoff can, for any > 0, be approximated arbitrarily closely by standard options (we assumed options for all maturities and strikes to be liquid). The corresponding trading strategy is to hold the portfolio until expiry, if the barrier is not reached and to sell the portfolio at the first hitting time τ. This is in fact a nearly perfect replication strategy, which can be seen by considering the limiting case for → 0: if the barrier is not reached, then the payoff of the replicating portfolio equals the payoff of the call option and thus the payoff of the DOC. On the other hand, with (14.13) we have T E G(F0 , 0) + 11{Ft >B(t)} GF (Ft , t)dFt Fτ 0 +
T
= E[(FT − K) Fτ ] − E 0
1 + B G (B, t)dLt Fτ , 2 F
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
SemiStatic Hedging Strategies for Exotic Options
357
which can be reformulated as τ 1 τ + 11{Ft >B(t)} GF (Ft , t)dFt + G (B, t)dLBt G(F0 , 0) + 2 0 F 0 T 1 + B = C(B(τ), τ; K, T ) − E GF (B, t)dLt Fτ , 2 τ where the righthandside is exactly the timeτprice of the portfolio. Note that at this point it is crucial that the prices of European options only depend on the asset price at time τ in order to interpret the expectations as prices. Since τ 1 τ + 11{Ft >B(t)} GF (Ft , t)dFt + G (B, t)dLBt = G(B(τ), τ) G(F0 , 0) + 2 0 F 0 and G(B(τ), τ) = 0 by definition, the value of the portfolio at time τ is 0. Thus we can close the positions at zero cost and replicate the payoff of the DOC. This replication strategy can be generalized to all kinds of terminal payoffs, barrier regions, rebates, and also to jumpdiffusion processes. We refer to [5] for details. An interesting difference to hedging strategies using the Put–Call–Symmetry is that here not only the correctness, but also the strategy itself depends on the model. More precisely: the number G+ F (B(t), t) has to be calculated using a model, while the strategy outlined before is independent of the exact specification of the local volatility, as long as Assumption 1 is fulfilled.
14.3.2. ModelDependent Strategies: Approximations The hedging strategies in the previous sections were perfect if the model was correct, i.e., the risk of the option could be completely eliminated. In particular, this gave further insight into the underlying models and into relations between barrier options and standard European options. However, the perfect replication could only be achieved at the cost of more or less severe assumptions like liquidity of arbitrary standard European options or restrictions on the model. In this section a different and in some sense more pragmatic approach to semistatic hedging is discussed. In the previous section we discussed an algorithm for a binomial model (although already Derman et al. generalized the approach to continuous time models). A key point of this generalized D–E–K algorithm is that the barrier is hit exactly and that the prices of the European options only depend on the asset price. This assumption is too restrictive, as for example stochastic volatility or price jumps cannot be accommodated in this framework. Therefore, a lot of research in recent years has been devoted to generalize the assumptions, but keeping the main idea of matching the barrier option value at the first hitting time. Examples of such studies are [4,44,46,58,59,63,64,75].
May 19, 2010
16:20
WSPC/SPIB913
358
b913ch14
FA
Albrecher and Mayer
To outline the basic ideas of these generalizations, we focus again on a DOC. As in the D–E–K algorithm, we take a long position in a call with the same maturity and strike as the DOC to cover the payoff at maturity in the case the barrier was not hit. Thus the aim is to take positions in other options, such that the value of the portfolio is 0 if the barrier is hit. Furthermore the additional options should not have any payoff if the asset price stays above the barrier during the lifetime of the barrier option. In a binomial world this could be done by solving equation system (14.9). For general asset price dynamics, this equation system has to be replaced by something more general and the method of choice of the above cited studies is to consider different scenarios for the portfolio value at the hitting time. To fix ideas, let denote the set of possible scenarios for the evolution of the barrier option. Then an element θ ∈ consists of • The time when the barrier is hit (the hitting time) • The undershoot of the asset price under the barrier • The form of the implied volatility surface at the hitting time (i.e., The prices of the used European options at the hitting time) Thus we have to solve Vθ · w + Cθ (Fτθ , τθ ; K, T ) = 0,
θ ∈ ,
(14.14)
where Vθ is the price vector of the options and Cθ (Fτθ , τθ ; K, T ) is the price of the call option at the hitting time τθ under scenario θ. Note that the asset price Fτθ ≤ B(τθ ) is allowed to undershoot the barrier. As is an infinite set in general, (14.14) consists of a continuum of linear equations, which is obviously not feasible. Hence the set has to be discretized to make (14.14) a finitedimensional equation system. This can be done by assuming a model and using MonteCarlo simulation (see e.g., [46, 63, 64, 75]) or by an apriori choice of “reasonable” scenarios [4, 59]. The implied volatility surface at the hitting time, i.e., Vθ and Cθ , can be calculated using a prespecified model, or can be allowed to vary in some reasonable manner (e.g., by allowing the parameters of the model to take a set of values [59]) in order to make the hedge robust with respect to model risk. After a discretization of (14.14), the number of equations, albeit finite, may be huge, if a large number of different scenarios is considered. However, since the number of scenarios is directly linked to the error due to discretization, it should in fact be chosen large in order to obtain robust results. As in general only a limited number of different options is liquid, we cannot hope to be able to solve even a discretized version (14.14) exactly. To overcome this problem, several approaches are feasible. [75] (see also [63, 64]) proposes to minimize a risk measure of the hedging error under the budget condition that the price of the hedging portfolio is smaller or equal to the one of the barrier option. Alternatively, in [46,58,59] it is suggested to modify the equation
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
SemiStatic Hedging Strategies for Exotic Options
359
to an inequality, such that the trading strategy becomes a superreplication and then minimize the cost of this hedge.
14.3.3. ModelIndependent Strategies: Robust Strategies In the preceding section the problem of making the hedging strategies robust with respect to model risk was already mentioned. Now let us sharpen the question: what can be deduced on the price of the barrier option if the market is solely assumed to be frictionless? In particular, robust super and subreplicating strategies for barrier options are considered. Those strategies are called robust because they over (respectively under) hedge the barrier option in any frictionless (and in particular arbitragefree) market model. For barrier options, these were first developed by Brown et al. [12] (for some recent extensions see [29, 30]). In order to outline the approach, assume again a forward market model and a constant barrier on the forward price (the ideas can be extended to timedependent barriers, but the price intervals implied by the hedges might increase). Consider again the DOC with strike K, maturity T , and barrier B. If K ≥ B, then for any γ˜ > K a subreplicating portfolio for the DOC is given by: • A long position in a call with strike K and maturity T • A short position in (γ˜ − K)/(γ˜ − B) puts with strike B and maturity T • A short position in (K − B)/(γ˜ − B) calls with strike γ˜ ∨ K The corresponding trading strategy is to hold the portfolio until expiry if the barrier is not reached and otherwise to unwind the positions at the first hitting time. To show the subreplication property of this semistatic strategy, let us again distinguish the two cases: • If the barrier is not hit, the payoff of the strategy is (ST − K)+ −
K−B ˜ + ≤ (ST − K)+ , (ST − γ) γ˜ − B
where the righthand side corresponds to the payoff of the DOC; • If the barrier is hit at time τ, then Ft = B − x, where x ≥ 0 denotes the (possible) undershoot. Hence for the value Vτ of the portfolio we find Vτ = C(B − x, 0; K, T ) −
γ˜ − K K−B P(B − x, 0; B, T ) − C(B − x, 0; γ, ˜ T) γ˜ − B γ˜ − B
≤ C(B − x, 0; K, T ) −
K−B γ˜ − K C(B − x, 0; B, T ) − C(B − x, 0; γ, ˜ T) γ˜ − B γ˜ − B
≤ 0,
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
360
FA
Albrecher and Mayer
where the first inequality follows from the Put–Call–Parity (14.3) and the second one from the convexity (in K) of the call price (or alternatively by a direct look at the final payoff of the last portfolio). The portfolio is thus a subreplicating one for any γ˜ and hence γ˜ can be chosen such that the initial value V of the portfolio is maximized. Since V = C(F0 , 0; K, T ) −
γ˜ − K K−B P(F0 , 0; B, T ) − C(F0 , 0; γ, ˜ T) γ˜ − B γ˜ − B
= C(F0 , 0; K, T ) − P(F0 , 0; B, T ) +
K−B ˜ T )), (P(F0 , 0; K, T ) − C(F0 , 0; γ, γ˜ − B
the optimal γ˜ is given by γ˜ = argmaxβ>K
P(F0 , 0; B, T ) − C(F0 , 0; β, T ) . β−B
If K < B, then a subreplicating portfolio for the DOC is given by: • A long position in a forward contract (with forward price F0 ) • A long position in F0 − K bonds • A short position in (B − K)/(γ − B) puts with maturity T and strike γ > B The trading strategy is the same as before and the subreplication property is again shown by distinguishing two cases: • If the barrier is not hit, we have for the payoff of the strategy (ST − F0 ) + (F0 − K) −
B−K (γ − ST )+ ≤ (ST − K), γ −B
where the righthand side corresponds to the payoff of the DOC, since B > K; • If the barrier is hit at time τ, then Ft = B − x. Thus the timeτvalue of the long position in the forward is exp(−r(T − τ)) (B − x − F0 ) and hence for Vτ we find Vτ = e−r(T −τ) (B − x − F0 ) + e−r(T −τ) (F0 − K) −
B−K P(B − x, τ; γ, T ) γ −B
≤ e−r(T −τ) (B − K) −
B−K P(B, τ; γ, T ) γ −B
≤ e−r(T −τ) (B − K) −
B − K −r(T −τ) e (γ − B) = 0, γ −B
where the first inequality follows from the monotonicity of the put price and the second one from the Put–Call–Parity (14.3).
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
SemiStatic Hedging Strategies for Exotic Options
361
Similarly as in the case K > B it can be shown, that the optimal choice for γ, i.e., the one maximizing the initial value of the portfolio, is given by γ = argminβ>B
P(F0 , β, 0, T ) . β−B
Superreplicating portfolios are in both cases found by omitting the barrier feature in the final payoff, i.e., they are given as the trivial superhedges consisting of • A long position in a call with maturity T and strike K, if K ≥ B • A long position in a call with maturity T and strike B and in B − K binary calls (a pathindependent option with payoff 11{ST ≥B} ) with strike B, if K < B Of course the sub and superhedging strategies also imply price bounds on the DOC: If K ≥ B we have C(F0 , 0; K, T ) −
γ˜ − K K−B ˜ T) P(F0 , 0; B, T ) − C(F0 , 0; γ, γ˜ − B γ˜ − B
≤ DOC(F0 , 0; K, T, B) ≤ C(F0 , 0; K, T ), while in the case K < B we have B−K e−rT (F0 − K) − P(F0 , 0; γ, T ) ≤ DOC(F0 , 0; K, T, B) γ −B ≤ C(F0 , 0; B, T ) + (B − K)BC(F0 , 0; B, T ), ˜ γ where BC(Ft , t; B, T ) denotes the timetprice of a binary call with strike B, and γ, are as before. The particular role of these sub and superreplication strategies is identified in the following result: Theorem 14.3.1 ( [12]). Assume that European options for all strikes with maturity T are liquid and that there are no options available with maturity less than T. Then the bounds on the price of the DOC are sharp. This means that there are forward price processes that are martingales and consistent with the prices of the European options for which the price of the DOC is given by the lower (resp. upper) bound. In those market models, the hedging strategies are perfect. The proof of this theorem is closely related to the Skorokhod problem (see also [67]).
14.4. HEDGING STRONGLY PATHDEPENDENT OPTIONS In this section we consider strongly pathdependent options. In particular we focus on lookback and Asian options. For lookback options we show how barrier options (and
May 19, 2010
16:20
WSPC/SPIB913
362
b913ch14
FA
Albrecher and Mayer
the corresponding hedging strategies) can be used to hedge them. For Asian options robust hedging strategies are shown, which are related to hedging strategies for basket options. A perfect (modeldependent) strategy for discretely sampled options (DSO), which include Asian or cliquet options, is outlined in Sec. 14.5.
14.4.1. Lookback Options Let us consider a floating strike lookback option (LO) with terminal payoff p(ST ) = ST − mT , where mT = inf 0≤s≤T Ss . This kind of lookback option can be hedged robustly in terms of barrier options, as was first observed by Carr et al. [17] (see also [26, 27] for static hedging strategies of similar insurance products). The hedging strategies are not exact but rather suband superreplications and hence impose bounds on the price of the lookback option. Along the way to obtain the hedging strategies for the lookback option, we also deal with some other exotic options with gradually increasing complexity. A rolldown call (RDC) with maturity T and strike K0 has two barriers H1 > H2 . In contrast to a doublebarrier option, these two barriers are both below the initial asset price and the strike. The payoff of the RDC is specified as follows: if neither of the two barriers is hit, then the payoff is (ST − K0 )+ . If the nearer barrier (H1 ) is hit, the strike of the option is rolled down to this barrier and the second barrier becomes an outbarrier. So hitting H1 the RDC becomes a DOC with strike H1 and outbarrier H2 . For some more details concerning rolldowncalls and also rollupputs, see [45]. For the purpose of hedging the lookback option, the definition should be extended as follows. Let H1 > H2 > · · · > Hn be a decreasing sequence of barriers all below the current spot price and below the initial strike K0 . If no barrier is hit, then the payoff is again (ST − K0 )+ . If the first barrier is hit, the strike rolls down to some level K1 ∈ [H1 , K0 ]. If the second barrier is hit, the strike rolls down to a certain strike K2 ∈ [H2 , K1 ]. This rolling down process is repeated until the asset price reaches (or undershoots) Hn — the outbarrier. Thus hitting Hn knocks out the option. The option defined above is called extended rolldown call (ERDC) and admits the following decomposition in terms of DOC’s: ERDC(F0 , 0; K0 , T, {Ki }1≤i≤n−1 , {Hi }1≤i≤n ) = DOC(F0 , 0; K0 , T, H1 ) +
n−1
DOC(F0 , 0; Ki , T, Hi+1 ) − DOC(F0 , 0; Ki , T, Hi ) .
(14.15)
i=1
This portfolio indeed matches the payoff of the ERDC exactly, since if Hi+1 < mt ≤ Hi (recall mt = inf 0≤s≤t Ss ), the sum in (14.15) starts with i + 1 and the leading term
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
363
SemiStatic Hedging Strategies for Exotic Options
DOC(F0 , 0; K0 , T, H1 ) is replaced by DOC(Ft , t; Ki , T, Hi+1 ). Thus the ERDC can be perfectly replicated by a finite number of barrier options. Note that the representation above is modelindependent and the semistatic hedge is perfect in all frictionless models. Therefore, to build a semistatic hedging strategy for ERDC’s using European options, the corresponding hedging strategies for DOC’s can be used. The next option we want to consider is the ratchet call. The ratchet call (RC) is an ERDC with initial strike K0 and the strikes Ki equal the barrier levels Hi for 1 ≤ i ≤ n. Furthermore it cannot be knocked out. Thus the RC can be written as follows: RC(F0 , 0; K0 , T, {Ki }1≤i≤n ) = DIC(F0 , 0; Kn , T, Kn ) + ERDC(F0 , 0; K0 , T, {Ki }1≤i≤n−1 , {Ki }1≤i≤n ), where DIC stands for a downandin call. Using the modelindependent representation (14.15) of the ERDC, we find: RC(F0 , K0 , 0, T, {Ki }1≤i≤n ) = DOC(F0 , 0; K0 , T, K1 ) + DIC(F0 , 0; Kn , T, Kn ) +
n−1
DOC(F0 , 0; Ki , T, Ki+1 ) − DOC(F0 , 0; Ki , T, Ki ) .
(14.16)
i=1
Consider now the floating strike lookback call, defined at the beginning of the section. Recall that the payoff at maturity T is given by ST −mT , where mT is the minimum price of the asset up to T . This type of lookback calls can be seen as RC with a continuum of rolldown barriers and strikes. Due to this continuum of rolldown barriers we are not able to find an exact hedge, but we give super and subhedges for lookback calls. It is clear that an RC with K0 = F0 and Kn = 0 undervalues the LO, because the strike of the lookback option at time t can only be below or equal the strike of the RC. Therefore a modelindependent subhedge for a LO is given by: LO(F0 , 0; T ) ≥ RC(F0 , 0; F0 , T, {Ki }1≤i≤n ) = DOC(F0 , 0; F0 , T, K1 ) + DIC(F0 , 0; Kn , T, Kn ) +
n−1
DOC(F0 , 0; Ki , T, Ki+1 ) − DOC(F0 , 0; Ki , T, Ki ) .
i=1
(14.17) Again the barrier options might be hedged using the strategies of Sec. 14.3. Adding more rolldown barriers obviously increases the quality of the subhedge, i.e., a tighter bound is obtained.
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
364
FA
Albrecher and Mayer
In order to find a superhedge for the LO, an ERDC with Ki = Hi+1 for 1 ≤ i ≤ n and Hn+1 = 0 can be used. Such an ERDC is a superhedge, because its strike is below or equal to the strike of the LO throughout the lifetime of the option. Hence we have LO(F0 , 0; T ) ≤ ERDC(F0 , 0; H1 , T, {Hi+1 }1≤i≤n , {Hi }1≤i≤n+1 ) = DOC(F0 , 0; H1 , T, H1 ) +
n+1
DOC(F0 , 0; Hi , T, Hi ) − DOC(F0 , 0; Hi , T, Hi−1 ) .
(14.18)
i=2
Another portfolio having the same payoff as the one stated above is given by: LO(F0 , 0; T ) ≤ C(F0 , 0; H1 , T ) − P(F0 , 0; H1 , T ) +
n−1
(Hi − Hi+1 ) DIB(F0 , 0; Hi , T )
i=1
+ Hn DIB(F0 , 0; Hn , T ),
(14.19)
where DIB is a downandin digital option that pays 1 at expiry if the barrier was hit before maturity. As for the subhedge, the bounds become tighter the more Hi are added. Note that (14.17) as well as (14.18) and (14.19) are modelindependent superand subhedging portfolios, respectively. For fixedstrike lookback options one can use similar arguments and digital options to replicate the payoff (see [13]). Then again, the hedging strategies for weakly pathdependent options can be used. It is worth noting that in [48] the best modelindependent bounds for lookback options is calculated. However, the ideas are similar to the ones used for the derivation of the robust strategies for hedging barrier options and we omit the details here.
14.4.2. Asian Options Let K be the strike, {ti }1≤n the monitoring times, and T the maturity of an Asian call (AC). Then the payoff p of the AC at T is given by
p(ST ) =
1 St − K n i=1 i n
+
+
n 1 = St − nK . n i=1 i
(14.20)
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
365
SemiStatic Hedging Strategies for Exotic Options
Taking the operator (·)+ inside the sum in (14.20) clearly gives an upper bound and hence a static superreplication strategy for the AC in terms of European calls, i.e.,
n
+ Sti − nK
n + = (St1 − K1 ) + · · · + (Stn − Kn ) ≤ (Sti − Ki )+ , (14.21)
i=1
i=1
n
whenever i=1 Ki = nK. A simple first choice for the strikes of the calls is Ki = K ∀ 1 ≤ i ≤ n, which is, however, not optimal. The cheapest (an hence best) choice was originally found in [74] for complete markets and later generalized in [1, 3] (see also [25, 80]) and it turns out that the concept of comonotonicity is a helpful tool to this end: Definition 14.3 (Stoploss transform). Let FX (x) be the distribution function of a nonnegative random variable X. Then the stoploss transform FX (m) is defined by ∞ (x − m)dFX (x) = E[(X − m)+ ], m > 0. FX (m) = m
Definition 14.4 (Comonotone random vector). Let (X1 , X2 , . . . , Xn ) be a nonnegative random vector with marginal distribution functions FX1 , FX2 , . . . , FXn . The vector is called comonotone, if the joint distribution function FX1 ,X2 ,...,Xn is given by FX1 ,X2 ,...,Xn (x1 , x2 , . . . , xn ) = min{FX1 (x1 ), FX2 (x2 ), . . . , FXn (xn )}. Note that the righthand side of the above equation is a copula (often called lower Frechetcopula). For a general introduction to this field and proofs of the properties used below, see [40, 41]. To simplify notation, let us assume that the marginal distribution functions are strictly increasing. Let (X1 , . . . , Xn ) be a nonnegative random vector with marginal distribution functions FXi and suppose (Y1 , . . . , Yn ) to be the comonotone vector with the same marginal distributions. Setting S C = ni=1 Yi the following holds: FS−1 C (x) =
n
FX−1i (x),
0 ≤ x ≤ 1.
(14.22)
i=1
A crucial result is the following: FSC (m) =
n
FXi FX−1i (FS C (m)) ,
m ≥ 0.
(14.23)
i=1
Let (Y1 , . . . , Yn ) be the comonotone vector with the marginal distribution functions FSt1 , . . . , FStn and S C the sum of the Yi ’s. The following inequality holds for all
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
366
FA
Albrecher and Mayer
n i=1
Ki = nK:
FSC (nK) = E
n
+ ≤
Yi − nK
n n E (Yi − Ki )+ = Sti (Ki ).
i=1
i=1
i=1
Using Eq. (14.23) and the relation Sti (Ki ) = exp(rti )C(F0 , 0; Ki , T ) we finally find: n
exp(rti )C(F0 , 0; FS−1 (FS C (nK)), T ) ti
i=1
≤
n
exp(rti )C(F0 , 0; Ki , T ).
i=1
As a consequence, the optimal choice for the call strikes in (14.21) is: Ki = FS−1 (FS C (nK)). t i
Note that n
FS−1 (FS C (nK)) = FS−1 C (FS C (nK)) = nK, t i
i=1
for all strictly increasing marginal cumulative distribution functions. Since the inverse function of FS C is given by (14.22) and the righthand side is strictly increasing in x, it is computationally straightforward to calculate FS C (nK). Thus this hedging strategy is also simple to evaluate. Numerical studies carried out in [1] for popular Lévy models and in [3] for stochastic volatility models suggest that the obtained bounds become tighter the deeper the options are in the money. Of course the concept of comonotonicity can also be applied to other types of options, for which the payoff depends on a sum of possibly dependent random variables, e.g., basket options. For more details we refer to [25, 49, 50]. In those papers, in particular, the (realistic) case is considered, when there are only finitely many strikes liquid in the market and one has to find the optimal combination of those. This last question and especially the problem of finding a subreplicating portfolio (which is associated to a lower price bound for theAsian option) is also addressed in [2]. Lower price bounds on the AC price were first considered by Curran [31] and Rogers and Shi [70], who pioneered a quite accurate method to determine lower price bounds in the BlackScholes model based on the following idea: Jensen’s inequality gives + n +
n 1 1 St − K ≥ E Sti Z − K , (14.24) n i=1 i n i=1 where Z is an arbitrary random variable. In the BlackScholes model the choice
n n1 Z= Si i=1
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
SemiStatic Hedging Strategies for Exotic Options
367
or variants thereof are very popular (because the distribution of the geometric mean is explicitly available in the BlackScholes model) and leads to tight lower price bounds, since the arithmetic and geometric average are strongly correlated (see [31, 66, 70, 77, 80]). However, in contrast to upper price bounds based on the concept of comonotonicity, this method in general does not imply a subreplication strategy. Nevertheless, in [2] it is shown that using Z = St1 together with (14.22) yields a robust subreplicating portfolio consisting of ni=1 e−r(T −ti ) /n calls with maturity t1 n and strike nK/( j=1 er(tj −t1 ) ). The corresponding trading strategy is to do nothing when S1 ≤ nK/ ni=1 er(ti −t1 ) or to buy ni=1 e−r(T −ti ) /n assets in the case that n r(ti −t1 ) . The cost for this trade is exactly the payoff of the options S1 > nK/ i=1 e in the portfolio plus Ke−r(T −t1 ) , which one should borrow. Then at each monitoring time ti one should sell e−r(T −ti ) /n assets and invest the gain in the riskless bank account. At maturity T of the Asian call, the payoff of the trading strategy is
n 1 St − K 11{ ni=1 er(ti −t1 ) St1 >nK} , n i=1 i which is clearly dominated by the payoff of the AC.
14.5. CASE STUDY: MODELDEPENDENT HEDGING OF DISCRETELY SAMPLED OPTIONS In Sec. 14.2 a hedging portfolio for pathindependent options was presented. As the payoff of discretely monitored options (DSO) depends only on a finite number of monitoring times, such options can be understood as pathindependent between the monitoring times and it is possible to construct a semistatic hedging strategy with adaptations of the portfolio only at the monitoring times. This kind of strategies was developed by Carr and Wu [24] and Joshi [55] (see also [54]). Similar to Sec. 14.2, we assume that standard European options are liquid for all strikes and monitoring times of the DSO. However, to apply the techniques for the pathindependent options we need the extra assumption that the asset price process is Markovian. In particular, the prices of standard European options may only depend on the current option price. We consider a recursion algorithm that can be used to price such options and to obtain some Greeks (namely the Delta and the Gamma) of the exotic option through the corresponding ones of plain vanilla options. This algorithm is based on the assumption that the price of the exotic option depends, in addition to the current price of the asset, only on some summary statistic of the historic asset prices at the monitoring times which is measurable with respect to the filtration generated by the asset price and updated only at the monitoring times of the option. If more than one statistic is needed
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
368
FA
Albrecher and Mayer
to describe the price of the exotic option, the method is still feasible, but for notational convenience we focus on the case with a single one. As an illustrating example, an AC with strike K, monitoring times t1 , . . . , tn , and maturity T = tn is considered, but any other similar DSOs (like e.g., cliquets, or discretely monitored barrier and lookback options) could serve as well. Let us denote the summary statistic of a generic DSO at time ti by Xi . Due to the assumptions on X, Xi is Fti measureable and Xt = Xi for ti ≤ t ≤ ti+1 , i = 0, 1, . . . , n − 1. Furthermore we assume that Xi+1 = f(Xi , Sti+1 ),
(14.25)
which is fulfilled for all DSOs we are aware of. For the AC, Xi is the running average Ai of the asset prices at time ti , i.e., 1 St , i j=1 j i
Xi := Ai =
∀ 1 ≤ i ≤ n.
Obviously here assumption (14.25) is fulfilled, since Ai+1 =
i 1 Ai + St , i+1 i + 1 i+1
∀ 1 ≤ i ≤ n − 1.
Due to (14.25) we have for the payoff of the generic DSO p(ST ) = g(Xn−1 , Stn ) for some terminal payoff function g. Now, since Xn−1 is known at time tn−1 , we can use (14.1) to find ∞ p(ST ) = g(Xn−1 , K∗ ) + gS (Xn−1 , K∗ )(S − K∗ ) + gSS (Xn−1 , x)(S − x)+ dx +
K∗
K∗
gSS (Xn−1 , x)(x − S )+ dx,
(14.26)
0
where gS and gSS denote the first and second derivative of g with respect to its second argument. Note that, as in Sec. 14.2, the above describes a static hedge for the DSO at the time tn−1 and therefore also its timetn−1 value Vtn−1 is settled. Vtn−1 (Xtn−1 , Stn−1 ) = g(Xn−1 , K ∗ ) + gS (Xn−1 , K ∗ )(Ftn−1 − K∗ ) +
∞
K∗
+
gSS (Xn−1 , x)C(Stn−1 , tn−1 ; x, T )dx
K∗
gSS (Xn−1 , x)P(Stn−1 , tn−1 ; x, T )dx,
0
where Ftn−1 is the timetn−1 forward price.
(14.27)
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
FA
SemiStatic Hedging Strategies for Exotic Options
369
For the AC we have p(ST ) =
n−1 1 An−1 + ST − K n n
+
=
1 (ST − (nK − (n − 1)An−1 ))+ n
and Vtn−1 (Xtn−1 , Stn−1 ) =
1 C(Stn−1 , tn−1 ; nK − (n − 1)An−1 , T ), n
where C(Ft , t; K, T ) := e−r(T −t) (Ft − K) for K < 0 with Ft denoting again the forward price. The Markov property together with (14.27) then implies that the value Vtn−1 of the DSO at time tn−1 only depends on Xtn−1 and Stn−1 and using again (14.25) we actually have (with a slight abuse of notation) Vtn−1 (Xtn−1 , Stn−1 ) = Vtn−1 (Xtn−2 , Stn−1 ). Thus we can use the righthand side of (14.27) as the new payoff function of a (modified) DSO and iterate the procedure until we reach the current time t0 and obtain a replicating portfolio. The associated hedging strategy is to hold this portfolio until expiry t1 of the standard European options involved and invest the payoff of these options to form the new replicating portfolio for the Europeantype option Vt2 (X1 , St2 ). This strategy is of course selffinancing. Note that the replicating portfolio does not change until t1 , and thus one can calculate the price of the DSO, as well as the derivatives with respect to the asset price, with this portfolio. Unfortunately the derivatives with respect to other parameters of the DSO price in general cannot be calculated in the same manner, since changing those parameters would also have an effect on the portfolio. It is worth noting that the weights of the standard options in the hedging portfolio at the current time t0 are determined by the Gamma of the DSO at time t1 . The derivation of this hedging strategy of course relies on the Markov assumption on the model and the liquidity of arbitrary standard European options. However, even if those assumptions may fail in practice, the approach gives an idea of how to obtain approximative hedging strategies. More precisely: if neither the Markov assumption nor the liquidity assumption is fulfilled, one might not be able to hedge the DSO perfectly. However, if alternatively a hedging strategy is chosen in order to minimize a certain risk measure for the hedging error in (14.26), the semistatic strategy might still outperform classical dynamic hedging strategies. Recently, this kind of approach was applied e.g., in [7].
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
370
FA
Albrecher and Mayer
14.6. CONCLUSION AND FUTURE RESEARCH Semistatic hedging strategies can be a valuable alternative to classic dynamic replication approaches in various situations, often leading to a better performance and reduced model risk. Although over the last years various results have been obtained, there are many open research questions. From a theoretical perspective, it might be interesting to extend the discussed sharp price bounds to other classes of exotic options, like cliquet options. Also, improved price bounds in terms of other liquid options beyond standard European options are needed. It could be rewarding to investigate which price bounds can be obtained if American put options, variance/volatility swaps, or credit default swaps are liquid. A recent paper in this direction is [23]. Another line of extension is to model the vanilla option prices themselves. Recently, in [73] dynamics for option prices were found that are consistent with the noarbitrage assumption and one might be able to use this kind of modelling to design more advanced trading strategies with options. Among further future research topics is the quantitative and systematic comparison of hedging performance of static and dynamic strategies, including historical backtesting of the strategies.
References [1] Albrecher, H, J Dhaene, M Goovaerts and W Schoutens (2005). Static hedging of Asian options under Lévy models. Journal of Derivatives, 12, 63–72. [2] Albrecher, H, P Mayer and W Schoutens (2008). General lower bounds for arithmetic Asian option prices. Applied Mathematical Finance, 15, 123–149. [3] Albrecher, H and W Schoutens (2005). Static hedging of Asian options under stochastic volatility models using fast Fourier transform. In Exotic Options and Advanced Lévy Models, A Kyprianou et al. (eds.), pp. 129–148. Chichester: Wiley. [4] Allen, S and O Padovani (2002). Risk management using quasistatic hedging. Economic Notes, 31, 277–336. [5] Andersen, LBG, J Andreasen and D Eliezer (2002). Static replication of barrier options: Some general results. Journal of Computational Finance, 5, 1–25. [6] Andreasen, J and P Carr (2002). Put call reversal. Working Paper, New York University. [7] Becherer, D and I Ward (2008). Optimal static hedging of equity and credit risk using derivatives. Working Paper, Humboldt University, Berlin. [8] BenHameur, H, M Breton and P François (2006). A dynamic programming approach to price installment options. European Journal of Opertational Research, 169, 667–676. [9] Bertsimas, D and I Popescu (2002). On the relation between option and stock prices: A convex optimization approach. Operations Research, 50(2), 358–374. [10] Black, F and M Scholes (1973). The pricing of options and corporate liabilities. Journal of Political Economy, 81, 637–654. [11] Breeden, DT and RH Litzenberger (1978). Prices of statecontingent claims implicit in option prices. Journal of Business, 51(4), 621–651.
May 19, 2010
16:20
WSPC/SPIB913
SemiStatic Hedging Strategies for Exotic Options
b913ch14
FA
371
[12] Brown, H, D Hobson and LCG Rogers (2001). Robust hedging of barrier options. Mathematical Finance, 11, 285–314. [13] Buchen, PW and O Konstandatos (2004). A new method of pricing lookback options. Mathematical Finance, 15, 245–259. [14] Carr, P and J Bowie (1994). Static simplicity. Risk, 7(8), 45–50. [15] Carr, P and A Chou (1997). Breaking barriers. Risk, 10, 139–145. [16] Carr, P and A Chou (2002). Hedging complex barrier options. Working Paper, New York University. [17] Carr, P, K Ellis and V Gupta (1998). Static hedging of exotic options. The Journal of Finance, 53, 1165–1190. [18] Carr, P and R Lee (2007). Realized volatility and variance: Options via swaps. Risk, May, 76–83. [19] Carr, P and R Lee (2010). Hedging variance options on continuous semimartingales. Finance and Stochastics (forthcoming). [20] Carr, P and R Lee (2008). Putcall symmetry: Extensions and applications. Mathematical Fincane 19(4), 523–560. [21] Carr, P and D Madan (2005). A note on sufficient conditions for no arbitrage. Finance Research Letters, 2, 125–130. [22] Carr, P and D Madan (2006). Towards a theory of volatility trading. In Risk Book on volatility, R Jarrow (ed.). Berlin: Risk. [23] Carr, P and W Schoutens (2008). Hedging under the Heston model with jumptodefault. International Journal of Theoretical and Applied Finance, 11, 403–414. [24] Carr, P and L Wu (2002). Static hedging of standard options. Working Paper, Courant Institute, New York University. [25] Chen, X, G Deelstra, J Dhaene and M Vanmaele (2008). Static superreplicating strategies for a class of exotic options. Insurance: Mathematics and Economics, 42(3), 1067–1085. [26] Coleman, TF, Y Kim, Y Li and M Patron (2007). Robustly hedging variable annuities with guarantees under jump and volatility risk. Journal of Risk and Insurance, 7, 347–376. [27] Coleman, TF, Y Li and M Patron (2006). Hedging guarantees in variable annuities (under both market and interest rate risk). Insurance: Mathematics and Economics, 38, 215–228. [28] Cousot, L (1969). Conditions on option prices for absence of arbitrage and exact calibration. Journal of Banking & Finance, 31, 3377–3397. [29] Cox, AMG and J Oblój (2008). Robust hedging of double touch barrier options. Working Paper, available at http://arxiv.org/abs/0808.4012. [30] Cox, AMG and J Oblój (2009). Robust pricing and hedging of double notouch options. Working Paper, available at http://arxiv.org/abs/0901.0674. [31] Curran, M (1994). Valuing Asian and portfolio options by conditioning on the geometric mean price. Management Science, 40(12), 1705–1711. [32] Davis, M and D Hobson (2007). The range of traded option prices. Mathematical Finance, 17, 1–14. [33] Davis, M, W Schachermayer and R Tompkins (2001). Pricing, noarbitrage bounds and robust hedging of installment options. Quantitative Finance, 1, 597–610. [34] Davis, M, W Schachermayer and R Tompkins (2002). Installment options and static hedging. Journal of Risk Finance, 3(2), 46–52. [35] Delbaen, F and W Schachermayer (2006). The Mathematics of Arbitrage. Berlin Heidelberg: Springer. [36] Derman, E, D Ergener and I Kani (1994). Forever hedged. Risk, 7, 139–145.
May 19, 2010
372
16:20
WSPC/SPIB913
b913ch14
FA
Albrecher and Mayer
[37] Derman, E, D Ergener and I Kani (1995). Static options replication. The Journal of Derivatives, 2, 78–95. [38] Derman, E, I Kani and M Kamal (1997). Trading and hedging local volatility. Journal of Fincancial Engineering, 6, 233–268. [39] Derman, E and NN Taleb (2005). The illusion of dynamic replication. Quantitative Finance, 5, 323–326. [40] Dhaene, J, M Denuit, MJ Goovaerts, R Kaas and D Vyncke (2002). The concept of comonotonicity in actuarial science and finance: Applications. Insurance: Mathematics & Economics, 31(2), 133–161. [41] Dhaene, J, M Denuit, MJ Goovaerts, R Kaas and D Vyncke (2002). The concept of comonotonicity in actuarial science and finance: Theory. Insurance: Mathematics & Economics, 31(1), 3–33. [42] Dupire, B (1997). A unified theory of volatility. Working Paper, BNP Paribas. [43] Engelmann, B, M Fengler and R Poulsen (2007). Static versus dynamic hedges: An empirical comparison for barrier options. Review of Derivatives Research, 9, 239–264. [44] Fink, J (2003). An examination of the effectiveness of static hedging in the presence of stochastic volatility. The Journal of Future Markets, 23(9), 859–890. [45] Gastineau, G (1994). Rollup puts, rolldown calls, and contingent premium options. The Journal of Derivatives, 1, 40–43. [46] Giese, AM and J Maruhn (2007). Costoptimal static superreplication of barrier options — an optimization approach. Journal of Computational Finance, 10, 71–97. [47] Green, RC and RA Jarrow (1987). Spanning and completeness in markets with contingent claims. Journal of Economic Theory, 41, 202–210. [48] Hobson, D (1998). Robust hedging of the lookback options. Finance and Stochastics, 2, 329–347. [49] Hobson, D, P Laurence and TH Wang (2005). Staticarbitrage optimal subreplicating strategies for basket options. Insurance: Mathematics & Economics, 37(3), 553–572. [50] Hobson, D, P Laurence and TH Wang (2005). Staticarbitrage upper bounds for the prices of basket options. Quantative Finance, 5(4), 329–342. [51] Hodges, SD and A Neuberger (2001). Rational bounds on exotic options. Working Paper, Warwick University. [52] Ilhan, A, M Jonsson and R Sircar (2008). Optimal staticdynamic hedges for exotic options under convex risk measures. Working Paper, available at http://ssrn.com/abstract= 1121233. [53] Ilhan, A and R Sircar (2005). Optimal staticdynamic hedges for barrier options. Mathematical Finance, 16, 359–385. [54] Joshi, MS (2002). The Concepts and Practice of Mathematical Finance. Cambridge: Cambridge University Press. [55] Joshi, MS (2002). Pricing discretely sampled pathdependend options using replication methods. Working Paper, available at www.quarchome.org. [56] Kraft, H (2007). Pitfalls in static superhedging of barrier options. Finance Research Letters, 4, 2–9. [57] Lindset, S and SA Perrson (2006). A note on a barrier exchange option: The world’s simplest option formula? Finance Research Letters, 3, 207–211. [58] Maruhn, J and EW Sachs (2006). Robust static superreplication of barrier options in the BlackScholes model. In Robust OptimizationDirected Design (RODD), AJ Kurdila, PM Pardalos and M Zabarankin (eds.), pp. 127–143. Springer.
May 19, 2010
16:20
WSPC/SPIB913
SemiStatic Hedging Strategies for Exotic Options
b913ch14
FA
373
[59] Maruhn, J and EW Sachs (2009). Robust static hedging of barrier options in stochastic volatility models. Mathematical Methods of Operations Research, 70(3), 405–433. [60] Merton, RC (1973). Theory of rational option pricing. Bell Journal of Economics and Management Science, 4, 141–183. [61] Mijatovi´c, A (2008). Local time and the pricing of timedependent barrier options. Working Paper, available at: http://arxiv.org/pdf/0809.1747. [62] Nachmann, D (1998). Spanning and completeness with options. Review of Financial Studies, 31, 311–328. [63] Nalholm, M and R Poulsen (2006). Static hedging and model risk for barrier options. Journal of Futures Markets, 26, 449–463. [64] Nalholm, M and R Poulsen (2006). Static hedging of barrier options under general asset dynamics: Unification and application. Review of Derivative Research, 13, 46–60. [65] Neuberger, A (2007). Bounds on the American option. Working Paper, Warwick Business School. [66] Nielsen, JA and K Sandmann (2003). Pricing bounds onAsian options. Journal of Financial and Quantitative Analysis, 38(2), 449–473. [67] Oblój, J (2004). The Skorokhod embedding problem and its offspring. Probability Surveys, 1, 321–392. [68] Peskir, G (2005). A changeofvariable formula with local time on curves. Journal of Theoretical Probability, 18, 499–535. [69] Poulsen, R (2006). Barrier options and their static hedges: Simple derivations and extensions. Quantitative Finance, 6, 327–335. [70] Rogers, LCG and Z Shi (1995). The value of an Asian option. Journal of Applied Probability, 32, 1077–1088. [71] Ross, S (1976). Options and efficiency. Quarterly of Economics, 90, 75–89. [72] Sbuelz, A (2005). Hedging double barriers with singles. International Journal of Theoretical and Applied Finance, 8, 393–407. [73] Schweizer, M and J Wissel (2008). Term structures of implied volatilities: Absence of arbitrage and existence results. Mathematical Finance, 18, 77–114. [74] Simon, S, MJ Goovaerts and D Dhaene (2000). An easy computable upper bound for the price of an arithmetic Asian option. Insurance: Mathematics & Economics, 26, 175–184. [75] Siven, J and R Poulsen (2009). Autostatic for the people: Riskminimizing hedges of berrier options. Working Paper, University of Gothenburg. [76] Stoll, HR (1969). The relationship between put and call prices. Journal of Finance, 24, 801–824. [77] Thompson, GWP (1999). Fast narrow bounds on the price of Asian options. Working Paper, University of Cambridge. [78] Toft, KB and C Xuan (1998). How well can barrier options be hedged by a static portfolio of standard options? Journal of Financial Engineering, 7, 147–175. [79] Tompkins, R (1997). Static versus dynamic hedging: An evaluation of hedge performance via simulation. Netexposure, 1(2), 1–28. [80] Vanmaele, M, G Deelstra, J Liinev, J Dhaene and MJ Goovaerts (2006). Bounds for the price of discrete arithmetic Asian options. Journal of Computational and Applied Mathematics, 185, 51–90.
May 19, 2010
16:20
WSPC/SPIB913
b913ch14
This page intentionally left blank
FA
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
FA
15
DISCRETETIME VARIANCEOPTIMAL HEDGING IN AFFINE STOCHASTIC VOLATILITY MODELS
JAN KALLSEN∗,§ , JOHANNES MUHLEKARBE†,¶ , NATALIA SHENKMAN‡, and RICHARD VIERTHAUER∗,∗∗ ∗
Mathematisches Seminar, ChristianAlbrechtsUniversität zu Kiel, ChristianAlbrechtsPlatz 4, 24098 Kiel, Germany † Fakultät für Mathematik, Universität Wien, Austria Nordbergstr. 15, 1090 Wien, Austria ‡ Lehrstuhl für Energiehandel und Finanzdienstleistungen Universität DuisburgEssen, Universitätsstraße 12, 45141 Essen § kallsen@math.unikiel.de ¶ johannes.muhlekarbe@univie.ac.at natalia.shenkman@unidue.de ∗∗ vierthauer@math.unikiel.de
We consider varianceoptimal hedging when trading is restricted to a finite time set. Using Laplace transform methods, we derive semiexplicit formulas for the varianceoptimal initial capital and hedging strategy in affine stochastic volatility models. For the corresponding minimal expectedsquared hedging error, we propose a closedform approximation as well as a simulation approach. The results are illustrated by computing the relevant quantities in a timechanged Lévy model.
375
May 12, 2010
17:47
WSPC/SPIB913
376
b913ch15
FA
Kallsen et al.
15.1. INTRODUCTION A classical question in Mathematical Finance is how the issuer of an option can hedge her risk by trading in the underlying. To tackle this problem in incomplete markets, we consider varianceoptimal hedging, cf. [5, 18, 22] and the references therein for a survey of the extensive literature. Varianceoptimal hedging of a contingent claim H means that one minimizes the expected squared hedging error E[(v0 + ϕ • ST − H)2 ] over all initial endowments v0 and trading strategies ϕ, where ϕ • ST represents the cumulated gains resp. losses from trading ϕ up to the expiry date T of the claim. In this chapter, we consider the above problem in affine stochastic volatility models. These generalize Lévy processes by allowing for volatility clustering and are capable of recapturing most of the stylized facts observed in stock price time series. For Lévy processes, varianceoptimal hedging has been dealt with using PDE methods by [6] and by employing Laplace transform techniques in [4,10]. The approach of [10] has subsequently been extended to affine models by [12,13,17] if the discounted asset price is a martingale and by [14] in the general case. However, whereas [4, 10] incorporate both continuous and discrete rebalancing, the results for affine processes have focused on continuous trading thus far. The present study complements these results by showing how to deal with discretetime varianceoptimal hedging in affine models. Since only finitely many trades are feasible in reality, this analysis is important in order to answer the following questions: (1) How should discrete rebalancing affect the investment decisions of the investor, i.e., to what extent should she adjust her hedging strategy? (2) How can one quantify the additional risk resulting from discrete trading, i.e., by how much does the hedging error increase? The general structure of varianceoptimal hedging in discrete time has been thoroughly investigated by [21]. However, examples of (semi) explicit solutions seem to be limited to the results of [4, 10] for Lévy processes and [2] for some specific diffusion models with stochastic volatility. Here, we show how to extend the Laplace transform approach of [10] to general affine stochastic volatility models. Similarly as in [12, 13], we focus on the case where the discounted asset price process is a martingale. Numerical experiments using the results of [10] and [14] indicate that the effect of a moderate drift rate on hedging problems is rather small. This article is organized as follows. In Sec. 15.2, we summarize for the convenience of the reader the general structural results of [21] on varianceoptimal hedging in discrete time, reduced to the case where the underlying asset is a martingale. Subsequently, we explain how the Laplace transform approach can be used in general discretetime
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
FA
377
DiscreteTime VarianceOptimal Hedging
models in order to obtain integral representations of the objects of interest. Section 15.4 turns to the computation of the integrands from Sec. 15.3 in affine stochastic volatility models. We show how to compute all integrands in closed form for the optimal initial capital and hedging strategy. This parallels results for continuoustime hedging in [4, 10] and [12, 13] and for discretetime hedging in [4, 10]. Somewhat surprisingly, the expressions for the corresponding hedging error turn out to be considerably more involved than in the continuoustime case and cannot be computed in closed form. We propose two approaches to circumvent this problem: First, we determine a closed form approximation, whose error becomes negligible as the number of trades tends to infinity. As an alternative, we put forward a simple Monte–Carlo scheme to approximate the hedging error via simulation. Section 15.5 contains some numerical examples for the timechanged Lévy models introduced by [3].
15.2. DISCRETETIME VARIANCEOPTIMAL HEDGING Let T > 0 be a fixed time horizon, N ∈ N, T0 := {t0 , t1 , . . . , tN }, and T := T0 \{0}, where tn = nT/N for n = 0, . . . , N. Denote by (, F, (Ft )t∈T0 , P) a filtered probability space with discrete time set T0 . For simplicity, we assume that the initial σfield F0 is trivial. As for an introduction to financial mathematics in this discrete setup, the reader is referred to the textbook of Lamberton and Lapeyre [15]. The logarithm X of the discounted stock price process S = S0 exp(Xt ),
S0 ∈ R+ ,
t ∈ T0 ,
is supposed to be the second component of an adapted process (y, X), where X0 is normalized to zero. The first component y models stochastic volatility or, more accurately, stochastic activity in the model. Throughout, we suppose that E[ST2 ] < ∞, as well as E[St2n Ftn−1 ] > 0,
t∈T,
(15.1)
to rule out degenerate cases. Our goal is to compute the varianceoptimal hedge for a given contingent claim H in the following sense. Definition 15.1. We say that (v0 , ϕ) is an admissible endowment/strategy pair, if v0 ∈ R and ϕ = (ϕt )t∈T is a predictable process (i.e., ϕt is Ft−1 measurable) such that ϕt St ∈ L2 (P). ϕ • ST := t∈T
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
378
FA
Kallsen et al.
An admissible endowment/strategy pair (v0 , ϕ ) is called varianceoptimal for a contingent claim with discounted payoff H ∈ L2 (P) at time T , if it minimizes the expected squared hedging error (v0 , ϕ) → E[(v0 + ϕ • ST − H)2 ] over all admissible endowment/strategy pairs (v0 , ϕ). In this case, we refer to v0 as the varianceoptimal initial capital and call ϕ varianceoptimal hedging strategy. As noted in the introduction, we restrict ourselves to the case where the stock price is a martingale. Assumption 15.2. The stock price process S is a squareintegrable martingale. In this case, the varianceoptimal capital and strategy can be represented as follows. Proposition 15.3. Let H ∈ L2 (P). Then the varianceoptimal endowment/strategy pair for H is given by v0 = V0 ,
ϕtn =
E[Vtn Stn Ftn−1 ] − Vtn−1 Stn−1 , E[St2n Ftn−1 ] − St2n−1
tn ∈ T ,
where Vtn := E[HFtn ],
tn ∈ T0
denotes the option price process of H. The corresponding minimal expected squared hedging error is given by J0 := E[VT2 − V02 ] − E[ϕtn (E[Vtn Stn Ftn−1 ] − Vtn−1 Stn−1 )]. tn ∈T
Proof. This follows from [21, Sec. 4.1] by making use of the martingale properties of S and V . Notice that if the initial capital is fixed at v0 ∈ R rather than being part of the optimization problem, the same strategy ϕ is still optimal if S is a martingale (cf. [21, Sec. 4.1]). However, the corresponding hedging error increases by (v0 − V0 )2 in this case.
15.3. THE LAPLACE TRANSFORM APPROACH In order to derive formulas that can be computed in concrete models, we use the Laplace transform approach, which has been introduced to varianceoptimal hedging by [10]. The key assumption on the contingent claim is the existence of an integral representation in the following sense.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
FA
379
DiscreteTime VarianceOptimal Hedging
Assumption 15.4. Suppose that the payoff function of the claim is of the form H = f(ST ) for some function f : (0, ∞) → R, such that f(s) =
R+i∞
sz l(z)dz, R−i∞
for l : C → C and R ∈ R such that x → l(R + ix) is integrable and E[exp (2RXT )] < ∞.
Example 15.5. Most European options admit an integral representation of this kind. For example, for the European call with payoff function f(s) = (s − K)+ , we have, f(s) =
1 2πi
R+i∞
sz R−i∞
K1−z dz, z(z − 1)
for any R > 1 by [10, Lemma 4.1]. More generally, the Bromwich inversion formula as in [10, Theorem A.1] ascertains that l is typically given by the bilateral Laplace transform of x → f(exp(x)), cf. [10] for more details and examples. Henceforth, we only consider contingent claims satisfying Assumption 15.4. In this case, Proposition 15.3 can also be written in integral form. Theorem 15.6. We have H ∈ L2 (P) and the corresponding option price process is given by Vtn =
R+i∞
V(z)tn l(z)dz,
tn ∈ T 0 ,
R−i∞
for the squareintegrable martingales V(z)tn := E[STz Ftn ],
tn ∈ T0 .
Moreover, the varianceoptimal hedging strategy for H can be represented as ϕtn
=
R+i∞
R−i∞
E[V(z)tn Stn Ftn−1 ] − V(z)tn−1 Stn−1 l(z)dz, E[St2n Ftn−1 ] − St2n−1
tn ∈ T .
Proof. The first assertion follows from Assumption 15.4 and Fubini’s Theorem along the lines of [13, Lemma 3.3 and Proposition 3.4]. The second can be derived analogously using the Cauchy–Schwarz Inequality and E[St2n ] < ∞, tn ∈ T . For the hedging error, Proposition 15.3 and Theorem 15.6 yield the following similar integral representation.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
380
FA
Kallsen et al.
Corollary 15.7. For tn ∈ T and z1 , z2 ∈ R + iR, let J1 (z1 , z2 ) := V(z1 )T V(z2 )T − V(z1 )0 V(z2 )0 , J2 (tn , z1 , z2 ) :=
E[V(z1 )tn Stn Ftn−1 ] − V(z1 )tn−1 Stn−1 E[St2n Ftn−1 ] − St2n−1 × (E[V(z2 )tn Stn Ftn−1 ] − V(z2 )tn−1 Stn−1 ).
If, for tn ∈ T , ∞ ∞ E[J2 (tn , R + ix1 , R + ix2 )]l(R + ix1 )l(R + ix2 )dx1 dx2 < ∞, −∞
(15.2)
−∞
the minimal expected squared hedging error is given by R+i∞ R+i∞ E[J1 (z1 , z2 )]l(z1 )l(z2 )dz1 dz2 J0 = R−i∞
−
R−i∞
tn ∈T
R+i∞
R−i∞
R+i∞
E[J2 (tn , z1 , z2 )]l(z1 )l(z2 )dz1 dz2 .
R−i∞
15.4. APPLICATION TO AFFINE STOCHASTIC VOLATILITY MODELS Theorem 15.6 and Corollary 15.7 show that in order to compute semiexplicit formulas of the discrete varianceoptimal capital, hedging strategy and hedging error, one must be able to compute conditional exponential moments of the process X. This suggests to consider models whose moment generating function E[exp(uXtn )] is known in closed form. Here we use affine processes in the sense of [7]. Assumption 15.8. Suppose that (yt , Xt )t∈T0 is the restriction to discrete time of a semimartingale which is regularly affine w.r.t. y in the sense of [7, Definitions 2.1 and 2.5]. This means that the characteristic function of (y, X) has exponentially affine dependence on y, i.e., there exist mappings j : T × iR2 → C, j = 0, 1 such that, for t ≥ s and (u1 , u2 ) ∈ iR2 , E[eu1 yt +u2 Xt Fs ] = exp(0 (t − s, u1 , u2 ) + 1 (t − s, u1 , u2 )ys + u2 Xs ).
(15.3)
Example 15.9. By [7, Theorems 2.7 and 2.12], a continuoustime semimartingale (y, X) is affine if and only if its local dynamics expressed in terms of the infinitesimal generator resp. the differential characteristics depend on y in an affine way. Moreover, the functions 0 , 1 can be determined by solving some generalized Riccati equations.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
FA
381
DiscreteTime VarianceOptimal Hedging
In [11], it is shown that a large number of stochastic volatility models from the empirical literature fit into this framework. Examples include the models of Heston [9] and BarndorffNielsen and Shephard [1] as well as their extensions to timechanged Lévy models by [3]. A particular specification of this general class of models is given by the following OUtimechange model: Xt = L t ys ds , 0
dyt = −λyt dt + dZt ,
y0 > 0,
for a mean reversion speed λ > 0, a Lévy process L with Lévy exponent ψL , and an increasing Lévy process Z with Lévy exponent ψZ , i.e., E[euLt ] = exp(tψL (u)),
E[euZt ] = exp(tψZ (u)),
∀u ∈ iR.
In this case, (y, X) is affine by [11, Sec. 4.4] and in view of [11, Corollary 3.5], we have 1 (t, u1 , u2 ) = e−λt u1 +
1 − e−λt L ψ (u2 ), λ
t
0 (t, u1 , u2 ) =
ψ Z (1 (s, u1 , u2 ))ds. 0
If y is chosen to be a GammaOU process with stationary Gamma(a, b) distribution (see [20] for more details), we have ψZ (u) = (λau)/(b − u) and 0 can be determined in closed form as well. By [12, Proposition 3.6], we have 0 (t, u1 , u2 ) b − 1 (t, u1 , u2 ) aλ L + tψ (u ) if bλ = ψL (u2 ), b log 2 b − u1 bλ − ψ L (u2 ) = b λt −aλ − 1) + t if bλ = ψL (u2 ), (e λu1 − ψ L (u2 ) where log denotes the distinguished logarithm in the sense of [19, Lemma 7.6], i.e., the branch is chosen such that the resulting function is continuous in t. To compute exponential moments of X such as V(z)t = E[STz Ft ], z ∈ R + iR, we need Eq. (15.3) to remain valid on a suitable extension of iR2 . The following sufficient condition is taken from [13]. Assumption 15.10. Suppose that for all tn ∈ T0 , the mappings (u1 , u2 ) → j (tn , u1 , u2 ), j = 0, 1 admit analytic continuations to the strip S := {z ∈ C2 : Re(z) ∈ (−∞, (M ∨ 0) + ) × ((2R ∧ 0) − , (2R ∨ 2) + )}, for some > 0 and M := sup{21 (T − tn , 0, r) : r ∈ [R ∧ 0, R ∨ 0], tn ∈ T0 }.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
382
FA
Kallsen et al.
The existence of the analytic extensions in Assumption 15.10 is difficult to verify in general. For affine diffusion processes, [8, Theorem 3.3] shows that it suffices to establish that solutions to the corresponding Riccati equations exist on [0, T ]. In the presence of jumps, the situation is more involved and one has to work on a casebycase basis. For timechanged Lévy processes, this has been carried out in detail by [12]. Example 15.11. By the proof of [12, Theorems 3.3, 3.4], Assumption 15.10 holds in the OUtimechange models from Example 15.9, if the Lévy exponents ψ L and ψZ admit analytic extensions to {z ∈ C : Re(z) ∈ ((2R ∧ 0) − , (2R ∨ 2) + )} resp. {z ∈ C : Re(z) ∈ (−∞, M + )} for some > 0. For example, if L is chosen to be an NIG process with Lévy exponent
ψL (u) = uµ + δ( α2 − β2 − α2 − (β + u)2 ), for µ ∈ R, δ, α > 0, β ∈ (−α, α) in the GammaOUtimechange model from Example 15.9, one easily shows that the Lévy exponents ψ Z and ψ L admit analytic extensions to {z ∈ C : Re(z) ∈ (−∞, b)} resp. {z ∈ C : Re(z) ∈ (−α − β, α − β)}. Consequently, checking the validity of Assumption 15.10 amounts to verifying M < b,
2R > −α − β,
2R ∨ 2 < α − β,
for M = ((1 − e−λT )/λ)2 max{ψL (R ∧ 0), ψ L (R ∨ 0)} ≥ 0 in this case. By [7, Theorem 2.16(ii)], Assumption 15.10 implies that the exponential moment formula (15.3) holds for all z ∈ S. In particular, S is squareintegrable. We proceed by providing sufficient and essentially necessary conditions that ensure the validity of the martingale Assumption 15.2 and the nondegeneracy condition (15.1). Assumption 15.12. Assume that the martingale conditions T T , 0, 1 = 1 , 0, 1 = 0 0 N N are satisfied and suppose that for T δ0 := 0 , 0, 2 , N
δ1 := 1
(15.4)
T , 0, 2 , N
we have δ0 , δ1 ≥ 0 and δ0 > 0
or δ1 yt > 0 a.s. for all t ∈ T .
(15.5)
Example 15.13. For OUtimechange models, the martingale conditions (15.4) read as ψL (1) = 0, i.e., exp(L) has to be a martingale. For example, in the NIGOU models from Example 15.11, this means
µ = δ( α2 − (β + 1)2 − α2 − β2 ).
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
FA
383
DiscreteTime VarianceOptimal Hedging
As for the nondegeneracy condition (15.5), the term δ0 + δ1 y is actually bounded away from zero in most applications. (1) In OUtimechange models satisfying the conditions of Example 15.11, 1 (s, 0, 2) = ψL (2)(1 − exp(−λs))/λ > 0, unless L is deterministic. More T/N over, we have 0 (T/N, 0, 2) = 0 ψZ (1 (s, 0, 2))ds, which is also positive by [19, Theorem 21.5]. Since (yt )t∈T0 is bounded from below by exp(−λT )y0 > 0, the term δ0 + δ1 y is bounded away from zero in this case. In particular, (15.5) is satisfied. (2) Now suppose that the Ornstein–Uhlenbeck process y is replaced by a squareroot process √ dyt = κ(η − yt )dt + σ yt dWt ,
y0 > 0,
where κ, η, σ > 0 and W denotes a standard Brownian motion. Subject to certain regularity conditions [cf. 12, Assumption 4.2], the proof of [12, Theorems 4.3, 4.4] and a comparison argument show that 1 (s, 0, 2) > 0 for s > 0. This in t turn yields 0 (t, 0, 2) = κη 0 1 (s, 0, 2)ds > 0 for t > 0 and hence δ0 , δ1 > 0. Since y is positive, this shows that δ0 + δ1 y is bounded away from zero for these CIRtimechange models as well and (15.5) holds. From now on, Assumptions 15.10 and 15.12 are supposed to be in force. Combined with Theorem 15.6, Assumption 15.10 allows us to compute the varianceoptimal initial capital v0 and the varianceoptimal hedging strategy ϕt at time t by performing single numerical integrations. Theorem 15.14. For tn ∈ T0 and z ∈ R + iR, we have V(z)tn = Stzn exp(0 (T − tn , 0, z) + 1 (T − tn , 0, z)ytn ). Moreover, for tn ∈ T , ϕtn
=
R+i∞ R−i∞
V(z)tn−1 Stn−1
exp(κ0 (tn , z) + κ1 (tn , z)ytn−1 ) − 1 l(z)dz, exp(δ0 + δ1 ytn−1 ) − 1
where, for j = 0, 1, T δ j = j , 0, 2 , N T T , 1 (T − t, 0, z), z + 1 − j , 1 (T − t, 0, z), z . κj (t, z) := j N N
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
384
FA
Kallsen et al.
Proof. The formula for V(z) follows immediately from Assumption 15.10 and [7, Theorem 2.16(ii)]. Analogously, we obtain E[St2n Ftn−1 ] − St2n−1 = St2n−1 (exp(δ0 + δ1 ytn−1 ) − 1)
(15.6)
and E[V(z)tn Stn Ftn−1 ] − V(z)tn−1 Stn−1 = V(z)tn−1 Stn−1 (exp(κ0 (tn , z) + κ1 (tn , z)ytn−1 ) − 1),
(15.7)
with
T , 1 (T − t, 0, z), z + 1 N T −0 T − t + , 0, z + 0 (T − t, 0, z), N T T , 1 (T − t, 0, z), z + 1 − 1 T − t + , 0, z . κ1 (t, z) = 1 N N
κ0 (t, z) = 0
By the martingale property of V(z), we have V(z)tn−1 = E[V(z)tn Ftn−1 ]. Together with [7, Theorem 2.16(ii)], this establishes the semiflow property e0 (T −tn−1 ,0,z)+1 (T −tn−1 ,0,z)ytn−1 = e0 (T −tn ,0,z)+0 (T/N,1 (T −tn ,0,z),z)+1 (T/N,1 (T −tn ,0,z),z)ytn−1 , for tn ∈ T and z ∈ R + iR. Insertion into (15.7) yields the assertion.
We now consider the expression for the minimal expected squared hedging error in Corollary 15.7. The first term J1 represents the variance of an unhedged exposure to the option. In view of Assumption 15.10, it can be computed by evaluating a double integral with the following integrand. Lemma 15.15. For z1 , z2 ∈ R + iR, we have E[J1 (z1 , z2 )] = V(z1 + z2 )0 − V(z1 )0 V(z2 )0 . Proof. This is due to the martingale property of V(z1 + z2 ).
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
FA
DiscreteTime VarianceOptimal Hedging
385
We now turn to the second term in the formula for the hedging error in Corollary 15.7. Suppose for the moment that (15.2) holds. By Eqs. (15.6) and (15.7), we have J2 (tn , z1 , z2 ) = V(z1 )tn−1 V(z2 )tn−1 ×
(eκ0 (tn ,z1 )+κ1 (tn ,z1 )ytn−1 − 1)(eκ0 (tn ,z2 )+κ1 (tn ,z2 )ytn−1 − 1) . eδ0 +δ1 ytn−1 − 1
In view of Corollary 15.7, it therefore remains to compute E[J2 (tn , z1 , z2 )] = (I(κ1 (tn , z1 ) + κ1 (tn , z2 ), tn , z1 , z2 )eκ0 (tn ,z1 )+κ0 (tn ,z2 ) + I(0, tn , z1 , z2 ) − I(κ1 (tn , z2 ), tn , z1 , z2 )eκ0 (tn ,z2 ) − I(κ1 (tn , z1 ), tn , z1 , z2 )eκ0 (tn ,z1 ) ) × S0z1 +z2 e0 (T −tn−1 ,0,z1 )+0 (T −tn−1 ,0,z2 ) , where I(u, tn , z1 , z2 ) := E
e(u+1 (T −tn−1 ,0,z1 )+1 (T −tn−1 ,0,z2 ))ytn−1 +(z1 +z2 )Xtn−1 . exp(δ0 + δ1 ytn−1 ) − 1
Unfortunately, I(u, tn , z1 , z2 ) can only be computed explicitly in some very special cases, unlike for continuoustime varianceoptimal hedging. For example, if N = 1, i.e., for static hedging, the sum in Corollary 15.7 only consists of the term J2 (T, z1 , z2 ), which is also deterministic in this case. Hence, we obtain E[J2 (T, z1 , z2 )] = V(z1 )0 V(z2 )0
(eν0 (z1 )+ν1 (z1 )y0 − 1)(eν0 (z2 )+ν1 (z2 )y0 − 1) , e0 (T,0,2)+1 (T,0,2)y0 − 1
for νj (z) = j (T, 0, z + 1) − j (T, 0, z), j = 0, 1, which allows to compute the static hedging error by evaluating a double integral. For Lévy processes, we have δ1 = 0. Hence the denominator in the expression for I reduces to a constant in this case and the expectations can be computed using (15.3). This leads to the formula obtained in [10]. For affine models, one can verify that as the number of trading times N tends to infinity, the argument of the expectation in the expression for I converges to an
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
386
FA
Kallsen et al.
expression of the form a + bytn−1 exp(uytn−1 + (z1 + z2 )Xtn−1 ). c + dytn−1 The expectation of this term can then be calculated, cf. the proof of [13, Theorem 4.2] for more details. If the set of trading times is finite, it does not seem possible to calculate I in closed form. We discuss two ways to circumvent this problem and tackle the computation of the hedging error. The first is to use the approximation exp(δ0 + δ1 yt ) − 1 ≈ δ0 + δ1 yt , which seems reasonable as the number N of trading dates tends to infinity, because both δ0 = 0 (T/N, 0, 2) and δ1 = 1 (T/N, 0, 2) converge to zero in this case. We obtain the following firstorder approximation. Theorem 15.16. Suppose that for any t ∈ T0 , the following holds. (1) The mappings (u1 , u2 ) → j (T −t, u1 , u2 ), j = 0, 1 admit analytic extensions to S := {z ∈ C2 : Re(z) ∈ (−∞, (M ∨ 0) + ) × ((2R ∧ 0) − , (2R ∨ 2) + )}, for some > 0 and M := M ∨ 21 (T/N, M/2, R + 1). (2)
E
exp(M yt + 2RXt ) < ∞. δ0 + δ1 yt
(15.8)
Then (15.2) is satisfied and for n = 1, . . . , N, z1 , z2 ∈ R + iR, u ∈ 1 (T − tn−1 , 0, z1 ) + 1 (T − tn−1 , 0, z2 ) + {κ1 (tn , z1 ) + κ1 (tn , z2 ), κ1 (tn , z1 ), κ1 (tn , z2 ), 0}, we have exp(uytn−1 + (z1 + z2 )Xtn−1 ) E δ0 + δ1 ytn−1 −uδ0 /δ1 1 δ1 δ0 e + us exp us + χ0 (s) + χ1 (s)y0 ds if δ0 , δ1 = 0, δ1 δ0 δ1 0 1 1 + us = 1 exp(ϑ0 (s) + ϑ1 (s)y0 )ds if δ0 = 0, δ1 0 s 1 exp(0 (tn−1 , u, z1 + z2 ) + 1 (tn−1 , u, z1 + z2 )y0 ) if δ1 = 0, δ0
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
FA
387
DiscreteTime VarianceOptimal Hedging
where, for j = 0, 1, T , 0, 2 , δ j = j N
δ1 χj (s) := j tn−1 , log(s) + us, z1 + z2 , δ0
ϑj (s) := j (tn−1 , log(s) + us, z1 + z2 ). Proof. In view of (15.5), we have exp(δ0 + δ1 yt ) − 1 > δ0 + δ1 yt , which combined with (15.6) yields
E[V(z1 )tn Stn Ftn−1 ] − V(z1 )tn−1 Stn−1 1
E[J2 (tn , z1 , z2 )]
0 which does not depend on ω and z1 , z2 . Consequently, (15.8) and Assumption 15.4 yield (15.2). The second part of the assertion now follows along the lines of the proof of [13, Theorem 4.2] under the stated assumptions.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
388
FA
Kallsen et al.
Example 15.17. In most applications the denominator in (15.8) is actually bounded away from zero (cf. Example 15.13). In this situation, (15.8) follows immediately from Condition 1 of Theorem 15.16 and [7, Theorem 2.16(ii)]. In view of Theorem 15.16, the hedging error can be approximated by a sum of triple integrals with known integrands. Notice that because 1 δ0 + δ1 yt +
1 (δ 2 0
+ δ 1 yt
)2
=
1 1 − , δ0 + δ1 yt δ 0 + 2 + δ1 y t
a secondorder approximation based on 1 exp(δ0 + δ1 yt ) − 1 ≈ δ0 + δ1 yt + (δ0 + δ1 yt )2 2 follows directly from Theorem 15.16. Instead of using the closedform approximation proposed above, one can eschew semiexplicit computations and instead calculate the hedging error using a Monte–Carlo simulation as in [6]: (1) Simulate K ∈ N independent trajectories (y(ωk ), X(ωk )), k = 1, . . . , K of (y, X) and compute the realizations S(ωk ) = S0 exp(X(ωk )) and H = f(ST (ωk )) of S and H. (2) Calculate the values of v0 and ϕtn (ωk ), tn ∈ T using numerical integration to evaluate the formulas from Theorem 15.14. (3) Compute the realized squared hedging errors J0 (ωk ) = v0 +
2 ϕtn (ωk )Stn (ωk ) − f(ST (ωk )) .
tn ∈T
(4) Use the empirical mean
1 K
K k=1
J0 (ωk ) as an estimator for J0 .
In addition to its simplicity, this approach has the advantage of approximating the entire distribution of the hedging error, rather than just its mean. On the other hand, computation time is increased.
15.5. NUMERICAL ILLUSTRATION In order to illustrate the applicability of our formulas and examine the effect of discrete trading, we now investigate a numerical example. More specifically, we consider the NIGGammaOU model from Examples 15.9 and 15.11. As for parameters, we use the
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
FA
389
DiscreteTime VarianceOptimal Hedging
values estimated in [16] using the generalized method of moments, adjusting the drift rate µ of L in order to ensure the martingale property of S: β = −16.0,
α = 90.1,
λ = 2.54,
δ = 85.9,
a = 0.847,
µ = 15.0,
b = 17.5.
By Example 15.11, Assumption 15.10 and the prerequisites of Theorem 15.16 are satisfied for European call options and R = 1.1. Henceforth, we consider a European call with discounted strike K = 100 and maturity T = 0.25 years. The results for the varianceoptimal initial hedge ratio ϕ0 for N = 1 (static hedging) and N = 12 (weekly rebalancing) are shown in Figs 15.1 and 15.2. For static hedging, the impact of discretization seems to be quite pronounced, in particular for outofthemoney options. Also notice that this effect turns out to be substantially bigger for the NIGGammaOU than for the Black–Scholes model. For weekly rebalancing, the effect of discretization on the initial hedge ratio already becomes marginal. More specifically, the difference between the discrete and continuoustime varianceoptimal hedging strategies is barely visible in Fig. 15.2. Figure 15.3 shows a simulated path of the discrete varianceoptimal hedges for N = 1, 3, 12, 60.
Figure 15.1 Varianceoptimal initial hedge ratios for N = 1.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
390
FA
Kallsen et al.
Figure 15.2 Varianceoptimal initial hedge ratios for N = 12. Hedging strategy (Strike = 100, T = 0.25, N = 1, 3, 12, 60 ) 1
0.9
0.8
Hedge ratio
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
N=1 N=3 N=12 N=60 0.05
0.1
0.15
0.2
0.25
Time
Figure 15.3 A simulated path of optimal hedge ratios for N = 1, 3, 12, 60.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
DiscreteTime VarianceOptimal Hedging
FA
391
We now turn to the minimal expected squared hedging error, which is depicted for N = 1 (static hedging) to N = 60 (daily rebalancing) in Fig. 15.4. As the number N of trading dates tends to infinity, the discrete hedging errors approach the respective continuoustime limits both in the Black–Scholes model and in the NIGGammaOU model. Naturally, this limit vanishes in the complete Black–Scholes model. As noticed above, the static hedging error for N = 1 can be computed without using any approximations. For N ≥ 2, the discretetime hedging error in the given NIGGammaOU model is approximated surprisingly well by the sum of the respective continuoustime hedging error and the corresponding discretetime hedging error in the Black– Scholes model. In fact, the maximal absolute difference is smaller than 0.045. If such an approximation can be used for the specific model at hand, computation time can often be drastically reduced by evaluating the formulas from [10, 12] instead of Theorem 15.16. Note that the discrete hedging errors in the NIGGammaOU model have been approximated using Theorem 15.16. Since the corresponding results for a simulation study using one million Monte–Carlo runs differ by less than 2.5% for N = 1, . . . , 60, we do not show them here. However, in Fig. 15.5, we use the results of the Monte– Carlo study to depict an approximation of the distribution of the hedging error for
Figure 15.4 Minimal expected squared hedging errors.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
392
FA
Kallsen et al.
Figure 15.5 Approximated distribution of the hedging error for N = 1, 12, 60.
N = 1, N = 12, and N = 60. Apparently, not only the variance of the hedging error, but also its law depend crucially on the rebalancing frequency.
Acknowledgments The first and fourth authors gratefully acknowledge financial support through Sachbeihilfe KA 1682/21 of the Deutsche Forschungsgemeinschaft.
References [1] BarndorffNielsen, O and N Shephard (2001). NonGaussian OrnsteinUhlenbeckbased models and some of their uses in financial economics. Journal of the Royal Statistical Society, Series B, 63, 167–241. [2] Bertsimas, D, L Kogan and A Lo (2001). Hedging derivative securities and incomplete markets: An arbitrage approach. Operations Research, 49, 372–397. [3] Carr, P, H Geman, D Madan and M Yor (2003). Stochastic volatility for Lévy processes. Mathematical Finance, 13, 345–382. ˇ [4] Cerný, A (2007). Optimal continuoustime hedging with leptokurtic returns. Mathematical Finance, 17, 175–203.
May 12, 2010
17:47
WSPC/SPIB913
DiscreteTime VarianceOptimal Hedging
b913ch15
FA
393
ˇ [5] Cerný, A and J Kallsen (2007). On the structure of general mean–variance hedging strategies. The Annals of Probability, 35, 1479–1531. [6] Cont, R, P Tankov and E Voltchkova (2007). Hedging with options in presence of jumps. In Stochastic Analysis and Applications: The Abel Symposium 2005 in honor of Kiyosi Ito, F Benth, G Di Nunno, T Lindstrom, B Øksendal and T Zhang, (eds.), pp. 197–218. Berlin: Springer. [7] Duffie, D, D Filipovic and W Schachermayer (2003). Affine processes and applications in finance. The Annals of Applied Probability, 13, 984–1053. [8] Filipovi´c, D and E Mayerhofer (2009).Affine diffusion processes: Theory and applications. Radon Series on Computational and Applied Mathematics, 8, 1–40. [9] Heston, S (1993). A closedform solution for options with stochastic volatilities with applications to bond and currency options. The Review of Financial Studies, 6, 327–343. [10] Hubalek, F, L Krawczyk and J Kallsen (2006). Varianceoptimal hedging for processes with stationary independent increments. The Annals of Applied Probability, 16, 853–885. [11] Kallsen, J (2006). A didactic note on affine stochastic volatility models. In From Stochastic Calculus to Mathematical Finance, Y. Kabanov, R Liptser, and J Stoyanov (eds.), pp. 343– 368. Berlin: Springer. [12] Kallsen, J and A Pauwels (2009). Varianceoptimal hedging for timechanged Lévy processes. To appear in Applied Mathematical Finance. [13] Kallsen, J and A Pauwels (2009). Varianceoptimal hedging in general affine stochastic volatility models. To appear in Advances in Applied Probability. [14] Kallsen, J and RVierthauer (2009). Quadratic hedging in affine stochastic volatility models. Review of Derivatives Research, 12, 3–27. [15] Lamberton, D and B Lapeyre (1996). Stochastic Calculus Applied to Finance. London: Chapman & Hall. [16] MuhleKarbe, J (2009). On UtilityBased Investment, Pricing and Hedging in Incomplete Markets. Dissertation, Technische Universität, München. [17] Pauwels, A (2006). Varianzoptimales Hedging in affinen Volatilitätsmodellen. Dissertation, Technische Universität, München. [18] Pham, H (2000). On quadratic hedging in continuous time. Mathematical Methods of Operations Research, 51, 315–339. [19] Sato, K (1999). Lévy Processes and Infinitely Divisible Distributions. Cambridge: Cambridge University Press. [20] Schoutens, W (2003). Lévy Processes in Finance. New York: Wiley. [21] Schweizer, M (1995). Varianceoptimal hedging in discrete time. Mathematics of Operations Research, 20, 1–32. [22] Schweizer, M (2001). A guided tour through quadratic hedging approaches. In Option Pricing, Interest Rates and Risk Management, E Jouini, J Cvitanic, and M Musiela (eds.), pp. 538–574. Cambridge: Cambridge University Press.
May 12, 2010
17:47
WSPC/SPIB913
b913ch15
This page intentionally left blank
FA
May 12, 2010
18:8
WSPC/SPIB913
b913Index
FA
INDEX
absolute ambiguity robust adjustment (AARA), 330 affine process, 376, 380 alternative asset classes, 52 Alternative real assets, 51–54, 69 ambiguity, 327–337, 339–344 ambiguity robust adjustment (AARA), 330, 333 ambiguity robust adjustment (RARA), 330 Angströmtype equation, 58 asset allocation, 147–149, 162, 165, 168, 170, 171 asset class, 3, 4, 8–13, 17–19 Autocorrelation, 65, 66 aviation, 53
calendar spread option, 118, 119 capitalmarket scenarios, 149, 170 CCO, 177 CDS/credit default swaps, 259–261, 265–267, 270, 271, 283, 285–290 CDX, 261, 266, 272, 275–277, 285, 289, 290, 292 CFXO, 177 CIRtimechange model, 383 closedend funds, 52 cointegration, 84–86 collateralized asset obligation, 179 collateralized commodity obligation, 177 collateralized debt obligation, 176 collateralized fx obligation, 177 collateralized trigger swap obligations, 177 comonotone random vector, 365 conditional value at risk optimization/CVaR optimization, 147, 157, 164–169, 171 confidence ellipsoid, 317, 319 constant proportion portfolio insurance (CPPI), 227, 228, 232 copula approach, 186, 187, 189, 190, 193 corporate bonds, 148, 158, 160, 162, 163, 165, 166, 170 correlated default times, 147, 150, 152, 158
backtesting, 88 barrier options, 345, 347, 350, 352–355, 359, 361–364 Black–Litterman expectation, 9, 10, 17 Black76, 98, 105, 114, 116, 117 Bonus (Guarantee) Certificates, 134 borrowing constraints, 227, 230 bottomup approach, 52, 54, 55, 68 Brownian motion, 263, 264, 271, 272, 275, 277, 285, 290
395
May 12, 2010
18:8
WSPC/SPIB913
b913Index
FA
396 CPDO/constant porportion debt obligation, 259–261, 285–292 CPPI, 201–203, 205–207, 209–211, 213–224 CPPI strategies, 295, 296, 300, 304, 305, 311, 321, 323, 324 CPPI/constant proportion portfolio insurance, 259–261, 267–270, 275, 278, 281, 282, 284, 285, 289, 291, 292 Credit default swap index/CDS index, 154, 155, 157, 161, 169 Credit default swap/CDS, 148, 149, 153–155, 157, 158, 160–163, 165, 169, 170 CreditGrades, 125 cross asset portfolio derivatives, 175–177, 182, 191, 195 cushion, 267–269, 275, 277, 282, 284, 305, 307 default risk, 123 defaultable knockout put options, 126 discount certificates, 126 discretely sample options, 347, 362, 367 dynamic asset allocation, 296 dynamic portfolio insurance, 201 Economic Scenario Generator, 56 efficient portfolios, 295, 296, 299, 300, 312, 313, 316, 324, 325 estimation risk, 295, 296, 298, 313, 314, 316, 320, 321, 324, 325 ETF, 201, 203, 210, 223, 224 European options, 348, 349, 355, 357, 358, 361, 363, 367, 369, 370 (extended) rolldown call, 362 expected shortfall, 308, 309, 312 expectedsquared hedging error, 375 explanatory variables, 72, 79–82, 84–86, 90 exposure, 303, 305, 307 extreme value theory (EVT), 282 floor, 303, 305, 307–309 forward freight agreement, 77, 78 freight derivatives, 71, 77, 78
Index freight futures, 71, 78, 79, 81, 86 freight rates, 74, 75, 77, 80–82, 90 Gamma distribution, 271 GammaOU process, 381 gap risk, 229, 230, 241, 246, 247, 251, 259, 268, 270, 277–279, 281–284, 287, 292 Global irradiance, 53, 55–58, 60–62 granger causality, 82, 83 guarantees, 228–230, 242 Heath–Jarrow–Morton, 94 HullWhite model, 56 index certificates, 124, 126, 135 inflation, 51–53, 55–57, 65 inflation protection, 52, 53, 55 infrastructure, 51–53 issuer risk, 123–125, 130, 132, 139 iTraxx, 261, 266, 272, 275–277, 285, 289, 290, 292 Jensen’s Inequality, 107–110 Lévy process, 264 Laplace transform approach, 376, 378 leverage, 261, 267, 268, 275, 285–288, 290, 291 LIBOR model, 98 limited partnership fund, 51, 53 lognormal (shifted), 126 lookback options, 346, 347, 364, 368 Macroeconomic factors, 56, 57 Margrabe Formula, 117 MarkovSwitching model, 3, 4, 11, 12, 19 Markov transition matrix, 4, 6, 7 maximum Sharpe ratio portfolio, 299, 301, 312, 314, 316, 321, 322 meanvariance optimization, 299, 300, 310, 311 method of moments, 11 minimum variance portfolio, 299, 301, 312, 316, 320, 322–324 modelindependent strategies, 359 Monte Carlo simulation, 13
May 12, 2010
18:8
WSPC/SPIB913
Index multiasset portfolio, 51, 52, 68 multiplier, 305–313, 315, 316, 320–324 multivariate variance gamma model, 289 NIG process, 382 Normalinverse gaussian onefactor copula/NIG onefactor copula, 147, 149, 152, 171 nthtotrigger basket, 177, 178, 183 oil, 72–74, 76, 77, 81–84, 86 optionbased portfolio insurance (OBPI), 228, 232 OU timechange model, 382, 383 Participation Guarantee Certificates, 132 pathindependent payoff, 347 photovoltaic, 51–56, 58, 61–64, 66–69 Portfolio, 51, 52, 56, 63–69 portfolio insurance, 259, 267 portfolio optimization, 13, 149, 156, 160, 162, 171, 231 power forward, 93–96, 98–103, 107, 108, 110–120 power markets, 93, 94, 96, 97 PutCallParity, 346, 349, 352, 360 ratchet call, 363 real return, 53 renewable energy, 51–53 robust counterpart, 295, 316, 317, 320 robust meanvariance optimization, 316 Samuelson effect, 97, 100 semistatic hedging strategies, 345, 352, 370 shipping, 51–54 Shipping market, 72, 75, 80 shortfall probability, 250, 251, 308, 309, 311, 312, 316, 322, 324 socially responsible investing (SRI), 3, 4 solar plants, 51, 53, 54
b913Index
FA
397 static hedging, 385, 389, 391 static hedging strategies, 362, 370 stoploss transform, 365 strong pathdependence, 347 structural approach, 185, 186, 188, 192, 193 subreplication (sub replicating strategies), 354, 359 subjective expected utility (SEU), 328 superreplication (super replicating strategies), 359, 361, 362, 365 sustainability, 3, 4, 6–8, 18 sustainability score, 6, 7, 18 swaptions, 270, 273–275, 279, 280 Term structure of interest rates, 57 TIPP, 203, 214, 220 total variance, 66 trading route, 73, 78 transaction costs, 227, 230, 242, 246, 247, 249–251 trigger derivative, 176, 178, 179, 188, 195 uncertainty set, 317–320 utility, 227–230, 232–234, 240–244, 246, 250 functions, 232, 233, 237, 242 loss, 230, 243–245, 250, 251 variance gamma process, 264, 270, 274 varianceoptimal hedging strategy, 378, 379, 383 varianceoptimal initial capital, 375, 378, 383 Vasicek process, 56 vector auto regression model, 72, 79, 80 vessel, cargo, 72 vulnerable options, 124 weak (strong) pathdependence, 347, 350, 364 wind power plants, 53