FOUNDATIONS FOR FINANCIAL ECONOMICS Chifu Huang
Sloan School of Management
Massachusetts Institute of Technology
Rob...
627 downloads
2893 Views
9MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
FOUNDATIONS FOR FINANCIAL ECONOMICS Chifu Huang
Sloan School of Management
Massachusetts Institute of Technology
Robert H. Litzenberger The Wharton School
University of Pennsylvania"'
PRENTICE HALL, Englewood Cliffs, New Jeney 07632
Huang.
Ch1fu.
L 1 tzanbargar. c•. P.
Fcund1t1on1
Rl!pr1nt.
for
I Chtfu Huang. Robart 1t.
f1nancta1 aconoatcs
Or1g1na11y
pub11shad:
Naw
York
:
NorthHo111nd.
c1988.
ISBN 0135008538 1. InvastmantsMathallt1cal IIIDdals. 2. FtnancaMatha!lat1ca1 IDdl\1. I. L1tzenbarger. Robart H. II. T1t11. H04515.HB3 1993 332dc20 93990 Inc1udal
1ndax..
CIP
•
C 1988 by Pm!lic:eHall, IDe. A Simon &. Schuster Company
Englewood Cliff•, New 1ency IJ7632
AU cifh11 re1�. ·No pat of Ibis book may be .ri:piOduc:ed, in any form or by any means, �pc�lim in wrilina fmm lhe puhlilher.
Piitatt.d m die Ulliled Slara rl. America 10 9 8 1 6 5 4 3
;.; I�BN 0 13 50 D.b 53 8 .
PrenticeHalliDtemllliooal (UK) Umited, l..oftfJo11 PrenticeHall ol Australia Ply. Umired, S�y PrenticeHall Canada IDe., TMorllo ��.. �spanoamericana, S. A., Muico : �HIIIJ: !)f India Privarc Umired, New Delhi : PrenticeHall.of J�P,an. IDe., Tokyo .SimOR &. Schuster.;A.sia Pie. Lld., SitJBapore
EdiiDi"a Pn;QI.ice.H.U\do Bllllil, Llda., Rio de JaMiro .
. .
.
. . �· ·., .
�
To Our Families
TABLE OF CONTENTS
1. 2.
3.
"·
5.
Preface
ix
Preferences Representation and
1
Risk Aversion
59
Mathematics of the Portfolio Frontier
83
Two Fund Separation and Linear Valuation
119
Allocative Efficiency and the Valuation of State Co ntingent Securities
6.
39
Sto chastic Dominance
Valuation of Complex Securities and Options with Preference
,.�
151
Restrictions
7.
Multiperiod Securities Markets 1:
Equilibrium Valuation
179
vii
viii
Table of Contents
8.
M ultiperiod Securities Markets II:
9.
Financial Markets with Differential
10.
Valuation by Arbitrage
223 259
Information Econometric Is s ues in Testing
299
the Capital As s et Pricing Mo del Index
361
PREFACE
This book evolved from lecture notes we have used to teach in troductory PhD courses in financial economics at the Massachusetts Institute of Technology, Stanford University, and the University of Pennsylvania. Its purpose is to provide the foundations for the study of modern financial economics. Rather than giving a sup erficial cov erage of a wide range of topics, we have chosen to concentrate our ·discussion on individuals' consumption and portfolio decisions under uncertainty and their implications for the valuation of securities.
Chapters 1 through 6 discuss twoperiod models, where the
consumption and portfolio decisions are made only once at the i.Jlitial
date of the economies. Chapter 1 analyzes an individual's behavior
under uncertainty. This chapter also shows the comparative statics of an individual's optimal portfolio choice in an economy with one _.. riskless and one risky asset when his initial wealth or attitude toward risk changes.
Moreover, we provide sufficient conditions for these
comparative statics to apply to economies with three or more assets.
In Chapter
2,
we discuss three concepts of Btoche&Btic domine&nce.
The concepts of stochastic dominance identify conditions that allow
risky assets to be ranked based on limited knowledge of individuals'
preferences. Chapter 3 shows mathematical properties of a portfolio frontier  the collection of portfolios that have the minimum variance lX
X
for different levels of expected rates of return. In Chapter
4,
distributional conditions on the rates of return on assets
we give
so
that
individuals will optimally choose to hold portfolios on the portfolio frontier.
As a consequence, expected rate of return on an asse t is
linearly related to its beta, which measures the contribution of the
asset to the risk of a welldiversified portfolio. This is the Capital Asset Pricing Model. We also discuss in this chapter the Arbitrage Pricing Theory, which relates the expected rate of return of an asset
to a number of random factors. Chapter
5
begins our description of a state contingent security
and its equilibrium valuation. A state contingent security pays one unit of consumption in one state of nature and nothing otherwise.
Markets are said to be complete if there is a state contingent secu
rity for every state of nature. An allocation of consumption among individ uals is said to be a Pareto optimal allocation if there is no other allocation that increases an individual's satisfaction without decreasing some other individual's satisfaction. We show how Pareto
optimal allocations can be achieved in complete markets as well as
in various other market structures. The allocational role of options,
in particular, is demonstrated. This chapter also prov ides the neces
sary and sufficient conditions on individuals' utility functions for all
Pareto optimal allocations to be achievable by holding the portfolio
of all assets and borrowing or lending and discusses the relation
ship between these conditions and an aggregation re�.ult in securities . In Chapter 6, we present general pricing rules for secu
markets.
rities that pay off in more than one state of nature and specialize
these rules with additional preference restrictions. In particular, we derive a dosed form solution for a call option written on a com mon stock when the random payoff's of the common stock and the aggregate consumption are jointly lognormally distributed and in dividuals' preferences are represented by power functions with the same exponent. We discuss in Chapter
7
how multiperiod dynamic economies
can be modeled. A multiperiod dynamic economy differs from a two period static economy in that trading can take place at more than
one date and individuals' expectations about future prices are there fore essential in an equilibrium specification. This leads to a notion
rrerace
xi
of a rational expectations equilibrium. The general equilibrium val uation principles in a multiperiod dynamic economy are essentially the same as those in a twoperiod static economy;
An important
feature of a multiperiod economy demonstrated in detail is that a Pareto optimal allocation can be achieved by trading dynamically in a limited number of longlitJed securities. Chapter
8
continues our
discussion of a mul tiperiod economy with emphasis on valuation by arbitrage. We show the connection between an arbitragefree price system and martingales. This connection allows us to compute prices of a derivative security in a simple way when the derivative security can be priced by arbitrage . As an example, we price a call option written on a stock when the stock price follows a binomial random walk.
A common feature of the economies in Chapters 1 through 8 is
that individuals are endowed with the same information. In Chap ter
9 we discuss
economies in which individuals have differential in
formation. We demonstrate that equilibrium properties can be very different from those in economies without differential information.
Chapter 10 examines econometric issues of testing the Capital Asset Pricing Model. Some test statistics are given geometric interpreta
tions in the context of Chapter
3.
Applications of information economics t o financial markets have gained significant importance in recent years. Our coverage in Chap ter
9 is
limited:.in scope.
Chapter 10 concentrates on econometric
issues in testing the Capital Asset Pricing Model. Empirical aspects
of many other theories developed in this book also deserve attention.
Separate books can be written on the general topic areas of Chapters
9 and 10. Our selection of subjects covered in these two chapters is intended to be an introduction.
""
Besides providing material for introductory PhD courses in fi nancial economics, this book can be used for a graduate /advanced undergraduate course in the economics of uncertainty. When supple mented with articles, this book can form the basis for a two semester course. Chapters 1 through
4
and Chapter 10 are recommended for
the first semester, while Chapters
5 through 9
are recommended for
the second semester. Although the level of presentation is rigorous in general, the necessary prerequisites are only intermediate level
xii
Preface
microeconomics, introductory econometrics, matrix algebra, and el ementary calculus. We owe a substantial debt to our academic colleagues who have contributed to the strong theoretical foundations of asset choice and valuation under uncertainty and to the empirical methodology for examining testable implications of the theory.
This book presents
and interprets materials in the existing literature and does not make original contributions of its own. In the end of each chapter, we try to give a brief attribution of the materials covered. But, undoubt edly, our attribution cannot be complete.
Many of our colleagues
provided either helpful comments on early drafts or encouragement throughout the years that this book was under preparation. Among them, special thanks go to Sudipto Bhattacharya and John Cox, from whom we have received continuous encouragement and invalu able suggestions on the selection of topics; and to Michael Gibbons and Craig MacKinlay, who have helped clarify some questions we had on the materials in Chapter
10.
Many of our students have provided
helpful comments and suggestions. Among them, special apprecia
tions go to Ayman Hindy, who read through every chapter in detail
and pointed out numerous mistakes in notation and derivations; to
Caterina Nelsen, whose editorial help has proved indispensible; and to Ajay Dravid and Tomas Philipson, who gave helpful comments in terms of style and topic selection. This book was completed when the first author was on a fel lowship provided by Batterymarch Financial Management, whose
generous support is greatly appreciated. The entire book is typeset
by 'lEX designed by Professor Donald Knuth of Stanford University.
Manjul Dravid helped us typeset some parts of this book. Our con versation with Hal Varian improved our understanding of some fine points of '!EX, for which we are thankful.
Last, but by no means
least, we are grateful to our families for their encouragement and help.
Chifu Huang Robert H. Litzenberger
FOUNDATIONS FOR FINANCIAL ECONOMICS
·
CHAP TER 1
PREFERENCES REPRESENTATION
AND RISK AVERSION
1.1.
As we mentioned in the preface, the main focus of this
book is on individuals' consumption and investment decisions under uncertainty and their implications for the valuation of securities. In dividuals' consumption and investment decisions under uncertainty are undoubtedly influenced by many considerations.
A commonly
�cepted theory of asse t choice under uncertainty that provides the underpinnings for the analysis of asset demands uses the e�pected util ity hypothesis. Under this hypothesis, each individual's consumption
and investment decision is characterized as if he determines the prob
abilities of possible asset payoffs, assigns an index to each possible consumption outcome, and chooses the consumption and investment policy to maximize the expected value of the index. More formally,
an individual's preferences have an expected utility representation if
there exists a function u such that random consumption
to random consumption
where
y if and
i is preferred
only if
E(u(i)] � E(u(y)],
E[·] is the expectation under the individual's probability belief.
1
2
Foundations for Financial Economics
In the first half of this chapter we give behavioral conditions that are necessary and sufficient for an individual's preferences to have an expected utility representation. We then go on to discuss the necessary and sufficient condition for an individual's preferences to exhibit risk aversion under the expected utility representation as sumption. Different measures of risk aversion will be proposed and used to analyze the comparative statics of an individual's portfolio behavior when faced with one risky asset and one riskless asset. Fi nally, we will discuss sufficient conditions for the comparative statics for the one risky and one riskless asset case to generalize to the case of multiple risky assets and a riskless asset. Before we proceed, we note that throughout this book we will use positive, negative, greater than, smaller than, increasing, decreasing, and etc. to mean weak relations. When a relation is strict, we will emphasize it by using a "strictly" to modify it, for example, by using "strictly positive." 1.2 Suppose for now that there are two dates, time 0 and time 1, and there is a single consumption good available for consumption only at time 1. Uncertainty in the economy is modeled by uncertain states of nature to be realized at time 1. A state of nature is a complete description of the uncertain environment from time 0 to time 1 . We denote the collection of all the possible states of nature by nand denote an element of nby w. At time 0, individuals know that the true state of nature is an element of n but do not know which state will occur at time 1. A consumption plan is then a specification of the number of units of the single consumption good in different states of nature. Let :r: be a consumption plan . We will use :Z:w to denote the number of units of the consumption good in state w specified by :z:. When there are five states of nature denoted by w1, . . . 1w5. Table 1 . 2. 1 tabulates a consumption plan :z:, which has 2 units of consumption in state w1, 3 units of consumption in states wz, etc. As we defined above, a consumption plan is a vector specifying units of consumption in different states of nature. Since the time 1 realized consumption is uncertain, a consumption plan :z: can also be viewed as a random variable and we will use i: to denote it when we use the "random variable" aspect of a consumption plan.
Preferences Representation and Risk Aversion
3
�w��l_·�1,L _w_; L_, w�:�l_ _w; �� � ��5� A Consumption Plan __
__
Table 1.2.1:
An individual is represented by his preference relation !: de fined on a collection of consumption plans. We will formally define a preference relation shortly. Roughly, a preference relation is a mech anism that allows an individual to compare different consumption plans. For example, given two consumption plans, z and :t, a pref erence relation enables an individual tD tell whether he prefers z to z' or z' to z. For concreteness, we would like an individual's pref erences to be represented by a utility function, or H, in the sense that the individual prefers z to z' if and only if H(z) � H(z'). We will see later in this chapter that, under some regularity conditions, a preference relation can always be represented this way. When the number of states is very large, a consumption plan z is a vector of large dimension and the function H will be complicated to analyze. It would be more convenient if there existed a function u that allowed comparison among consumption plans that are certain and a probability P that gave the relative likelihood of states of nature such that the preference relation can be represented as an ezpected utility in the sense that consumption plan z is preferred to consumption plan z' if and only if the expected utility of z is greater than the expected utility of z', that is,
l u(zw)dP(w) l u(z�)dP(w). �
Denoting the expect ation operator under equivalently written as
P
(1.2.1)
by E[·J, (1.2.1) can be
E[u (x)] � E[u (x')] ,
(1.2.2)
where we have used the random variable aspect of z and z'. No� that a consumption plan is certain if the number of units of consumption does not vary across different states of nature. Note also that in the above expected utility representation, if z and r are both certain or sure things, that is Zw = z
and z� = z'
Vw
E
n,
4
Foundations for Fin ancial Economics
for some constants z and z', then E[u(i)] = u(z) and E[u(z')] = u(z'). In this sense, u compares consumption plans that are certain. Certainly not all preference relations have an expected utility representation. Indeed , we have to put a fair amount of structure on a preference relation to achieve this purpose. In general, there are two approaches for a preference relation to have an expected utility representation, depending on whether one treats the probabilities of the states of nature as objective or subjective. The former approach was introduced by von Neumann and Morgenstern (1953) and the re sulting function u is thus called the von HeumannMorgenstern util ity function. The latter approach was taken by Savage (1972), who views probability assessments as an integral part of an investor's pref erences and thus purely subj ective. However, the distinction between subjective and objective probability assessments is inconsequential to our purpose in this book. Hence for the analysis to follow, we will not distinguish between them and always call the function u defined on sure things a von NeumannMorgenstern utility function. Before discussing the representation of a preference relation by an expected utility, in the next section , we give a definition of a preference relation and discuss conditions under which a preference relation can be represented by a utility function H. Formally, let X be the collection of consumption plans under consideration. A binary relation � on X is a collection of pairs of consumption plans (z, y). H (z, y) is in the relation, we write z � !I and say z is preferred to !I· H (z, y) is not in the relation, then we write z 't.. !I and say z is not preferred to !I · A binary relation is transitive if z � !I and !I � v imply z � v; that is, if z is preferred to !I and !I is preferred to v, then z is preferred to v. A binary relation is said to be complete if for any two consumption plans z and y, we either have z � !I or !I � z; that is, any two consumption plans can always be compared. A preference relation is a binary relation that is transitive and complete. We can also define an indifference relation and a strict preference relation. Formally, given a preference relation �, two consumption plans z and !I are said to be indifferent to each other if z � !I and !I � z, denoted by z !I· The consumption plan z 1 .3.
"""
6
P.!eferences Representation and Risk Aversion
is said to be strictly preferred to y, denoted by :t > y, if :t � !I and � :t. Note that a strict binary relation and an indifference relation can also be similarly defined for any given binary relation. 11
1 .4 . W hen X has a finite number of elements, a preference relation t can always be represented by a utility function . This assertion can be proved in a straightforward manner. The readers are asked to furnish a proof in Exercise 1 . 1 . Here we shall give an example to demonstrate the essential idea. Suppose that there are three consumption plans in X, denoted by 2:1, 2:2, and :ts. Pick any consumption plan, say :ts, and define
H(zs) = 6,
where 6 is an arbitrary constant . Next take :t1. Since a preference relation is complete, 2:1 and :ts can be compared . We define H(z!)
=
{
6+ 1 if :t1 > zs; 6 1 if zs > z1; if zs"""2:1. b
That is, compare 2:1 and :ts. If z1 is strictly preferred to zs, we assign a value strictly larger than 6 to H(z1); and similarly for other cases. Without loss of generality, suppose that z1 > zs. Finally, we compare z2 with z1 and zs, and define
H(z2)
=
if Zs > Z2j if z2 """zs; + i if z1 > z2 > zs; b+ 1 if Z2 """Z1 j 6+2 if Z2 > Z1 .
lb1:
Here we compare z2 with z1 and zs and assign values to H(z2) natural way. It should now be transparent that H(zn)
�
H(zm) if and only if Zn
t
Zm,
n,m = 1 ,2, 3 .
in
a
That is, Has defined above represents the preference relation t. When X has a countable number of elements, the above idea can be carried out ·in a similar way to conclude that a preference relation can always be represented by a utility function.
•
6
Foundations for Financial Economics
1.5. Matters are not as simple when an individual expresses his preferences on an uncountably infinite number of consumption plans. In such event, there exist wellknown examples of preference relations that cannot be represented by utility functions. The so called Lexicographic preference relation is one such example; we refer readers to Exercise 1.2 for a brief description. Thus, for general X, additional conditions on a preference relation will be needed for an expected utility representation to exist . It turns out that the additional condition needed is purely technical in nature and is stated in Exercise 1.2. Interested readers should consult Debreu (1954 ) and F ishburn (1970) for details.
1.6. Now we turn to the representation of a preference rela tion by an expected utility. Let P be a probability defined on the state space n, which can be either objective or subjective. ( For the technically inclined readers, we are a bit informal here. When n has uncountably infinite elements, a probability is actually defined not on n but rather on a collection of subsets of n that satisfies a certain structure.) A consumption plan is a random variable, whose probabilistic characteristics are specified by P. We can define the distribution function for a consumption plan x as follows:
Fz(z) = P{w En:
x111
�
z}.
If a preference relation � has an expected utility repr�sentation with a utility function u on sure things, the expected utility derived from xis
oo + E[u(i)] = oo u( z)dFz(z ) . J
From the above relation, we see that if two consumption plans z and x' have the same distribution function, they will yield the same expected utility and are indifferent to each other . This demonstrates that the primitive objects on which an individual expresses his or her preferences are probability distributions of consumption. Note that two consumption plans having the same distribution function can have very different consumption patterns across states of nature.
Preferences Represen tation and Risk Aversion
7
1. 7. To simplify matters, we shall assume that an individual only expresses his preferences on probability distributions defined on a finite set z. In other words, the collection of consumption plans X on which an individual expresses his or her preferences must have the property that Xw E
Z
'Vw
E
0, 'Vx
E
X.
For example, if Z = {1, 2, 3}, then the units of consumption in any state can only be 1, 2, or 3. This assumption can be justified, for . example, when the consumption commodity is not perfectly divisible and the supply of the commodity is finite . In this case, we can represent a consumption plan x by a function p( · ) defined on Z, where p(z) is the probability that x is equal to z. Thus p(z) � 0 for all z E Z and E.ez p(z) = 1. The distribution function for the consumption plan x discussed in Section 1.6 is then
Fz(z') = L p(z), •S.'
and
E[u(z)]
=
L u(z)p(z).
zEZ
One can also think of a consumption plan as a lottery with prizes in Z. The probability of getting a prize z is p(z). We denote the space of probabilities on Z by P and its elements by p, q, and r. If pEP, the probability of z under pis p(z). 1.8. The following three behavioral axioms are necessary and sufficient for a binary relation defined on P to have an expected utility representation.
Aziom
1.
tis a preference relation on P.
Aziom e . For all p, q, r a )r > aq + (1 a)r.
E
P and a
E
(0, 1],
p >
q implies
ap + (1 
8
Foundations for Financial Economics
This axiom is commonly called the BU6Btitution aziom or the inde pendence aziom. Think of p, q, r as lotterieB and ap + ( 1 a)r as a compound lottery: First an experiment with two outcomes ( say head and tail ) is carried out, where the probability of a head is a. If a head shows up, the lottery p is performed . If a tail shows up, the lottery r is performed . The motivation for this axiom is the following: The difference between ap + (1a)r and aq+ (1a)r is what happens if a head shows up, so how an individual feels about ap + (1 a)r versus aq + (1a)r should be determined by how he feels about p versus q. In other words, satisfaction of consumption in a given event does not depend on what the consumption would have been if another event had occurred. 
For all p, q, rE P, if p > q > r then there exists such that ap + (1 a)r > q > bp + (1 b)r. Aziom !1.
a, bE (0, 1)

This is called the Archimedean aziom. It roughly says that there is no consumption plan p so good that for q > r a small probability b of p and a large probability (1  b) of r is never worse than q. Similarly, there is no consumption plan r so bad that for p > q a large probability a of p and a small probability (1 a) of r is never preferred to q. It is called an Archimedean axiom because of the resemblance to Archimedes' principle : No matter how small z > 0 is and how big z' is, there is an integer k such that kz > z'. When p, q, and r are sure things, say when p (z) = 1 , q(z') = 1, and r (z") = 1, the Archimedean axiom says that there exists a lottery awarding z with a probability a and awarding z" with a probability (1a) which is strictly preferred to the sure consumption level z'; and there exists a lottery awarding z with a probability b and awarding z" with a probability (1  b) so that the sure consumption z' is strictly preferred to this lottery. We will show in the following section that a binary relation � on P has an expected utility representation if and only if it satisfies the above three axioms. Before we do that, we will first record some very intuitive properties of a binary relation when the three axioms 1.9.
Preferences Represen tation and Risk A version
9
are satisfied . Their proofs are straightforward , and we leave them for the readers as Exercise 1.3 at the end of this chapter. We will need the following notation. For z E Z, let P. be the probability distribution degenerate at z in that p
•
(z') =
{ 01
if z' if z'
= z;
'I z.
That is, P. represents the sure consumption plan that has z units of consumption in every state. S uppose henceforth that t is a binary relation on P that satisfies the above three axioms. Then >
q and 0 �a< b �1 imply that bp + (1b)q > ap + (1a)q. 2. p t q t r and p > r imply that there exists a unique a• E [0, 1] such that qa•p + (1a•)r. 3 . p > q and r > s and a E [0, 1] imply that ap + (1  a)r > aq + (1 a)s. 4. p q and a E [0, 1] imply that pap + (1a)q. 5. p q and a E [0, 1] imply that ap + (1 a)raq + (1 a)r, for all rEP. 6. There exist Z0, z0EZ such that P.o t p t P.o for all pEP. 1. p
.
The first property says that if p is strictly preferred to q, then any compound lottery on p and q with a strictly higher weight for pis strictly preferred to a compound lottery with a lower weight for p. The last property can be seen as follows. Suppose that t always prefers more to less. Since Z has only a finite number of elements, there exist a maximum z0 and a minimum z0• A sure consumption plan P.o is certainly preferred to any other consumption plan , and conversely for P.o . Other properties can be interpreted similarly. 1.10. Now we will prove that t has an expected utility repre sentation. We take cases. Case 1. P. o  P.o. Then p  q for all p, qE P. Therefore any u(z) = k for a constant k will be a utility function for sure things. Case 2. P.o > P.0• For p E P, define H(p) = a, where a is a number in [0, 1] such that aP.o + (1  a)P.o  p. That is, we
10
Foundations for Fin ancial Economics
define H(p) to be the weight a compound lottery on P... and Pa., assigns to P... to make it indifferent top. We know from property 2 of the previous section that a is unique. So H(p) is well defined for all p E P. Note that by the definition of H and property 1 of the previous section, we have H(p) � H(q) if and only if
hence, if and only if p t q. Therefore, H is a utility function that represents t. This is not good enough, however. We want to show that there exists a function u defined on Z such that
H(p)
=
L: u(z)p(z).
• ez
We will achieve this by construction. First, repeated use of property 5 of the previous section implies that, for all p, q E P and a E [0, 1], •
ap + (1 a)q a{H(p)P... + (1 H(p))P .,] + (1 a)[H(q)P... + (1 H(q))P .,] (aH( p) + (1 a)H(q))P... + (1 aH(p) (1 a)H(q))P •
(1.10.1)
• .,.
Note that by the definition of H we know that H(p) and H(q) are greater than 0 and less than 1. Hence, aH(p)+(1a)H(q) is between 0 and 1 and the righthand side of the second indifference relation of (1.10.1) is a welldefined compound lottery. Since H represents t, it follows that
H(ap + (1  a)q) = H((aH(p) + (1 a )H(q))P... + (1 aH(p)  (1  a)H(q))P.,.) aH(p) + (1 a)H(q), =
where the second equality follows again from the definition of H. We thus conclude that H must be linear in that
H(ap + (1  a)q) = aH(p) + (1 a)H(q).
11
Preferences Represent ation and Risk Aversion
Next we define a function
u
on Z by
u(z) =: H(Pz) Vz E Z.
(1.10.2)
We claim that this function is a von NeumannMorgenstern utility functio n. Before we prove our claim, we remark that the above defi nition of u is a natural one. The von NeumannMorgenstern utility function u is a function on sure things. Thus we define u(z) to be the utility, according to H, for the sure consumption plan P.. Here is the proof for our claim. Let p E P. It is easily seen that
Since H represents
t,
zEZ
=
l:p(z)H(Pz) zEZ
= l:p( z)u(z), zEZ
where the second equality follows from the repeated use of the lin earity of H, and the third equality follows from the definition of u. Thus any binary relation on P satisfying the three axioms has an ex pected utility representation. Finally, it can be shown that u is only determined up to a strictly positive linear transformation in that if u is also a von NeumannMorgenstern utility function, then there exist two constants c > 0 and d such that u = cu + d. You will be asked to provide a proof of this in Exercise 1.4. Conversely, it is easily verified that if a binary relation t has an exp ected utility representation in that there exists u such that, for
p,q EP,
ptq
if and only if
L u(z)p(z)
zEZ
�
L u(z)q(z) , zEZ
then t satisfies the three axioms of Section 1.8. We leave the proof of this assertion to the readers in Exercise 1.4.
12
Foundations for Fin ancial Economics
1. 11. We proved the expected utility representation theorem for the case where Z is a finite set. When Z is an infinite set, for example when Z contains all the positive real numbers, the representation theorem is no longer true. We need a fourth axiom called the sure thing principle. In words, it basically says that if the consumption plan p is concentrated on a set BE Z such that every point in B is at least as good as q, then p must be at least as good as q. With this fourth axiom as well as some technical conditions, we have a representation theorem for general Z. We refer interested readers to Fishburn ( 1970 ) for details.
When consumption occurs at more than one date, say at times t = 0, 1, . . . , T, our previous discussion can be generalized in a straightforward manner. Let Z be a collection of T+1tuples z = (zo, .. . , %T), where ze denotes the number of units of consumption at time t for sure. Suppose that Z is a finite set. A probability p on Z is a mapping with the following properties: 1. 12.
1. p (.z) E [0, 1]
2. E.e_. p(.z)
Vz E Z;
and
1. An individual expresses his preferences over the probabilities defined on Z, or, equivalently, an individual expresses his preferences on lot teries whose prizes are consumption at times t = 0, 1, ... , T. Denote the collection of probabilities on Z by P. Now mimic the analysis of Sections 1. 8 to 1.10 to conclude that the binary relation� is a prefer ence relation satisfying the Substitution axiom and the Archimedean axiom if and only if there exists a von NeumannMorgenstern utility function u on sure things such that for all p, q E P, p � q if and only if =
L u(zo, ... , %r)p(z=zo, ..., ZT) � L u(zo, ..., zT)q(z=zo, ... , ZT),
•EZ
•EZ
where p(.z = zo, ..., ZT) is the probability that consumption from time 0 to time T is (.zo, . . . , .zT). For tractability, most analysis of the equilibrium valuation of risky assets in later chapters uses von NeumannMorgenstern utility functions that are timeaddititJe. That is, there exist functions ue (·)
Preferences Representation and Risk A version
such that
T
u(zo, ... , ZT ) = L uc(zc). t=O
18
(1.12.1}
This is a rather strong assumption. It basically says that what an individual consumes at one time will not have any effect on his desire to consume at any other time . For example , it says that having a big lunch will not affect one's appetite for a seven course dinner. So, readers should be cautioned to note that results reported there are colored by this timeadditivity assumption. 1.13. One implication of the expected utility theory is that a von NeumannMorgenstern utility function is necessarily bounded when probability distributions of consumption involve unbounded consumption levels. This is a consequence of the Archimedean ax iom. To see this, suppose that u is unbounded . Without loss of generality, let u be unbounded from above, and let Z contain all positive levels of consumption. Then there exists a sequence of con sumption levels {z.a}:":1 such that z11 + oo and u(z11) � 2". Now consider a consumption plan p such that p(zn) = 21,., n = 1, 2, . . .. This consumption plan has unbounded consumption levels. The ex pected utility of this consumption plan is
Now let q, r E P be such that p > q > r. We know immediately that the expected utilities associated with q and r must be finite. It is then easily seen that the Archimedean axiom can not be satisfied . The boundedness of a utility function is somewhat discomfort ing since many utility functions used in economic applications are unbounded . For example, the log utility function is unbounded from above and from below. Any power function, u(z) = z16, is bounded from below and unbounded from above if b < 1 and bounded from above and unbounded from below if b > 1. There are, however, ways to get around this kind of difficulty. For example, if we only consider consumption plans that concentrate on a finite number of consump tion levels, then the above problem certainly will not arise. This will
14
Foundations for Fin ancial Economics
be the case in many of our discussions in the subsequent chapters. There we usually take the state space n to have finite elements. Then a consumption plan naturally takes a finite number of consumption levels . The number of consumption levels is at most equal to the number of states of nature! In such an event, the consumption levels are certainly bounded. Another possible resolution when Z contains all the positive consumption levels is to use a preference relation that exhibits ris/c aversion and consider consumption plans that have finite expecta tions . We will see in later sections of this chapter that risk aversion in this case implies that u is concave. Being a concave function, u is differentiable at some point, say b > 0. It then follows that
u(z) � u(b) + u '(b) (z b)
Vz,
(1.13.1)
where u'(b) denotes the derivative of u at b. Now let z be a consump tion plan having a finite expectation. Then the expected utility of z is
E[u(z) ] � u(b) + u'(b) (E[z] b) < co, where we have used (1.13.1) . That is, if u is concave,
expected util ities �sumption plans having finite expectations will be finite even when u is unbounded! Thus in applications, we can comfort ably use unbounded utility functions as long as they are concave and random consumption plans considered all have finite expectations. When Z is composed of units of consumption at more than one date, the above analysis applies easily to the timeadditive utility func tions of (1.12.1) when u, is concave and the random consumption at each time t has a finite expectation. 1.14. Among the three ax1oms of Section 1.8, the one that is often violated in empirical experiments is the Substitution or In dependence axiom. The best known example of this is the Allais Paradox. Consider the two pairs of lotteries in Figure 1.14.1 . Lot tery pt gives $1 million for sure. Lottery pz pays $5 million with 0.1 probability, $1 million with 0.89 probability, and $0 with 0.01 probability. Lotteries Ps and P4 of the second pair are interpreted similarly.
16
Preferences Represen tation and Risk A version
$ $
$5m
1m
/ �$0
5m
0.11
$1m
~ $0
0
$
Figure 1.14 .1: The Allais Paradox
Most individuals choose lottery Pl over P2 when faced with the first pair of lotteries. They prefer $1 million for sure to a high prob• ability of getting $1 million coupled with low probabilities of getting $5 million and $0. On the other hand, most individuals choose Ps over p,. when faced with the second pair of the lotteries. This behav ior, however, is a violation of the S ubstitution axiom. To see this, we first note that, Pl
P2
,._
,._
0.11($1m)+ 0.89($1m)
(
(1.14.1)
)
1 10 0.11 ($Om)+ ($5m) + 0.89($1m). 11 11
Transitivity implies that

0.11 ($1m)+ 0.89($1m)
We claim that
Pl
(
(1.14.2)
)
>
1 10 0.11 ($Om)+ ($5m) + 0.89($1m). 11 11 (1.14.3)
>
10 1 ($Om)+ ($5m). 11 11
(1.14.4)
16
Foundations for Fin ancial Economics
Suppose otherwise. That is, 10 1 ($Om) ($5m) + 11 11
t
Pl·
(1.14.5)
Then the Substitution axiom implies a contradiction to (1.14.3) . By the Substitution axiom, (1.14.4) implies that 0.11($1m) + 0.89($0m)
This is equivalent to
>
0.11
( 1 ($Om) + 10 ($5m)) + 0.89($0m).
P4
11 >
11
Ps,
a contradiction to the experimental results. Thus the behavior of choosing P1 over P2 and choosing Ps over P4 is inconsistent with the Substitution axiom. The experimental violation of the Substitution ax1om makes one cautious in using expected utility analysis for descriptive purposes, as we will do in later chapters in drawing conclusions about relations among equilibrium prices of risky assets by assuming that htdividuals are expected utility maximizers. The ultimate test of the •reasonableness" of the descriptive conclusions, however, is whether the descriptive conclusions, to the extent that they are empirically testable , conform with the observable data. This is a subject to be discussed in Chapter 10. The expected utility analysis can also be defended at another level by drawing on the work of Machina (1982). It is shown there that the basic concepts, tools, and results of expected utility analysis can be carried over to cases where the Substitution axiom is violated provided that the following condition is satisfied: An individual's preferences can be represented by a utility function H as discussed in Sections 1.4 and 1.5 and H is differentiable in a certain sense. Details of Machina's work, to which we refer interested readers, will not be discussed here. 1.15.
1.16. Throughout the rest of this chapter, we will assume that individuals are expected utility maximizers. For brevity, whenever

Preferences Representation and Risk A version
17
we say a utility function, we mean a von NeumannMorgenstern utility function unless otherwise specified. As we are dealing with economies under uncertainty, it is im portant to characterize an individual's behavior when he is facing risk. The following section gives necessary and sufficient conditions for a utility function to exhibit behaviors that exhibit risk aversion behavior. Moreover, in order to discuss the comparative statics of an individual's behavior when his attitude towards risk changes, we need to have measures of risk aversion. These are the subjects to which we now turn. 1.17. An individual is said to be risk averse if he is unwilling to accept or is indifferent to any actuarially fair gamble. An individual ·:: is said to be strictly risk averse if he is unwilling to accept any actu '''arially fair gamble. Consider the gamble that has a positive return, :: ·111, with probability p and a negative return, h2, with probability · ·
(1 p).
The gamble is actuarially fair when its expected payoff is zero,
(1.17.1) Let u(·) be the utility function of an individual. From the definition of (strict) risk aversion, we have:
u(Wo}(>) � p u(Wo + h1) + (1 p) u(Wo + h2) , i. where Wo denotes the individual's initial wealth. ?
Using the definition of a fair gamble as in relation , lation (1.17.2 } may be rewritten as
u(p(Wo+h1}+ (1p}(Wo+h2)) (>)
�
(1.17.2) (1.17.1), re
p u(Wo+ht)+ (1p) u(Wo+ h2).
The above relations demonstrate that risk aversion implies a concave utility function and that strict risk aversion implies a strictly concave utility function. A reversal of the above steps demonstrates that a concave utility function implies risk aversion and that a strictly concave utility function implies strict risk aversion. Figure 1.17 .1 illustrates the fact that the expected utility of gamble is strictly less than the utility of its expected payoff if and only if utility function is strictly concave.
18
Foundations for Financial Economics
utw0+h1) u tw0l
pu(Wo+h1H(Ip) utw0+h2l
u(w0+h2)
Figure 1.17.1: A
Concave Utility Function
1.18. Consider a portfolio choice problem of a risk averse in dividual who strictly prefers more to less (has a strictly increasing utility function). If the individual invests a; dollars in the jth risky asset and (Wo  E; a;) dollars in the risk free asset, his uncertain end of period wealth; W, would be ;
;
or equivalently
W
=
W0(1 + rJ) + 2:a;(r; r,), ;
where
Wo
=
initial wealth, r1 = the riskless interest rate, r; = the random rate of return on the jth risky asset, a; = the dollar investment in the jth asset.
19
Preferences Representation and Risk Aversion
Thus the individual's choice problem
is
E[u(Wo(1 + r/) + I: a,.(ii r/))].
max {a;}
j
(1.18.1)
We assume that there exists a solution to (1.18.1). Since u is 1:oncave, the first order necessary conditions are also sufficient. They are
;Jiote that, since
u
E[u'(W)(ii r,)]
=
o
Vj.
( 1.18.2)
is strictly increasing, (1.18.2) implies that the
' )robability that ii  r1 > 0 must lie in ( 0, 1). 1.19.
An individual who is risk averse and who strictly prefers .re to less will undertake risky investments if and only if the rate lf:return on at least one risky asset exceeds the riskfree interest pte:·. To see this, we note the follo�ng. For the individ�al to in�est _ or even short sell the riSky assets as an opt1mal cho1ce, ...hiDg m it la necessary that the first order conditions evaluated at no risky l;avestments be nonpositive: lquivalently,
•. assumption the individual strictly prefers more to less, therefore, t{·) > 0, and the above condition is equivalent to
ai �
0Vj
only
if
E[ii  r1] � 0
V j.
Thus an individual with a strictly increasing and concave utility func
·tion
will avoid any positive risky investment only if none of the risky usets has a strictly positive risk premium. When one or more risky usets have strictly positive risk premiums, the individual will take part in some risky investments. That is 3
j, such that
a,. > 0
if
3
j' such that
E[ri'  rtl > 0.
(1.19.1)
20
Foundations for Financial Economics
Note that j' may not equal j, because when there is more than one risky asset, E[r;  r/ ] > 0 does not necessarily imply a; > 0. However, when there is only one risky asset , condition (1.19.1) indicates that a positive risk premium implies positive investment in that asset . 1.20.
Henceforth until specified otherwise, consider a strictly risk averse individual who prefers more to less in an economy where there is only one risky asset and one riskless asset and where the risk premium of the risky asset is strictly positive. ( Here we note that this individual will strictly prefer more to less. ) Note that Sec tion 1.19 demonstrates that a strictly positive risk premium waul� induce a strictly positive risky investment . This section examines the minimum risk premium that is required to induce the individual to invest all of his wealth in the risky asset . Let i and a denote the rate of return of the single risky asset and the amount invested in it, respectively. For an individual to invest all his wealth in the risky asset it must be that
E[u'(Wo(1 + i))(i r1) ] � 0. Taking a first order Taylor series expansion of u'[Wo(1 + i) ] around u'[Wo(1 + r,) ], multiplying both sides by the risk premium, and taking expectations gives
E[u'(Wo(1 + i))(i r,)] = u'(W o(1 + r1))E[i r1l + u"(Wo(1 + r1))E[(i r1)2 ] W0 + o(E[(i r1)2]), where o(E[(ir/)2 ]) denotes terms of smaller magnitude than
r/)2 ] .
E[(i
Risk is said to be small when E[(i r1)2] is small and terms involving E[(i  r,)s] and higher orders can be ignored . Ignoring the remainder term, the minimum risk premium required to induce full investment in the risky asset may be determined by setting the righthand side of the above relation to 0 and getting
(1.2 0.1)
21
Preferences Represen tation and Risk Aversion
where RA(·) =  u" ( · )/ u' ( · ) is the measure of absolute risk aversion defined by Arrow (1970) and Pratt (1964) . Note that for small risks, absolute risk aversion is a measure of the intensity of an individual's aversion to risk. From (1.20.1), the higher an individual's absolute risk aversion, the higher the minimum risk premium required to in duce full investment in the risky asset. Intuitively, the curvature of an individual's utility function would be related to the minimum risk premium required to induce full investment in the risky asset. The absolute risk aversion is a measure of the curvature of an individ ual's utility function. Note that von NeumannMorgenstern utility functions are only unique up to a strictly positive affine transforma tion ( addition of a constant and multiplication by a strictly positive scalar ) . Therefore, the second derivative alone cannot be used to characterize the intensity of risk averse behavior. The ArrowPratt measure of absolute risk aversion is invariant to a strictly positive affine transformation of the individual's utility function. 1.21.
The characteristics of an individual's absolute risk aver sion allow us to determine whether he treats a risky asset as a normal gOod when choosing between a single risky asset and a riskless asset. An individual's utility function displays decreasing absolute risk aversion when RA(·) is a strictly decreasing function. Similarly, dRA(z)fdz = 0 Vz implies increasing absolute risk aversion, and a constant dRA(z)fdz Vz implies constant absolute risk aversion. Obviously, a single utility function may display more than one of the above characteristics over different parts of its domain. Several interesting behavioral properties of utility functions that display the same sign of dRA(z)fdz over the entire domain of RA(·) were derived by Arrow ( 1970) . Arrow showed that decreasing absolute risk aver sion over the entire domain of RA (·) implies that the risky asset is a normal good; i.e., the (dollar) demand for the risky asset increases as the individual's wealth increases. Increasing absolute risk aversion implies that the risky asset is an inferior good, and constant absolute risk aversion implies that the individual's demand for the risky asset is invariant with respect to his initial wealth. That is dRA(z) dz
da 0 dWo >
V
Wo;
22
Foundations for Fin ancial Economics
dRA(z) 0 u > v z � dao 0 VWo; dz dW < dRA(z) da =0 VWo. = 0 V z � dz dWo The proof for only the decreasing absolute risk aversion case is presented as the proofs for the other two cases follow the same structure. At an optimum, we have:
E[u'(Wo(1 + 'I) + a(i r1))(i r1)]
=
0.
The change in the individual's optimal risky asset investment with respect to a change in initial wealth can be determined by:
1. Differentiating the first order condition for an optimum (that is determined for a given initial wealth level ) with respect to his or her initial wealth; 2. Setting the first derivative equal to zero to move along the indi vidual's optimal portfolio path; and 3. Solving for the implicit relationship between the change in risky asset investment and the change in initial wealth that would move the individual's risky asset investment along an optimal path as his initial wealth changes. This process is referred to as implicit differentiation of a with respect to Wo and gives
da _ E[u"(W )(i rt)](1 + rt)
dWo
E[u"(W)(i r1)2]
(1.21.1)
where as usual W = W0(1 + rt) + a(i rt) denotes the individual's end of period (random) wealth . The denominator is positive because strict risk aversion implies u"(W) < 0; therefore, sign (dafdWo) =sign {E[u"(W)(i r,)]}. Under decreasing absolute risk aversion, in the event that i � r f, we have W � Wo(1+rt) since the amount invested in the risky asset is strictly positive . Thus
(1.21.2a)
Preferences Representation and Risk A version
Similarly, in the event that i < r1, we have W
r1 must lie in (0, 1) for an optimal solution to exist. Substituting the first order condition E[u'(W )(i r,)] = 0 into (1.21.4) gives the desired result. Note that the property of nonincreasing absolute risk aversion implies that the third derivative of the individual's utility function is strictly positive:
dRA(z) dz
_ 
u
"'
( z ) u' ( z ) + [u"(z)]2 $ 0 � u > O, [u'(z)]2
,
(1.21.5)
where we have used u"' to denote the third derivative of u. This follows because u'(z) > 0 and [u"(z)]2 > 0. Thus dRA(z)/dz $ 0 implies u"'(z) > 0. 1.22. The property of decreasing absolute risk aversion is re lated to the dollar demand for the risky asset. Thus, an individual having a utility function displaying decreasing absolute risk aversion may actually increase , hold constant, or decrease the proportion of his wealth invested in the risky asse t as his wealth increases. The ArrowPratt measure of relative risk aversion is RR(z) = R A(z)z. Under increasing relative risk aversion, that is, when dRR(z)/dz > 0
Foundations for Fin ancial Economics
24
the wealth elasticity of the individual's demand for the risky asset is strictly less than unity. That is, the proportion of the individual's initial wealth invested in the risky asset will decline as his wealth increases. Under constant relative risk aversion, dRR (z)/dz = 0 Vz, the wealth elasticity of demand for the risky asset would be unity, and under decreasing relative risk aversion, dRR(z)jdz < 0 Vz, the wealth elasticity of the demand for the risky asset would be strictly greater th an unity. The wealth elasticity of the demand for the risky asset , '7 1 may be expressed as
Vz,
da Wo dWo a
'7 = _
=
1+
(da/dWo)Wo  a . a
(1 .22 . 1)
Substituting the righthand side of (1.21.1) for (da/dWo) into the righthand side of relation ( 1 .22.1) and rearranging terms gives
Note that the numerator of the second term on the righthand side of the above relation may be expressed as
E[ u" ( W ) (W0 ( 1 + r 1 ) + a( r  r 1 )) (r  r1 )] = E[ u"(W)W ( r  r1 )] . Thus, the elasticity coefficient may be expressed as f7
= 1+
E[u"(W) W ( r  r 1)]  a E[u" (W) ( r  r 1 )2 ] 
.
(1 .22.2)
The denominator of the second term on the righthand side of (1 .22.2) is positive because u"(z) < 0 . Whether the demand elasticity for the risky asset is elastic, '7 > 1 , unitary elastic, '7 = 1, or inelastic , '7 < 1, depends on the sign o f the numerator. That is sign ( '7  1) = sign { E[ u" (W )W (r  rt)J l . Under increasing relative risk aversion, in the event in the event
;: � rJ, ;: < 'I ·
Preferences Represen tation and Risk A version
25
Multiplying both sides of the above relations by u' (W) (i  r , ) re duces the relations to
( 1 .22.3a )
in the event that i � r b and
( 1 . 22.36)
in the event that i < r1. Relations ( 1.22.3a) , ( 1 .22 .3b ) , and the fact ' that at a portfolio optimum E[u'(W) (i  r1 )] = 0 give
E[u"(W)W(i  rt)J
0, B > O, B 1 1 _1 . u ' (z) = z  l , u " (z ) =  z _ .!. a B Note the following: 1 1 RA (z) = z  1 , RR (z) = B B and dRA (z) dRR(z) 1 =  z _2 dz B < 0, dz = 0 . Thus, for an individual having a narrow power utility function the proportion of wealth invested in the risky asset is invariant with re spect to changes in his initial wealth level. Finally, consider an e:ztended power utility function: ( ) 1 A u z = ( A + Bz) 1 _ .!. s , B > O, A "f. O, z > max!  , O) , BI B u' (z) = (A + Bz)  l , u" (z) =  (A + Bzt i  1 .
28
Foundations for Financial Economics
Direct computation yields
RA(z ) =
1
A + Bz ' dRA (z) B dz  (A + Bz ) 2 z RR(z) A + Bz ' =
< O,
{ �� �
if and only if A > 0; if and only if A = 0; 0 if and only if A < 0. Extended power utility functions exhibit decreasing or increas ing relative risk aversion depending upon whether A is negative or positive. Note that when A = 0 an extended power reduces to a narrow power utility function. Finally, we remark that the exponent of a (extended) power function is equal to the coefficient of z in the above examples. This is just for convenience in writing the derivatives. They do not have to be equal.
A dRR(z) dz  (A + Bz)2 _
1 .24 . The analysis in Section 1 .20 is essentially a local analysis. When risk is small, the higher the individual 's absolute risk aversion at Wo ( 1 + r1 ) , the higher the risk premium which is required for him to invest all his wealth in the risky asset. Thus, in a sense, RA ( ) is a local measure of risk aversion. Pratt ( 1964) showed that RA ( ·) is also a measure of risk aversion in a global sense. That is, if there are two individuals i and k with � ( z) � Ri (z) 'V z , then individual i will pay a larger insurance premium than k to insure against a (not necessarily small) random loss. Individual i is then said to be more risk auerse than individual k. In the framework of Section 1. 20, it will be demonstrated below that if individual i is more risk averse than individual k and they are endowed with the same initial wealth, then the risk premium required for individual i to invest all his wealth in the risky asset is larger than that required by individual k. The above statement is valid in a global sense where risk does not h ave to be small. Suppose that ·
E[u � (Wo ( 1 + i )) (i  r1 )]
=
0,
( 1 .24.1)
Preferences Represen tation and Risk A version
29
where u�:(·) denotes individual k 's utility function. Relation (1.24.1) is a necessary and sufficient condition for E[i rJ] to be the minimum risk premium that induces individual k to put all his wealth into the risk y asset. H we can show that
(1.24.2) where u, ( ) is individual i's utility function, then we are done. Re lation (1.24.2) implies that individual i's optimal investment in the rlsky asset is less than or equal to Wo . This in turn implies that ihe required risk premium for i to invest all his wealth into the risky Asset is greater than that required by k . 1. 25. We will first prove a useful relation between u, and U J:: . We claim that there exists a strictly increasing and concave function G such that
(1.25.1)
and only i f � (z) � R� (z) , V z . We will prove the necessity part first. Since u�: is strictly increasing, we define G (y) = u, ( uk" 1 (y)), where u;1 denotes the inverse of u�: . Substituting y = u �: (z) into the definition, we immediately have (1.25.1). We still have to show that G ( ·) is strictly increasing and concave. Differentiating (1.25.1) once gives if
(1.25.2)
By the fact that uH z) > 0 and u� (z) > 0, G must be strictly increas ing in the domain of its definition . Now differentiating (1.25.2) with respect to z gives
(1.25.3) Dividing (1.25.2) into (1.25.3) gives i
RA (z )
_
 
G"(u�: (z))
,
G ' (u�: (z )) u�:(z)
1: + RA(z) .
(1.25.4)
By the hypothesis that .RiA ( z) � R�(z) Vz, (1.25.4) implies that G is concave. This completes the proof for the necessity part .
30
Foundations for FinlllJ cial Economics
Next suppose that there exists a strictly increasing and con cave function G such that ( 1 .25.1) holds. By differentiation we have ( 1.25.2 ) and ( 1.25.3 ) . Dividing ( 1 .25.2 ) into ( 1 .25.3 ) gives ( 1 .25.4 ) . Since G is strictly increasing and concave, we immediately conclude that R�(z) � R�(z), Vz, which concludes the proof for the sufficiency part. 1.26 .
Now we are ready to show that if Ui = G(uk) , the mini mum risk premium required for i to invest all his wealth in the risky asset is higher than that for k. We will first fix some notation . Let X and Y be two random variables. We use E[X ! Y � 0] to denote the conditional expectation of X given that Y � 0; similarly for E[ X! Y < 0] . Then we have by definition,
E[ X] = E[ X l Y
�
O] P( Y
�
o ) + E[XJ Y
O, C < 0, and z � max [O,  (A/B)] , or A > 0, B < O, C > 0 and 0 � z <  (A/ B) for ( 1 .27. 1) and where A > 0, B < 0 and z � 0 for (1.27.2). The conditions on A, B and C are to insure that the utility functions are strictly concave and increasing. The proof for necessity of (1 .27.1) or (1 . 27.2) is tedious and we refer interested readers to Cass and Stiglitz 's treatment. In what follows, we shall demonstrate the sufficiency of ( 1 .27. 1) or (1 .27.2) .
32
Foundations for Financial Economics
1.28 . Let a denote the proportion of initial wealth, Wo, invested in the riskless asse t, and let b; denote the proportion of the remainder Wo(1  a) invested in the jth risky asset . A risk averse individual solves the following programming problem:
max
{a,b;,.\}
where
l
e
= E[ u(W) ] + .\( 1  L b; ) , .
'
denotes the Lagrangian, and
W
=
Wo(1 + ar1 + (1  a) L bi i;). i
The necessary and sufficient conditions for an interior optimum are:
E[ u' (W )Wo( r,  L bi r;)] i E[u' (W)Wo(1  a)r; ] and
=
.\
=
(1.28.1)
0,
v
( 1 .28.2)
j,
L:: 6; 1. i (1.28.1) into (1.28.2) and using (1.28.3) =
Substituting
E[u' (W0 ( 1 + r1 + (1  a) L: bz ( iz  r,)))(r;  r,)] l
Now suppose that
(1 .27.1 ) is true.
Then
(1.28.4)
(1.28.3) gives, Vj, =
0 . ( 1 .28. 4)
implies, Vj,
E[ (A + BW0 (1 +r1 + (1 . a) L b; ( i;  r1 ))) c ( i;  r1) ] = 0. (1.28. 5) i Consider the same individual with a different initial wealth WJ. Let ' a , bj be the individual's optimal portfolio decisions with respect to W�. Note that, by the strict concavity of the utility function, a' and b' are uniquely determined . We must have, Vj,
;
33
Preferences Representation and Rid A version
We claim that
0!1
b; and « , bj are related by
W.0' (1  Clt' )b1'· = A + BWJ (1 + rt) Wi0 (1  C!t )b1· A + BWo(1 + rt)
v
j.
(1 .28.7)
To see this, we substitute (1 .28.7) into either (1 .28.5 ) or ( 1.28.6 ) and check that the equality is satisfied. Summing (1.28.7) over all j and using (1.28.3 ) we get
(1.28.8) Now substituting
( 1.28.8 ) into ( 1 .28.7) , we have bj = b;
v
j,
which was to be shown . Along the same line of argument, it is easily checked that (1.27.2) also implies two fund monetary separation. 1.29.
The solutions of
functions:
( 1.27.1 )
include extended power utility
1 u ( z)  (C + 1 (A + Bz ) C+ l , )B
C ;t  1
and log utility functions:
u( z ) The solutions of
(1 .27 .2)
=
ln(A + Bz ) .
are exponential utility functions:
u( z) =
A exp
B
{ Bz} .
Narrow power utility functions (A = 0) and log utility func tions exhibit constant relative risk aversion. Thus, the proportions invested in the riskless asset Cit and in the risky asset mutual fund (1  Cit ) are invariant to different levels of initial wealth .
34
Foundations for Fin an cial Economics
The exponential utility functions exhibit increasing relative risk aversion. Thus as initial wealth increases , the proportion invested in the risky asset mutual fund will decrease. Other comparative statics can be derived. Their analysis is a direct corollary of the comparative statics in the two asset case, there fore we leave them for interested readers.
Exercises 1.1. Suppose that X has a countable number of elements. Show that a binary relation on X is a preference relation if and only if it can be represented as a realvalued function. 1.2. Suppose that !: is a binary relation. A subset Y of X is said to be a !: order dense subset of X if for all x, x1 E X such that x � x', there exists y E Y such that x !: y !: x' . In words , given any two consumption plans that are in the binary relation !: and can be strictly compared, there always exists an element of Y that lies between the two. The set Y is a countable !: o rder dense subset of X if it is a !:order dense subset of X and has a countable number of elements. A representation theorem for arbitrary X is as follows: The binary relation !: can be represented by a utility function H if and only if !: is a preference relation and there is a countable !: order dense subset Y of X . Now let X = [0, 1 ] x [0, 1 ] and define (xi , x2) !: (y� , ':12) i f x 1 > ':11 or if x1 = ':11 and x2 > ':/2 . 1.2. 1 . Verify that !: is a preference relation. ( This is the socalled Lexicographic preference relation . ) 1 .2.2. Show that if (x1 , x2) !: (yi , ':/2) and (yi , ':/2 ) !: (x1 , x2) then X 1 = ':/1 and X2 = ':/2 · 1 .2.3. Demonstrate that !: cannot be represented by a realvalued function . ( Hint : Show that there does not exist a counter !:order dense set and use the representation theorem above for arbitrary X .) 1 .3. Prove the properties of Section 1 .9.
Preferences Represen tation and Risk A version
35
1. 4. Show that a von NeumannMorgenstern utility function is de termined up to a strictly positive linear transformation . Show also that if a binary relation � has an expected utility represen tation in that there exists u such that, for p, q E P, p
�
q
if and only if
L u(z)p(z) � L u(z)q(z) ,
zEZ
zEZ
then � satisfies the three axioms of Section 1 .8 . 0 0 (z)/dz > implies that da/dWo < YWo and 1 .5. Prove that dRA 0 0 dRA (z)/dz = implies that da/ dWo = 'v'Wo in the context of Section 1.21. 0 1.6. Show that dRR (z)/dz < 'v'z E R implies 'I > 1 and that 0 dRR (z)/dz = 'v'z E R implies 'I = 1, where 'I is the wealth elasticity of demand for the risky asset defined in ( 1.22. 1 ) . 1.7. Define absolute risk tolerance to be the inverse of the Arrow Pratt measure of absolute risk aversion. Show that solutions to ( 1.27 .1) and (1.27.2 ) all exhibit linear absolute risk tolerance. 1.8. Fix an individual with an increasing and strictly concave util ity function u and consider the gamble of ( 1.17 .1) . Define the insurance premium z to be the maximum amount of money the individual is willing to pay to avoid the gamble. That is, z is the solution to the following u(Wo  z)
=
pu(Wo + h 1 ) + ( 1  p)u(Wo + h2 ) .
Obviously, z depends upon the initial wealth , Wo , and we will denote this dependence by z(Wo ) . S how that, when the risk is small, dRA(z) dz(Wo) 0 . O 'v' z if 'v' �0 ' < < dz dWo dRA(z) dz(Wo) = 0 = 0 'v'z if 'v' Wo ; dz dWo dRA (z) dz(Wo) 0 > O 'v'z if > 'v' �O· dz dWo 1.9.
Show that utility functions of ( 1.27 .2 ) imply two fund monetary separation.
36
Foun dations for Fin ancial Economics
Remarks. Discussions in Sections 1 .3 to 1 . 15 are adapted from Kreps ( 1981 ) . For discussions of a continuous preference relation and its representation by a continuous realvalued function, see Debreu ( 1964 ) . For representation of a preference relation by a continuous von NeumannMorgenstern utility function, see Granmont ( 1972 ) . The example of the unbounded expected utility of Section 1 . 1 3 is a generalization of the socalled St. Petersburg Paradoz . Arrow ( 1970) is an excellent source for a discussion of this and other related issues. Nielsen ( 1985, 1986) develops a set of axioms under which a preference relation can be represented by a possibly unbounded expected utility function . Discussions of Sections 1 . 17 through 1.22 are freely borrowed from Arrow ( 1970 ) . Some of the results of Sections 1.24 and 1 .25 are contained in Pratt ( 1964 ) . The sufficiency proof of Section 1 .28 is adapted from Rubinstein ( 1 974 ) . We will discuss a stronger measure of risk aversion in Chapter 2 that gives some intuitively appealing comparative statics in a port folio problem with two risky assets. Readers interested in different measures of risk aversion should see Pratt and Zeckhauser ( 1987 ) . Machina ( 1982 ) discusses measures of risk aversion when an individ ual's preferences can be represented by expected utility only locally. For some negative results on the comparative statics for portfolio choice problems when there are many assets see Hart ( 1975 ) . References
Allais, M. 1953 . Le comportement de l'homme rationnel devant le risque, critique des postulates et axioms de l'ecole Americaine. Econometrica 21:503546. Arrow, K. 1970. Essays in the Theory of RiskBearing. Amsterdam: NorthHolland. Bernoulli, D. 1954. Exposition of a new theory on the measurement of risk ( translation from the 1 730 version ) . Econometrica 22:2336.
Preferences Representation and Risk Aversion
37
Cass, D., and J . Stiglitz. 1970, The structure of investor prefer ences and asset returns, and separability in portfolio allocation: A contribution to the pure theory of mutual funds. Journal of Economic Th eory 2 : 1 22 160.
Debreu, G . 1954. Representation of a preference ordering by a nu merical function. In Decision Processes. Edited by R. Thrall, C . Coombs, and R. Davis. John Wiley & Sons, New York . Debreu , G . 1964. Continuity properties of paretian utility. In terna tional Economic Review 5 :285293. Fishburn, P. 1970. Utility Th eory for Decision Making . John Wiley & Sons. New York. Granmont , J . 1972. Continuity properties of a von Neumann Mor genstern utility. Journal of Economic Theory 4:4557. Hart, 0. 1975. Some negative results on the existence of comparative statics results in portfolio theory. Review of Economic Studies 42:61562 1 . Kreps, D. 1981 . Single Person Decision Theory. Lecture Notes. Graduate School of Business, Stanford University. Stanford , Cal ifornia. Machina, M. 1 98 2. Expected utility analysis without the indepen dence axiom. Econ ometrica 50:277323. Nielsen, L . 1985. Unbounded expected utility and continuity. Math ematical Social Sciences 8 : 2012 16 . Nielsen, L. 1986. Corrigendum. Mimeo. Economics Department, University of Texas at Austin. Pratt, J. 1964. Risk aversion in the small and in the large. Econo metrica 32:122136. Pratt, J ., and R. Zeckhauser. 1987 . Proper risk aversion. Econo metrica 5 5 : 143154. Rubinstein , M. 1974 . An aggregation theorem for securities markets . Journal of Fin ancial Economics 1 :225 244. Savage, L. 1972 . Foundations o f Statistics. Dover. New York. Von Neumann, J . , and 0. Morgenstern. 1953. Theory of Games and Economic Behavior, Princeton University Press. Princeton , New Jersey.
CHAP TER 2
S TOCHAS TIC DOMINANCE
2. 1. In Chapter 1, we discussed the relationship between the risk premium and ArrowPratt measures of risk aversion in a context ' where there are only two securities , one riskless and one risky. The first part of thiS chapter addresses the following question: Suppose that there are two risky securities. Under what conditions can we unambiguously say that an individual will prefer one risky asset to another when the only information we have about this individual is either that he is nonsatiable or that he is risk averse? To answer this question , we will introduce two concepts of stochastic dominance. One concept turns out to be useful in comparing the riskiness of risky asse ts. This concept, however, does not allow us to compare any two risky assets, and in the terminology of Section 1.3, it does not define a complete order among risky ass e ts. We will analyze the comparative statics of an individual's problem with a portfolio of one risky and one riskless asset when the riskiness of the risky asset increases. These comparative statics depends upon Arrow Pratt measures of risk aversion in a complicated way. We will �lso demonstrate through an example that the Arrow Pratt measures of risk aversion are too weak to study the comparative statics of
40
Foundations for Fin ancial Economics
a.n individual's portfolio problem when faced with two risky assets . A stronger measure of risk aversion that gives intuitively appealing comparative statics is then discussed . 2.2. We will sa.y tha.t risky asset A dominates risky asset B in the sense of First Degree Stochastic Dominan ce , denoted by A r�D B, if a.ll individuals having utility functions in wealth tha.t a.re increasing a.nd continuous either prefer A to B or a.re indifferent between A a.nd B. Intuition suggests tha.t if the probability of asset A's rate of return exceeding any given level is not smaller tha.n tha.t of asset B's rate of return exceeding the same level, then any nonsa.tia.ble individual will prefer A to B . It turns out tha.t this condition is not only sufficient but also necessary. Here we note tha.t a.n individual is nonsa.tia.ble if a.nd only if his utility function is strictly increasing. For ease of exposition, we shall assume tha.t the rates of return on A a.nd B lie in the interval [0, 1] throughout this chapter . However, the results to be shown a.re valid in more general contexts when some regularity conditions a.re satisfied . Let FA (·) a.nd FB (·) denote cumulative distribution functions of the rates of return on A a.nd B respectively. Suppose tha.t:
V z E [0, 1] .
(2.2.1)
Since FA (·) a.nd FB (·) a.re cumulative distribution functions, they a.re continuous from the right a.nd FA (1) = FB (1) = 1 . However, FA (O) ma.y not be equal to FB (O) since FA(·) a.nd FB (·) can have different masses a.t zero. Relation (2.2 . 1 ) is graphed in Figure 2.2 . 1 . For a.ny z, the prob ability tha.t the ra.te of return on asset A is less than z is less tha.n tha.t for asset B. Putting it differently, the probability tha.t the rate of return on asset A is greater tha.n z is greater tha.n tha.t for asset B. Note tha.t (2.2.1) does not mean tha.t asset A has a. realized rate of return tha.t is a.lwa.ys greater than tha.t of asset B. The follow ing example illustrates this point. Suppose tha.t rates of return on assets A a.nd B ca.n only take on three possible values: 0, 1/2, a.nd 1 with cumulative probabilities described in Table 2.2. 1 . It is easily seen tha.t condition (2 .2.1) is satisfied . Ex post , when asset A has a.
Stochastic Domin ance
F (z )
0
Figure 2.2 .1:
z First Degree Stochastic Dominance
z
FA (z) FB (z) Table 2.2.1:
Dominance
0 1/4 1/2
1/2 3/ 4 4 /5
1 1 1
An Example of First Degree Stochastic
realized ra.te of return 0, asset B ca.n ha.ve a. realized ra.te of return 1 . H the rates of return on A a.nd B were independent, then the probability tha.t the ra.te of return on A is 0 a.nd the ra.te of return on B is 1 would be 0.05 . Let u( ) be a.ny continuous a.nd increasing utility function representing a. nonsa.tia.ble individual's preferences a.nd assume with out loss of generality tha.t this individual has unit initial wealth. H the individual invests in asset A, his expected utility is E[u ( 1 + r A )] a.nd similarly for asset B , where we use r A and rB to denote the rates of return on asset A a.nd asset B , respectively. We wa.nt to show tha.t 2. 3.
·
42
Foundations for Fin ancial Economics
(2. 2.1) implies that E[u(1 + rA)]
{
where
}��
u( 1 + z) dFA ( z)
� �
E[u(1 + iB )] , or equivalently,
{
���
u( 1 + z) dFB (z),
{ u( 1 + z) dFA (z) = u( 1 ) FA ( O ) + J[o,1)
lal u( 1 + z)dFA (z) ,
and similarly for the integral on the right side. Integration b y parts gives:
{ u(1 + z) d[FA(z)  FB (z) j = u( 1) (FA ( O)  FB ( O )) lro,lJ +
11 u( 1 + z) d[FA (z)  FB (z)]
= u( 1) (FA ( O )  FB (O) ) + u(1 + z) ( FA (z)  FB (z))
 11 [FA (z)  FB (z)] du(1 + z) .
Since FA (1)
=
FB ( 1 )
=
=
I�
1 , we get
{ u( l + z) d[FA (z)  FB (z) ] lro,IJ
 11 [FA(z)  FB (z)] du ( 1 + z).
(2.3.1)
Now by the hypothesis that FA (z)  FB (z) � 0 and the fact that u(·) is increasing, the sign of (2.3.1) is nonnegative. Therefore ( 2.2 . 1) is sufficient for A P�D B. Next we estab lish that ( 2.2.1) is also a necessary condition for A P�D B. Suppose that A r�D B, and FA (z) > FB (z) for some z E [0, 1 ) . Note that z '1 1 , as FA (1) = FB ( 1 ) = 1 . Since cumulative distri bution functions. are increasing and right continuous, there must exist an interval [z, c] E [0, 1] such that:
FA (z)
>
FB (z)
V z E [z, c] .
(2.3.2)
43
Stoch astic Dominance s
( y)
y
0
Figure 2.3.1:
Relation (2.3 .2)
' If
we put a(y) = FA (!I)  FB (y) , then ( 2.3.2) can be depicted in Figure 2.3. 1 . ... Note that s(y) is the integrand of the righthand side of (2.3 . 1 ) . B y (2.3.2) , we know that s(y) is strictly positive on [x, c] . H we can construct a continuous and increasing utility function that strictly increases only on [1 + x, 1 + c] , then (2.3 .1) would h ave a negative sign, which contradicts the hypothesis that A P�D B. This can be accomplished by defining Y
z E [0, 1] ,
(2.3.3)
where 1 [z , c ] (t) equals 1 if t E [x, c] and 0 elsewhere . It is quickly checked that u ( · ) is increasing and continuous. Moreover,
U'( 1 + z) = 1 [l+ z ,Hc] ( 1 + z) =
{ 01
if 1 + z E [1 + x, 1 + c] , (2 3 . otherwise , ·
·
4)
that is, the marginal utility is equal to 1 on [1 + x, 1 + c] and 0 elsewhere, as graphed in Figure 2.3.2.
Foundations for Financial Economics
u• ( l + z)
     
0
X
z
c
Figure 2.3.2: Relation ( 2.3.4)
Now using ( 2.3.1 ) and ( 2;3.2 ) , we get
( u ( 1 + z) d[FA (z)  Fs (z)] = J[o,t)
=
which contradicts
A

11 [FA (z)  Fs (z)] du( 1 + z) 1a [FA (z)  Fs (z)] dz 0,
FSD
B;
FA (z) � Fs (z ) 'Vz E [0, 1] ; 'A  o.  =d rs  + a,  a > From 3 above, if
A
F D
� B, then asset
�i,�gh an expected rate of return as asse t B. ;;'Gowever.
A must have at least as The converse is not true,
Suppose now that the only information we have about l'&n individual is that he is risk averse. Under what conditions can ' ire unambiguously say that he prefers risky asset A to risky asset B? This subject is examined in this section. Note that risk averse ';in dividuals may have utility functions that are not monotonically i, ncreasing. We will say that risky asset A dominates risky asset B in the aense of Second Degree StochaBtic Dominance , denoted by A s �D B, if all risk averse individuals h aving utility functions whose first deriva tives are continuous except on a countable subset of [1, 2] prefer A to B. We claim that A s �D B if and only if 2.5.
( 2 .5.1 ) and
S( y )
=: fa" (FA(z)  Fs (z))dz �
0
'Vy E [0, 1] .
(2.5.2)
46
Foundations for Financial Economics
F (z)
0
Figure
y
2.5 .1: An Example of S(y)
z
If FA ( · ) and FB ( ·) are as drawn in Figure 2.5.1, then S(y) is the sum of the two shaded areas, and the signs of the areas are as indicated . 2.6. We will prove the sufficiency part first. From (2.3.1) , we
have (
A��
u(1 + z) d [ FA (z)  FB (z)] =
 1 1 [ FA (z)  FB (z)] du(1 + z) . a
Integration by parts gives
 11 [FA (z)  FB (z)] du(1 + z) =  1 1 u1 (1 + z) dS(z) u'(1 + z) S(z) l � + 1 1 S (z)du1 ( 1 + z) .
= 
Since
S ( O) = (
J[o,o] = 0,
(FA (z)  FB (z)) dz
(2 .6.1)
47
Stochastic Domin an ce
and 8 ( 1) = {
J(o,1] 1
( FA (z)  Fs (z) ] dz
=
1
=
( FA (z)  Fs (z)) zl�
= =
[ FA (z)  Fs (z) ] dz
 11 z dFA (z) 11 z dFs (z)
11 z dFs (z)  11 z dFA (z)
+
E (i s ]  E (iA ] = 0,
(2.6.2)
[�here the last equality follows from (2.5.1 ) , (2.6.1) can be written as
,.
1 . f u( 1 + z) d(FA (z)  Fs (z) ] = f S(z)du1 (1 + z) lo J(o,1]
�
0, (2 .6 .3)
:•here the inequality follows from (2.5.2 ) and the fact that u' is de ��easing. We h ave proved the sufficiency part . .:' :.';: 1 '
Now we prove the necessity part. Suppose first that :A s�D B. Since linear utility functions are admissible, we can take i.'� linear strictly increasing utility function and a linear strictly de ;' ereasing utility function. Using the definition of A s�D B, 2. 7.
{ z dFA (z) = f z dFs (z) , J(o,1] l(o,1]
or, equivalently, E (iA ] = E[ rs ] , which is (2.5.1) . Next repeating the integration by parts done in Section 2 . 6 , we get the equality of (2.6.3 ) . We claim that S(·), which is continuous, is negative. Suppose this is not the case. It then follows from the continuity of S( · ) that there must exist an interval [a, 6] with a 'I b such that S (z) > 0
V z E [a, b] .
If we can find a concave utility function whose first derivative is continuous except possibly on a countable set and strictly decreases
48
Foundations for Financial Economics u'
( I+
z)
ba
0
a Figure 2.7.1: Relation (2.7.3)
z
b
only on [1 + a , 1 + 6] , then the inequality of (2.6.3) will be violated and thus the hypothesis that A s �D B will be contradicted . We can accomplish this by defining
u ( 1 + z) =
�a• 1.1 l[Ha,HbJ (l + t) dt dy
V z E (0, 1) , (2.7.2)
which is continuously differentiable and concave with
u'(l + z) = This derivative is graphed Thus,
I
�.Q
in
1 1 l [l+a,l+6) ( 1 + t ) dt.
(2.7.3)
Figure 2 .7.1.
u ( l + z ) d(FA (z)  FB (z)] =
=
1 1 S (z) du'( l +
h 
I S ( x ) dx lra,6J
z
0. It is dear from ( 2.10.5 ) that V( ·) may not be concave. Thus when the riskiness of the risky asset increases, a risk averse individual ' may well increase his investment in the risky asset!
{
2.11.

"'( ) }
In Section 1 .26, it is demonstrated that a more risk averse ·. individual will require a higher risk premium on the risky asset for him to invest all his wealth in it than will a less risk averse individual, when there are two assets: one risky and one riskless. Equivalently, a more risk averse individual will never invest more in the risky asset than a less risk averse individual. Recall that individual i is said to be more risk averse than individual lc if R� ( z) � R! (z) V z . Unfortunately, this result does not extend to a setting where the two traded assets are both risky. This is illustrated by the example presented below, which is adapted from Ross ( 198 1 ) . Consider a two asset portfolio problem where iA and iB denote the rates of returns on the two risky assets. Denoting i = iA  iB , we assume that, for every realization of iB , E[ili B] � 0. This implies that asset A is more risky than asset B and also has a higher expected
52
Foundations for Financial Economics
rate of return . In particular , we assume for the time being that and rB are independent with
z = { 2 1
 { 01
rB =
z
with probability ! , with probability !; with probability ! , with probability ! .
Let u k () be an increasing concave utility function for individual k with
Also let G (·) be a concave function with
Now define From Section 1 .25, we know that individual i with utility function Ui is more risk averse than individual k. Assume without loss of generality that both individuals have unit initial wealth. It is easily verified that a = 1/4 satisfies the following relation
z
E [u�( l + rB + a ) z] = o . That is, individual A. However
E[u� ( l +
;:B
+
k
optimally invests 1/4 units of his wealth in asset
�z)z] = E[G' ( uA: ( 1 + = 5 > 0.
;:B
+
�z) u� ( l + rB + �z)z]
This implies that individual i ' s utility will b e increased if h e invests strictly mor� than 1/4 units of wealth in asset A. Even though indi vidual i is more risk averse in the ArrowPratt sense than individual k, he will n ot take a less risky position.
53
Stochastic Domin ance
2.12. Ross ( 19 81 ) proposes another measure of risk aversion. In dividual i is said to be strongly m ore risk a1.1erse than individual k if ' . u� ( z ) uH z ) (2 . 1 2 . 1) mf , � sup , . (z ) • uk ( z ) uk •
Relation (2.12.1) implies that for arbitrary '
u
� ( z ) > uH z ) u� ( z) uH z) Rearranging gives
'
z,
we have
·
u� ( z ) u�(z)   > . u
H z)
( 2. 12.2) � (z ) which implies that individual i is more risk averse than individual k in the sense of ArrowPratt. The following example shows that (2. 12. 1 ) is strictly stronger than (2. 12.2) . Let u; ( z ) =  e  Q• and u �: ( z ) =  e  h , with a > b. It is easily verified that i is more risk averse in the sense of ArrowPratt than lc. However, and
u
�' ( z2 ) u %( zz )
imply that when
zz  z1
=
u
(�) 2 b
e  (G  b) •2
is large we will have
contradicting (2. 1 2 . 1 ) . Thus the strong measure of risk aversion is strictly stronger than the ArrowPratt measure . 2.13.
Now we recall from Section 1 .25 that if individual i is more risk averse in the sense of ArrowPratt than individual k, then there exists a monotone concave function G such that Ui = G (u�: ) . When individual i is strongly more risk averse than individual k, we have a similar characterization.
54
Foundations for Fin anc¥.1 Economics
We claim that i is strongly more risk averse than k if and only if there exists a decreasing concave function G and a strictly positive constant ..\ such that (2.13 . 1)
Consider the sufficiency part. Differentiate (2.13. 1) to get
and
u�' (z ) = ..\u � (z ) + G " (z) � ..\ u: (z ) V z,
where the inequalities follow from the fact that G is decreasing and concave . The above two relations imply (2.13 .2)
It is easily seen that (2.13.2) implies (2.12.1). Conversely, let i be strongly more risk averse than k. By the definition of the strong measure of risk aversion , there exists ..\ > 0 such that (2.13 .2) holds. Define G by (2 . 13. 1 ) . Differentiating we get G' (z) = uHz)  ..\u� (z ) � 0 and
G " (z) = u�'(z )  ..\u:(z) � 0,
where the inequalities follow from (2.13.2) . Thus we have proved our claim that i is strongly more risk averse than k if and only if (2.13.1) holds. 2.14 . Now let us again consider the two risky asset portfolio problem of Section 2 . 1 1 . Assume that (2. 12.1) holds and that indi vidual i is strongly more risk averse than individual k. Let a be such that (2. 14 . 1) E [u� ( 1 + r B + ai)i] = 0.
55
Stoch astic Domin an ce
Then a is the amount individual k should optimally invest in asset A. Using (2. 13 . 1 ) and (2.14.1), we obtain E[uH1 + r B + ai)i]
= E[A u� (1 + rB + ai)i + G'(1 + rB + ai)i] = E[G' ( 1 + rB + ai)i] .
Using iterative expectations and the definition of covariance , we can �ewrite the above relation as: '
'·
=
;�.
L
= E [ E[G' ( 1 + rB + ai)ilrBl] E cov (G'(1 + iB + ai)1 ilrB )
E[G'(1 + rB + ai)i]
[
+ E [G'(1 + rB + ai) lrB ] E [il rBJ
]
� E[Cov (G' ( 1 + rB + ai), iirB )] � 0,
�here the first inequality follows from the hypothesis that E [ ilrB ]
l' and the fact that G'
�
� 0, and the second inequality follows from �e concavity of G. This implies that individual i 1 s utility can be increased by investing an amount smaller than a in asset A. That �� . �, i will opt�mally c�oose a less ris�y portfolio . Thus t � e stronger . . Jineasure of r1sk avers1on g1ves the rtght comparat1ve stat1cs. lr
��:·
�1�·
Exercises
f•. 1. A risky asset A is said to third degree stochastically dominate
risky asset B if all investors exhibiting decreasing absolute risk aversion prefer A to B 1 denoted by A T �D B. Provide a suffi cient condition strictly weaker than that for the second degree stochastic dominance on the distribution functions for A T �v B. 2.2. Suppose that there are two risky assets with rates of return r1 and r21 which are independent and identically distributed . Show that the equally weighted portfolio is an optimal choice for any risk averse investor . 2 .3. Suppose that there are five_ states of nature denoted by Wn n = 1 1 2, . . . 1 5, all of which are of equal probability. Consider two risky assets with rates of returns rA and rB as follows:
S6
Foundations for Financial Economics
state rA
Wl
0.5 0.9
wz
0.5 0.8
ws
0.7 0.4
w,.
0.7 0. 3
W&
0.7 0.7
f rB Explain which asset a risk averse investor will choose . 2.4. Suppose that there are two risky 888ets with random rates of returns rA and rB , respectively. Assume that rA and rB are independent and have the same mean. We know further that
rB g rA + f and that rA and f are independent. Does this imply that rB dominates rA in the sense of second degree stochastic dominance? Show that if these are the only assets available to a risk averse expected 'lftilitymaximizing individual, this indi vidual will invest more in asset A than in asset B.
Remarks. Machina (1982) generalized some of the Ross (198 1) results on the stronger measure of risk aversion. Roim (1983) showed that the comparative statics of Section 2.14 . cannot be generalized to a multiassets case. The sufficiency proofs of r�D and s�D are adapted from Hadar and Russel (1969) . For a bibliography on the subject of stochastic dominance, see Bawa (1981) . Exercise 2.4 was provided to us by Richard Khilstrom.
References
Bawa, V. 198 1 . Stochastic dominance: A research bibliography. Bell Laboratories Economics Discussion Paper # 196. Hadar, J . , and W. Russell. 1969. Rules for ordering uncertain prospects , American Economic Review 59: 2534. Hanoch, G . , and C. Levy. 1969. Efficiency analysis of choices involv ing risk. Review of Economic Studies 36:335346. Machina, M. 1982. A stronger characterization of declining risk aver sion. Econometrica 50:10691079. Roell, A. 1983. Risk aversion and wealth effects on portfolios with many independent assets. Mimeo. Johns Hopkins University. Baltimore, Maryland.
Stochastic Domin an ce
Roes,
57
S. 198 1 . Some stronger measures of risk aversion in the small and large with applications. Econometric a 49:621638. Rothschild , M., and J. Stiglitz. 1970. Increasing risk I: A definition. Journ al of Econ omic Theory 2:225243 . ;· Rothschild, M . , and J. Stiglitz . 1971 . Increasing risk II: Its economic & consequences. Journal of Economic Theory, 3:668 4.
C H AP TER 3 MATHEMAT I C S O F T H E P O RTFOLIO F RO NTIER
3.1. I n Chapter 2 we demonstrated that when risky asset A second degree stochastically dominates risky asset B, risky asset A must have the same expected rate of return as risky asset B and a lower variance. When there are more than two assets and when port folios can be formed without restrictions, if there exists a portfolio of assets that second degree stochastically dominates all the portfo lios which have the same expected rate of return as it has, then this domin ant portfolio must have the minimum variance among all the portfolios. This observation is one of the motivations to characterize those portfolios which have the minimum variance for various levels of expected rate of return. 3 . 2 . The meanvariance model of asset choice has b een used extensively in finance since its development by Markowitz (1952) more than two decades ago. A preference for expected return and an aversion to variance is implied by monotonicity and strict concavity
59
60
Foundations for Financial Economics
of an individual's utility function. However, for arbitrary d istribu tions and utility functions, expected utility cannot be defined over j ust the expected returns and variances. Nevertheless, the mean variance model of asset choice is popular b ecause of its analytical tractability and its rich empirical implications. Two technical moti vations, besides the one stated in Section 3 . 1 , exist and are briefl y reviewed below. 3 . 3 . An individual's utility function may be expanded as a Taylor series around his expected end of period wealth,
u( W )
= +
u(E[ W ] ) + u' (E[ W ] )( W  E[ W ] ) u " (E[ W ] )( W  E[ W j)2 + R3 ,
�
where
and where u(n) denotes the nth derivative of u. Assuming that the Taylor series converges and that the expecta tion and summation operations are interchangeable, the individual's expected utility may be expressed as ( 3 . 3 . 1) where (3.3.2) and where m" ( W ) denotes the nth central moment of W . Relation (3.3.1) indicates a preference for expected wealth and an aversion to variance of wealth for an individual having an increas ing and strictly concave utility function. However, relation (3 .3 .2) illustrates that expected utility cannot be defi n ed solely over the ex p ected value and variance of wealth for arbitrary distributions and preferences, as indicated by the remainder term which involves higher order moments.
61
Mathematics of the Portfolio Frontier
3 . 4 . For arbitrary distributions, the meanvariance model can be motivated by assuming quadratic utility. Under quadratic util ity, the third and higher order derivatives are zero and, therefore, = 0 for arbitrary distributions. Hence, an individual 's ex pected utility is defined over the first two central moments of his end of peri () d wealth,
E[R3]
W, E[u(W)] = E[W]  �E[W2] = E[W]  � ((E[W])2 + a2(w)) .
Thus, when expected rates of return and variances are finite, quadratic utility is sufficient for asset choice to be completely de scribed in terms of a preference relation defined over the mean and variance of expected returns. Unfortunately, quadratic utility dis plays the undesirable properties of satiation and increasing absolute risk aversion. The satiation property · implies that an increase in wealth beyond the satiation point decreases utility. Increasing abso lute risk aversion implies that risky assets are inferior goods. Thus, economic conclusions based on the assumption of quadratic utility function are often counter intuitive and are not applicable to indi viduals who always prefer more wealth to less and who treat risky investments as normal goods. 3.5.
For arbitrary p references, the meanvariance model can be motivated by assuming that rates of return on risky assets are multivariate normally distributed. The normal distribution is com pletely described by its mean and variance. Under normality, the third and higher order moments involved in can be expressed as functions of the first two moments, and is, therefore, solely a function of the mean and variance. Normal distributions are also stable under addition; i.e., the rate of return on a portfolio made up of assets having returns that are multivariate normally distributed is also normally distributed. The lognormal distribution is also com pletely described by its mean and variance; however, it is not sta ble under addition. That is, a portfolio made up of assets having returns that are multivariate lognormally distributed is not lognor mally distributed. Thus, for utility functions that are defined over a
E[R3] E[R3]
62
Foundations for Financial Economics
normally distributed end of period wealth, the assumption that asset returns are multivariate normally distributed implies that demands for risky assets are defined over the mean and variance of portfolio rates of return. However, for other classes of utility functions such as u(z) ln(z) , expected utility is not defined over nonpositive wealth levels. Unfortunately, the normal distribution is u nbounded from below, which is inconsistent with limited liability and with eco nomic theory, which attributes no meaning to negative consumptio n. Fortunately, multivariate normality is only a sufficient distributional condition for all individuals to choose meanvariance efficient port folios, not a necessary condition.
=
3 . 6 . Based on the above, the meanvariance model is not a general model of asset choice. Its central role in financial theory can be attributed to its analytical tractability and the richness of its empirical predictions. This chapter develops the analytical relations between the means and the variances of rates of return on feasible portfolios. This will provide the basis for the development of more general conditions for meanvariance asset choice and meanvariance asset pricing models in Chapter 4. 3 . T. We suppose that there are N � 2 risky assets traded i n a frictionless economy where unlimited short selling is allowed and th at the rates of return on these assets have finite variances and unequ al exp ectations , unless otherwise mentioned . It is also assumed th at the random rate of return on any asset cannot be expressed as a linear combination of the rates of return on other assets. Under this assumption, asset returns are said to be linearly independent and their variance covariance matrix V is nonsingular. The variance covariance matrix is also symmetric because Cov (r; , ri ) Cov (ri , r; ) , for all i , j . Such a symmetric matrix is said to b e positive definite if for arbitrary Nvector of constants w, with w f:: 0, w T Vw > 0, where T denotes "transpose" and where w f:: 0 means therE;! is at least one element of w that is not zero. V is a positive definite matrix because w T Vw is a portfolio variance even when the portfolio weights do not sum to unity and because variances of risky portfolios
=
Mathematics of the Portfolio Frontier
63
are strictly positive. 3.8. A portfolio is a frontier portfolio if it has the minimum variance among portfolios that have the same expected rate of return. A portfolio p is a frontier portfolio if and only if w P, the N vector portfolio weights of p, is the solution to the quadratic program:
1 min 2 w TVw {w}

(3 . 8. 1 )
s.t.
w T e = E[rp] and wT 1 = 1 , where e denotes the N vector of expected rates of return on the N risky assets , E [rp] denotes the expected rate of return on portfolio p, and 1 is an N vector of ones. The programming problem given in (3.8.1) minimizes the port folio variance subject to the constraint that the portfolio expected rate of return is equal to E[rp] and that the portfolio weights sum to unity. Note that short sales (i.e., negative portfolio weights) are permitted. Therefore, the range of expected returns on feasible port folios is unbounded. (This follows from the assumption that assets do not have identical expected rates of return.) Forming the Lagrangian, wP is the solution to the following: (3.8. 2 ) where ..\ and 7 are two positive constants. The first order conditions are aL
aw
aL a ..\
= Vwp  ..\ e  71 = 0,
= E [ rp ]  wPT e = 0,
aL  = 1  wpT 1 = 0, a,
(3. 8.3a) (3 .8.3b) (3 .8.3c)
where 0 is a N vector of zeros. Since V is a positive definite matrix, it follows that the first order conditions are necessary and sufficient for a global optimum.
Foundations for Financial Economics 3.9. Solving (3.8 .3a) for Wp gives wp
= >. (v 1 e ) + 7(v 1 t ) .
(3.9.1)
Premultiplying both sides of relation (3.9 . 1) by e T and using (3.8.3b) gives ( 3.9.2a) Premultiplying both sides of relation (3.9.1 ) by 1 T and using (3.8 .3c) _ gives 1 = .\ ( 1 T V 1 e ) + ')' ( 1 T V  1 1 ) . (3.9.2b) Solving (3.9.2a) and (3.9.2b) for >. and ')' gives
=
OE[ip ]  A D
1=
B  AE[rp ] D
>.
where
A= B=
( 3.9 . 3a) (3.9.3b)
1 T v 1e = e T v  1 1 , e T v  1 e,
0
= l T V 1 1 , D = BO  A2 • Since the inverse of a positive definite matrix is positive definite, B > 0 and 0 > 0. We claim that D > 0. To see this, we note that (Ae  B 1 ) T V  1 (Ae  B 1 )
=
B(BC  A2)
The lefthand side of the above relation is strictly positive, since V  l i s positive definite. Hence the righthand side is strictly positive. By the fact that B > O, we have BO  A2 > 0, or, equivalently, D > 0. Substituting for >. and ')' in relation (3.9.1) gives the unique set of portfolio weights for the frontier portfolio having an expected rate of return of E[ip ] : ( 3.9.4) where
and
65
Mathematics of the Portfolio Frontier
Note that relations (3.8.3a) , ( 3.8 .3b) , and (3.8.3c) are necessary and sufficient conditions for Wp to be the frontier portfolio having an ex pected rate of return equal to E[rp ] · Therefore, any frontier portfolio can be_ represented by ( 3.9.4) . On the other hand, any portfolio that can be represented by (3.9.4) is a frontier portfolio. The set of all frontier portfolios is called the portfolio frontie r . Now we claim that g is the vector of portfolio weights corre sponding to a frontier portfolio having a zero expected rate of return and that g + h is the vector of portfolio weights of a frontier p ortfo lio having an expected rate of return equal to 1 . To see this, we first substitute zero for E[rp] of relation (3.9.4) to get and then substitute 1 for E[rp ] of (3.9.4) to get
Wp = g
+h
· 1 = g + h.
Next we claim that the entire portfolio frontier can be g e n erated by forming portfolios of the two frontier portfolios g and g + h. Let q be a frontier portfolio having an expected rate of return E[rq] · From (3.9.4) , we know that Consider the following portfolio weights on g and g + E[rq] , E[rq ] } , whose portfolio weights on risky assets are
h
{1 
That is, the portfolio { 1  E[rq ] , E[rq ] } on g and g + h generates the frontier portfolio q. Since the portfolio q is arbitrarily chosen, we have shown that the entire portfolio frontier can be generated by the two frontier portfolios g and g + h. 3 . 10 .
Note that the arguments in the last section showing that the portfolio frontier is generated by g and g + h only use the fact
66
Foundations for Financial Economics
that the two frontier p ortfolios g and g + h do not h ave identical exp ected rates of return. The following much stronger statement is in fact valid: The portfolio frontier can be generated by any two distinct frontier portfolios. To see this, let Pl and P2 be two distinct frontier p ortfolios, and let q be any frontier p ortfolio. We want to show that q is a p ortfolio generated by Pl and p2 . Since E[rpJ is not equal to E[rp2 ] , there exists a unique real number a such that
(3. 10 .1) Now consider a portfolio of Pl and P2 with weights { a, (1  a) } . We have awp 1 + (1  a)wp 2 = a( g + hE [ rp J ) + (1  a) (g + hE [rp 2 ] ) = g + h(aE[ rpJ + ( 1  a) E[rp2 ] ) = g + hE[rq J
(3 . 10.2 )
where the first equality follows from the fact that Pl and p 2 are frontier p ortfolios, the third equality follows from (3 . 10.1) , and the fourth equality follows from the fact that the weights for a frontier p ortfolio are uniquely determined. Thus, we h ave demonstrated that the portfolio frontier can be generated by any two distinct frontier portfolios. 3. 11. The covariance between the rates of return on any two frontier portfolios p and q is
Cov (rp , rq ) = w; vwq =
� (E [rp]  A/C) (E[rq]  A/C) + 1 /C,
( 3.11.1 )
where we h ave used the definition of covariance and the portfolio weights for a frontier portfolio given in relation ( 3.9.4 ) . The definition of the variance of the rate of return of a portfolio and ( 3 . 1 1 . 1 ) give u 2 ( rp ) _ ( E[rp]  A/C) 2 = 1' 1/ C DjC 2
( 3.11.2 a )
67
Mathematics of the Portfolio Frontier
Figure 3 .1 1. 1: Portfolio Frontier in the u (r) E[r]
space
which is a hyperbola in the standard deviationexpected rate of re turn space · with center (0, A/C) and asymptotes E[rp] = A/C ± ..fl57Cu (rp ) , where u2 (rp) and u (rp) denote the variance and the standard deviation of the rate of return on the portfolio p , respec tively. Relation ( 3 . 1 1 .2a) can equivalently be written as
( 3.11.2b) which is a parabola in variancei!xpected rate of return space with vertex ( 1 /C, A/C) . The portfolio frontier in meanstandard devia tion space is graphed in Figure 3 . 1 1 . 1 , and the p ortfolio frontier in meanvariance space is graphed in Figure 3 . 11.2. The portfolio hav ing the minimum variance of all p ossible p ortfolios, or the minimum variance portfolio , is at ( .Jl7G, A/C) in Figure 3 . 1 1 . 1 . This follows directly from ( 3 . 1 1 .2a) . 3. 12.
mvp,
The minimum variance portfolio, denoted henceforth by has a special property : The covariance of the rate of return on
68
Foundations for Financial Economics
E [ r l
A /C
Figure 3 . 11 . 2 :
Sp ace
Portfolio Frontier in the u 2 (r) E[r]
the minimum variance p ortfolio and that on a n y portfolio (not only those on the frontier ) is always equal to the variance of the rate of return on the minimum variance p ortfolio. To see this, let p be any p ortfolio. We shall demonstrate below that (3.12 . 1) We consider a p ortfolio of p and and with minimum variance. Then, following program:
mvp
with weights a and 1  a be the solution to the
a must
The first order necessary and sufficient condition for solution is:
a
to be the
Since mvp is the minimum variance portfolio, a = 0 must satisfy relation (3 .12.2) . Thus we have relation (3.12.1) . In words, the co variance between the rate of return on any portfolio and that on the
69
Mathematics of the Portfolio Frontier
minimum variance portfolio is equal to the variance of the rate of return on the minimum variance portfolio. Note that in the argu ments proving that relation (3.12.1 ) holds for any portfolio p, the assumption that risky assets do not h ave identical expected rates of return is never used. Thus relation (3.12.1 ) h olds even when risky assets h ave identical expected rates of return. 3 . 13 .
Those frontier portfolios which h ave expected rates of return strictly higher than _ that of the minimum variance portfolio, A/C, are called efficient portfolios . Portfolios that are on the p ortfo lio frontier but are neither efficient nor min�mum variance are called inefficient portfolios. For each inefficient portfolio there exists an efficient one having the same variance but a higher expected rate of return. Let wi , i = 1, 2, . . . , m , be m frontier portfolios and C4 , i = 1 , 2 , . . . , m, be real numbers such that E�1 ai = 1. Then denoting the expected rate of return on p ortfolio i by E[ri] for i = 1, 2, . . . , m, we have m m
2: aiwi = 2: ai (g + hE[ri ]) i =l
 i=l
m
= g + h L ai E[ri ] ·
i= l Recalling (3.9.4 ) , the second line of the above expression i s a frontier p ortfolio having an expected rate of return equal to E� 1 ai E[ri ] · Thus any linear combination o f frontier p ortfolios i s o n the frontier. If p ortfolios i = 1, 2, . . . , m are efficient portfolios, and if ai , i = 1 , 2, . . . , m are nonnegative, then m m (3.13. 1) L ai E[ri] � L ai 0A = A/C. i= l i= l Formally stated, any convex combination of efficient p ortfolios will be an efficient portfolio. The set of efficient portfolios is thus a convex set. 3 . 14 .
One important property of the portfolio frontier is that for any portfolio p on the frontier, except for the minimum variance
70
Foundations for Financial Economics
portfolio, there exists a. unique frontier portfolio, denoted by zc(p) , which has a zero covariance with p. Setting the covariance between two frontier portfolios p and zc (p) , given in relation (3. 1 1 . 1) , equal to zero: Cov (rp , rzc(p ) ) =
� ( ( E [rp ]  AjC) ( E [rzc (p) ]  A/C) + D jC2 ) = 0
and solving for the expected rate of return on
zc(p) , we get:
(3 . 14.1 ) · ·
( 3. 14.2 )
In fact , ( 3.1 4 .2 ) defines zc(p) . The uniqueness assertion follows from the fact that ( 3.14.2 ) defines E[r z c (p ) ] uniquely, which in turn deter mines zc(p) uniquely by ( 3.9.4 ) . It is also easily seen from (3.11.1) that the covariance of the minimum variance portfolio and any other frontier portfolio is equal to 1/C , which is strictly positive. There fore, there does not exist a frontier portfolio that has zero covariance with the minimum variance portfolio. 3 . 1 5 . Equation ( 3.14.2 ) gives us a clue to the location of zc(p ) for p other than the minimum variance portfolio. If p is an efficient portfolio, then ( 3. 14.2 ) implies that
thus
zc(p) is an inefficient portfolio, and vice versa. Geometrically zc(p) can be located by the following fact: The
intercept on the expected rate of return axis of the line tangent to the p ortfolio frontier in the standard deviationexpected rate of return space at the point associated with any frontier portfolio p ( except the minimum variance portfolio ) is E[rz c (p ) ] · Alternatively, in the varianceexpected rate of return space, the intercept on the E[r] axis of the line joining any frontier portfolio p and the mvp is equal to E [rzc (P ) l · To see these, we first differentiate ( 3.11 .2a ) totally with respect to u(rp) and E[rp] to obtain dE[rp] du (rp ) 
u(rp) · D c . E [rp]  A '
( 3.15.1 )
71
Mathematics of the Portfolio Frontier
The Location o f a Zero Covariance Portfolio in the u (r)E [r] Space
Figure 3 . 1 5 . 1 :
which is the slope of the portfolio frontier at the point (u (rp ) , E[rp ] ). The expected rate of return axis intercept of the tangent line is
(3. 15.2)
where the second equality follows from (3. 1 1 . 2a) , and the third equal ity follows from (3.14.2) . The geometric picture of relation (3.15 .2) can b e seen in Figure 3.15 . 1 . Next, i t i s easily seen that the line joining a frontier portfolio p and the mvp, in the u 2 (r)E [r] plane, can be expressed as (3.15 .3)
72
Foundations for Financial Economics
E [r )
The Location of a Zero Covariance Portfolio in the u 2 (r)E[r] Space
Figure 3 . 1 5 . 2 :
S ubstituting u 2 (r) = 0 into (3.15.3) , we h ave that the intercept of the line on the expected rate of return axis is equal to E[rzc(p ) ] · This result is presented in Figure 3 .15. 2 . Finally, let p be a p ortfolio which is not o n the portfolio frontier as depicted in Figure 3 .15.3, which we note is in the u 2 (r)E[r] plane . We claim that the intercept on the expected rate of return axis of the line joining p and mvp is equal to the expected rate of return on a portfolio, q , that has zero covariance with p and the minimum variance among all the zero covariance portfolios with p. To see this, we note that w q is the solution to the following program: min  wqT Vwq Wq 2 s.t. w ; vwp = 0 , 1
WTl q
= 1.
(3 . 15 .4)
73
Mathematics of the Portfolio Frontier
Using the Lagrangian method, we can easily verify that
( 3 . 15.5)
where w mvp denotes the vector of portfolio weights of the minimum variance portfolio. That is, the minimum variance zero covariance portfolio of p is a linear combination of p and mvp. Since u 2(rp) > 1 / C , q is constructed by short selling portfolio p and buying the minimum variance portfolio. The expected rate of return of q is T wq e _  1 _
E[rp]
Ca2 (rp)
Cu2 (rp)
1 
E[rp ]  Au2 (rp) 1  Cu2 ( rp )
A
Cu2 (rp ) C
( 3 . 15 .6)
which is easily verified to be the intercept on the expected rate of return axis of the line joining p and mvp in the u2 (r)E[r] plane. In Exercise 3 .6 we ask the reader to show that the p ortfolio frontier generated by two assets or portfolios having distinct expected rates of return passes through these two assets or p ortfolios. Consider the portfolio frontier generated by p and mvp, which is demonstrated graphically in Figure 3 . 1 5 . 3 . Note that this p ortfolio frontier lies inside the portfolio frontier of all assets and touches it at a single point, the mvp. S ince any linear combination of two frontier port folios is on the frontier, any linear combination of p and mvp is on the p ortfolio frontier generated by p and mvp. It follows that the portfolio q is on the portfolio frontier generated by p and mvp as it is a linear combination of p and mvp by (3.15.5) . We have demonstrated the existence of a zero covariance p ortfolio for any frontier portfolio other than the minimum vari ance p ortfolio. A characterization of the relationship between the expected rate of return on any portfolio q, not n ecessarily on the frontier, and those of the frontier portfolios is given below. 3 . 16.
74
Founda tions for Financial Economics E [r] ,
Figure 3 . 1 5 .3: Minimum Variance Zero Covariance Portfolio of a Nonfrontier Portfolio
Let p be a frontier p ortfolio other than the minimum variance portfolio, and let q be any portfolio. The covariance of rp and r q is Cov (rp , rq ) = w;vwq = .XeTv 1Vwq + 11 T v 1Vwq = AeTwq + I1TWq = .X E[rq] + 1. (3 . 16 . 1)
where the second equality follows from the fact that p is a frontier portfolio and relation (3.9.1) , and where the fourth equality follows from the definition of E[rq ] and the fact that Wq is a vector of port folio weights. S ubstituting (3.9.3a) and (3.9.3b) for A and 1 , respec tively, into (3.16.1) we get E[ q] r
AE[rp]  B D ) + Cov ( rq , rp CE[rp]  A CE [rp ]  A D/C 2 A C E[rp]  A/C D Cov (rq , rp) 1 [ E [ rp ]  A/ C ] 2 + + C D/C CE[rp ]  A u 2 (rp ) 
=
[
_
]
75
Mathematics of the Portfolio Frontier
= E [izc(p) ] + {Jqp =
=
(
E[rp] 
� + E [r��:/ C )
E[izc (p) ] + {Jqp ( E [rp ]  E [ rzc (p) J) (1  {Jqp )E[rzc (p)] + {Jqp E[rp] ,
(3. 16.2)
where_ we have used (3 . 1 1 .2a) in the second equality, where {Jq p = Cov (rq , rp ) ju2(ip ) , and where the fourth equality follows from rela tion (3 .15.2) . The expected rate of return on any p ortfolio q can be written as a linear combination of the expected rates o f return on p and on its zero covariance p ortfolio, with weights {Jq p and 1  {Jqp · Note that since zc(zc(p) ) = p for any frontier portfolio p other than the mvp, we can also write (3.16.2) as (3.16.3)
From the fact that E[rp] i= E[rz c (p) ] , there exists a unique number, say a, such that E[rq] = aE[rp] + ( 1  a) E[rz c (p ) ] · Therefore , relations (3.16.2) and (3. 16.3) imply that (3.16.4)
and we can write (3 . 16.5)
Relations (3.16.2) , (3 . 16.3) , and (3 . 16.5) are equivalent relations. 3 . 1 7. The relationship among the three random variables rq , ip and izc (p) can always be written as
(3.17.1)
with Cov (rp , lq ) = Cov (rzc (p) , lq ) = E[lq] = 0, where (fJo , fJb fJ2 ) are coefficients from the "multiple regression" of rq on ip and rzc (p) . Since ip and rz c (p) are uncorrelated , we have (3. 17.2)
76
Foundations for Financial Economics
It then follows from (3.16.5) that f3o = 0, and thus we can always write the return on a portfolio q as: (3 .17. 3) with Cov (rp , fq ) = Cov (rzc (p ) , f'q) = E[lq ] = 0. This relation will be particularly useful in Chapter 4. 3 . 1 8 . In previous sections, we have characterized properties of the frontier portfolio when a riskless asset does not exist. When a riskless asset does exist , some simple results follow. Let p be a frontier portfolio of all N + 1 assets, and let Wp denote the N vector portfolio weights of p on risky assets. Then wP is the solution to the following program:
where we still use e to denote the N vector of expected rates of return on risky assets, and r1 is the rate of return on the riskless asset . Forming the Lagrangian, we k now that Wp is the solution to the following :
The first order necessary and sufficient conditions for Wp to b e the solution are and Solving for Wp , we have w
p
=
y  l (e 
r
r!
r]J 1) �� ]{ � ' E[ p
(3 .18.1)
77
Mathematics of the Portfolio Frontier
E [r]
�0+�� � CT ( r J
Figure 3.18.1: A Portfolio Frontier when r1
0 as A2  B C < 0. The variance of the rate of return on portfolio p is (3 .18.2 ) where the second equality follows from substituting relation (3 . 1 8 .1 ) . Equivalently, we can write if E[ip ] ?: r1 , if E [ip]
E[re] _
=
A/C 
( 3 . 1 1 . 2a )
=
A/C
Vii ,
D/C2 Tj A/C ,
(3 . 1 8 .4)
_
( 3 . 1 8 . 4) ,
t h at i s , we use t h e result in Section 3 . 1 5 to conclude that TJ . Now using relations
E [re]  T J a(re)
== A _ ( C
==
and
we get
E [rzc ( e ) )
=
D/C2 _ C ( r 1  A/C) r1 ) T J  A/C Vii H Cr1  A VH , VII Cr1  A .
.

=
which is to be shown.
Any p ortfolio on t h e line segment r fe is a convex combination of
p ortfolio e and the riskl ess asset . A ny portfolio on t h e half line r f q..
Vii a(rp)
other than t hose on r 1 e i nvolves shortselling the riskless
asset and investing the proceeds in portfolio
e.
checked that any p ortfolio on the half line r f shortselling p ortfolio as set .
e
It can also easily be

VIIa( fp)
involves
and investing the proceeds in the riskless
79
Mathematics of the Portfolio Frontier
E [r l
r1
=
A/C
/
/
/
/
/
/
/.
�
�
/
/
/
mvp
F igure 3 . 1 8 . 3 : A Portfolio Frontier when r1 = A/C
_
� r1 > AjC. The story here is a little different) The port folio frontier of all assets is graphed in Figure 3 .1 8.2. Any portfolio on the halfline r1 + v'Hu (ip) involves shortselling portfolio e1 and investing the proceeds in the riskless asset. On the other h and, any portfolio on the halfline r1  v'Hu (ip) involves a long p osition in portfolio e1 • � rf = A/C . In this case,
H = B  2Ar1 + Cr1 2 = B  2A (A/C) + CA 2 jC 2 BC  A2 D = > c
c
o.
Recall that E [ip] = A/C± y757C u (ip) are the two asymptotes of the p ortfolio frontier of risky assets. The portfolio frontier of all assets is graphed in Figure 3 .18.3. In the previous two cases i t i s very clear how the portfolio fron tiers of all assets are generated from looking at the figures. Portfolio frontiers are generated by the riskless asset and the "tangency" port folios e and e ' , respectively. In the present case, there is no tangency
80
Foundations for Financial Economics
portfolio. Therefore the portfolio frontier of all assets is not gener ated by the riskless asset and a portfolio on the portfolio frontier of risky assets. The question is how the p ortfolio frontier of all assets is generated. S ubstituting r1 = A/C into relation ( 3 . 1 8 . 1 ) and premultiplying Wp by l T , we get
=0
.
Therefore any portfolio on the portfolio frontier of all assets involves investing everything in the riskless asset and holding an arbitrage portfolio of risky assets  a portfolio whose weights sum to zero. 3 . 19 . When there exists a riskless asset , a relation similar to ( 3.17.2 ) can be derived. Let q be any portfolio, with Wq the portfolio weights on the risky assets. Also, let p be a frontier portfolio, with Wp the p ortfolio weights on risky assets. We assume that E [rp] # r 1 . Then Cov ( rq , rp) = w ;vwp
Using ( 3.18.2 ) , we obtain
(E [rq]  r1) ( E [rp]  r1) H
( 3 . 19 . 1 ) This relation holds independent of the relationship between r 1 and AjC . Given ( 3.19.1 ) and Section 3.17, we can readily write with Cov ( rp , Eqp) = E [E'qpj = 0, for any portfolio portfolio p other than the riskless asset .
( 3.19.2 ) q
and any frontier
Mathematics of the Portfolio Frontier
81
Exercises 3 . 1 . Let there be two securities with rates of return r; and f, . Sup
3.2.
3.3.
3 .4. 3.5.
p ose that these two securities have identical expected rates of return and identical variances. The correlation coefficient be tvveen r; and r1 is p. Show that the equally weighted portfolio achieves the minimum variance independently of p . Suppose that the riskless borrowing rate i s higher than the risk less lending rate. Graphically demonstrate the portfolio fron tier of all assets. Next suppose that borrowing is not allowed. Graphically demonstrate the portfolio frontier of all assets. Let p be a frontier portfolio, and let q be any portfolio having the same expected rate of return. Show that Cov(rp , rq) = Var(rp) and, as a consequence, the correlation coefficient of rp and rq lies in (0, 1] . Let /; , j = 1, 2, . . , n, be efficient frontier portfolios. Show that if E [rp ] = L:j= 1 W iE [r,j ] , where L:j= 1 w; = 1, then E [ r,..c (p) ] :f. L:j=l w; E [r,.. c (/j )J., unless all /; are the same. Let f; , j = 1, 2, . . . , n , be efficient frontier portfolios, and let ). ; be the L agrangian multiplier of (3.8 .2) that is associated with E [iJj ] . S how that if E [rp ] = L:j= 1 wi E [r1j ] , where L: j= 1 w ; = 1 , then E[i,.. c ( p) ] = lP L:j= 1 w; ). ; E [r,..c (/j ) ] , where Ap is the La grangian multiplier of (3 .8.2) that is associated with E [rp] , where .
Wj
� 0.
3.6. S how that the portfolio frontier generated by two assets or port folios having distinct expected rates of return p asses through these two assets or portfolios. Remarks. Most of the discussions in Sections 3.8 through 3. 17 are
freely adapted from GonzalezGaverra ( 1973) and Merton (1972). Many results of GonzalezG averra ( 1973) were later independently derived by Roll ( 1977). Chamberlain (1983) characterizes the complete family of distri butions that are necessary and sufficient for the expected utility of final wealth to be a function only of the mean and variance of fi nal wealth or for meanvariance utility functions. Epstein (1985) shows that meanvariance utility functions are implied by a set of
82
Foundations for Financial Economics
decreasing absolute risk aversion postulates.
References
Chamberlain, G. 1983. A characteriz ation of the distributions that imply meanvariance utility functions. Journ al of Economic The ory 29:185201. Epstein, L. 198 5. Decreasing risk aversion and meanvariance anal ysis. Econometrica 53 :945962. GonzalezGaverra, N. 1973. Inflation and capital asset market prices: Theory and tests. Unpublished PhD dissertation, Stanford Uni versity. Markowitz, H . 1 9 5 2 . Portfolio Selection. Journal of Finance 7:779 1. Merton, R. 1972 . An analytical derivation of the efficient portfolio frontier. Journal of Financial and Quantitative Analysis 7:18511872. Roll, R. 1977. A critique of the asset pricing theory's tests. Journ al of Financial Economics 4:129176 .
CHAPTER 4 TWO F U N D S EPARAT I O N AND LINEAR VALUATION
4.1.
In Section 3.10 we saw that all portfolios o n the portfo lio frontier can be generated by any two distinct frontier portfolios. Thus, if individuals prefer frontier portfolios, they can simply hold a linear combination of two frontier portfolios or mutual funds. In that case, given any feasible portfolio, there exists a portfolio of two mutual funds such that ind�yiduals prefer at least as much as the original portfolio. This phenomenon is termed two (mu tual) fun d separation. In the first part of this chapter, we develop conditions on as set returns that are necessary and sufficient for two fund separation given that individuals are risk averse and that variances of asset re turns exist. When asset returns exhibit two fund separation, it turns out that the two separating mutual funds must be on the portfo lio frontier. It then follows that an individual's optimal portfolio is a frontier portfolio. In equilibrium, markets h ave to clear. Thus, the market portfolio, a convex combination of individuals' optimal portfolios, is also on the portfolio frontier. As long as the market 83
84
Foundations for Financial Econ omics
p ortfolio is not the minimum variance portfolio, Section 3 . 16 implies that there exists a linear relation among the expected asset returns with the expected return on the market portfolio as the pivotal vari able. This is the Capital Asset Pricing Model  the subject of the second part of this chapter. Finally, we will turn to the Arbitrage Pricing Model, where a linear relation among expected W!set returns holds app roximately for most of the . a ssets in an economy with a large number of assets,� when, roughly, there are no arbitrage opportunities (in the limit) . 4 . 2 . We start our formal discussion by giving a definition. A vector of asset rate of returns r = (r;)f=1 is. said to exhibit two fun d separation i f there exist two mutual funds a 1 and a 2 such that for any portfolio q there exists a scalar � such that
(4.2 . 1) for all concave u ( · ) . We shall W!sume that all assets are risky, until specified other wise. Moreover, we W!sume that asset returns have finite second mo ments and that no two asset returns are perfectly correlated, which imply that the variancecovariance matrix of asset returns exists and is positive definite. Then portfolio frontier exists, and every fron tier portfolio is uniquely determined in that there is a unique set of p ortfolio weights associated with each frontier portfolio. The above points were discussed in Chapter 3. Now suppose that the vector o f W!Set rates o f return, r, exhibits two fund separation. We first claim that the separating mutual funds a 1 and a 2 must be frontier portfolios. To see this, we note that by the definition of two fund separation, for any portfolio q there must exist a scalar � such that
(4.2.2) for all concave
u (·) .
This is equivalent to
( 4.2.3)
85
Two Fun d Separation an d Linear Valuation
From Section
2.8 w e must then have (4 .2 .4)
and
( 4.2 .5)
That is, the dominating portfolio Xo1 + (1  ..X)o2 must have the same expected rate of return as q and a smaller variance. Suppose, for example, that o2 is not a frontier portfolio. Then there must exist a portfolio & that has a variance strictly smaller than that of any portfolio formed by o 1 and o2 • This contradicts the hypothesis that o1 and o2 are separating portfolios. Hence o1 and o2 must be on the portfolio frontier. 4.3.
Next we observe that because o1 and 02 are frontier port folios, any linear combination of them is also on the frontier. Hence, for any portfolio q , the dominating portfolio formed from two sepa rating portfolios is the frontier portfolio that has the same expected rate of return as portfolio q . As the portfolio weights of any fron tier portfolio are uniquely determined and any two distinct frontier portfolios span the whole portfolio frontier, whenever two fund sep aration obtains , it must be that any two distinct frontier portfolios can be separating portfolios. In particular, we can pick any frontier portfolio, p =ft m vp and its zero covariance portfolio, zc(p) , to be the separating p ortfolios. Sections 3.16 and 3.17 allow us to write, for any portfolio q ,
rq where (3qp that
= ( 1  (3qp ) rzc(p ) + {3qprp + lqp = ;zc (p ) + (3qp ( rp  rzc (p) ) + Eqp >
(4.3 . 1)
= cov( rq , rp )jvar( rp ) and Eqp is such that E( lqp ) = 0. Note (4.3.2)
is the rate of return of the dominating portfolio. Next we claim that the necessary and sufficient condition for two fund separation is v
q.
( 4.3 .3)
86
Foundations for Financial Economics
4 . 4 . F irst we will demonstrate the n ecessity of ( 4.3.3) for two fund separation. We note that as Q (/Jqp ) is the return on the domi nating portfolio, a = 0 is a solution of the following program:
max {a}
E[u(aiq + (1  a) Q ({Jqp ) )]
A necessary condition for a = 0 to be a solution is V concave u (·) .
(4.4.1)
Suppose (4.3.3) does not hold for some q . We denote E[lqp i Q ({Jqp ) ] by m q (Q ) and the cumulative distribution function of Q ({Jqp) b y F (· ) . As E [lqp] = 0, i t follows that there must exist a real number z such that
(4.4.2) This can be seen as follows. First, the two sides of the equality of (4.4.2) can not be zero for all z, as this will contradict the hypothesis that ( 4.3.3) does not hold for some q. Second, if the equality of ( 4.4.2) does not hold for all z , that is,
then
which contradicts the fact that E[lq p] = utility function that is piecewise linear:
0. Now consider a concave if y < if y �
z; z,
(4.4.3)
87
Two Fun d Separation and Linear Valuation u,
(y)
z
0
F igure 4 . 4 . 1 : The Utility Function Defi ned in ( 4.4.3)
y
where K1 > K2 are strictly positive constants. This utility function is graphed in Figure 4.4. 1 . Then
E[u� (Q (,8qp))lqp] = E[E[u� (Q (,8qp))lqpiQ (,8qp]J = E[u� (Q (,8qp)) E[lqpi Q (,8qp)]J z oo = K 1  m q (Q)dF(Q) + K2 r+ mq (Q) dF(Q) Jz oo = (K1  K2) mq (Q)dF(Q ) i: 0,
1
l�
which contradicts (4.4.1) . Thus (4.3 .3) i s a necessary condition for two fund separation . 4 . 5 . Now we will show th at ( 4.3.3) is also sufficient for two fund separation. The separating portfolios are p and its zero covariance portfolio. If q is a portfolio, then the dominating portfolio is (1 ,8qp)rzc(p) + {3qprp . To see this, let u(·) be any concave utility function.
88
W e have
Foundations for Financial Economics
E[u(iq)J = E[u(Q (,8qp) + Eqp)] = E[E[u(Q(,8qp) + lqp)IQ(,8qp)]J � E[u(Q(,8qp))J,
where the inequality follows from Jensen's inequality because u(·) is concave. Thus , (4 .3.3 ) is sufficient for two fund separation. As we observed earlier, in this case, any two distinct frontier portfolios can be separating portfolios. Note that the condition for two fund separation can be summa rized in words based on (4.3.1) and (4 .3.3) . The rate of return on any feasible portfolio is equal to the rate of return on a particular portfo li o plus a random noise term that has a zero conditional expectation given the rate of return on the particular portfolio. This particular portfolio has the same expected rate of return as the given feasible p ortfolio and is formed from two fixed mutual funds. As expected rates of return on the two fixed mutual funds are different , portfolios formed from them can have any expected rate of return desired . 4.6. We will say that a vector of asset returns i exhibits one fund separation if there exists a ( feasible) portfolio a such that every risk averse individual prefers a to any other feasible portfolio. That is, when one fund separation obtains, there must exist a mutual fund a such that for any portfolio q, we have, for all concave u( · ),
This implies and
Var (ra )
�
Var(rq)
b y way of second degree stochastic dominance. As q i s arbitrary and a is fixed, all assets must have the same expected rate of return. Furthermore, the separating portfolio must be the minimum variance portfolio. Otherwise, pick q to be the minimum variance portfolio. Then
var (rq) va r (ra),
0, since zc(m) is inefficient. Relation (4 .10.1) is drawn in Figure 4.10.1. The higher the f3;m for asset j, the higher its equilib rium expected rate of return. The equilibrium expected rate of return on a risky asset depends upon the covariability of its rate of return with the rate of return on the market portfolio. The expected rates of return on all risky assets and on all feasible portfolios lie
93
Two Fund Separation and Linear Valuation
0
Figure
4.10.1:
A Security Market Line
/3 j m
along a line, which is termed the security market line, as shown in Figure 4.10 . 1 . When the market portfolio is a n inefficient portfolio, we still have a linear relation between E[r;] and f3im · However, the slope of the line in the E[r;]  f3;m plane is negative. In this case, the higher the covariability between an asset's rate of return with that on the market portfolio, the lower its equilibrium expected rate of return. We can rewrite (4. 10.1) as follows:
(4. 10.2) Then the higher a risky asset return's covariability with the zero covariance portfolio with ·respect to the market portfolio, the higher its equilibrium expected rate of return will b e . Figure 4.10.2 graphs relation (4. 10.2) when E [ rm]  E[rzc(m)l < 0. 4 . 1 1 . In the previous two sections, we derived a linear restric tion on equilibrium expected rates of return on risky assets when two fund separation obtains. In the analysis, it was assumed that the market portfolio was not the minimum variance portfolio. In
94
Foundations for Financial Economics
0
Figure 4 . 1 0 . 2 :
l3j z c ( m )
A Security Market Line
this section, we shall analyze a special case of two fund separation and show that the market portfolio is an efficient portfolio. Thus a relation like (4.10.1) is valid in equilibrium. Let us assume that individuals' utility functions are increasing and strictly concave. Note that monotonicity and strict concavity imply strict monotonicity. In addition, we assume that the rates of return on assets are multivariate normally distributed. We show first that, under these conditions, each individual will choose to hold an efficient portfolio. We have observed in Section 4.7 that a linear combination of multivariate normal random variables is a normal random variable. Thus the rate of return on any portfolio chosen by an individual is normally distributed. Let p denote the portfolio chosen by individual i . The expected utility of individual i's choice is ( 4 . 1 1 . 1)
where z denotes a standard normal random variable. It is clear from relation (4. 1 1 .1) that an individual's expected utility can be parame terized by the expected rate of return and standard deviation of the portfolio he chooses to hold. In other word, individuals' preferences
95
Two Fund Separation and Linear Valuation
can be completely specifi e d by their preferences for expected rates of return and standard deviations of feasible portfolios. Now we show that strict monotonicity of an individual's utility function implies that he prefers a higher expected rate of return, ceteris paribus. Defining Vi (E[rp] , u(rp)]
:=
E [ ui (Wj ( l + E [rp] + u(rp)z)) J
and partially differentiating Vi with respect to E [rp] gives: aVi (E[rp] , u(rp) ) = E[ u,� (W i) �oi ] a E [rp]
>
0'
( 4 . 1 1.2)
where
( 4 . 1 1 .3) and where the strict inequality follows from the strict monotonicity of ui (· ) . That is , individual i prefers a. higher expected rate of return, cete ris paribus.
Next we claim that the strict concavity of an individual's utility function implies that he prefers a portfolio with a lower standard deviation, ceteris paribus. To see this, we partially differentiate Vi with respect to u(rp) to get a Vi ( E [rp] , u (rp)) 0 = E[u,'. ( wi ) z �il . au(rp)  i ) , z) , = W0i Cov ( u I (W i
(4 .1 1 .4)
where we have used the defi n ition of covariance to write the second equality. Note that W i is perfectly correlated with z. By the strict concavity of u i (·) , we know that uH · ) is strictly decreasing . Hence it follows that uH W i ) and z are strictly negatively correlated, and Cov ( uH W i) , z) < 0. So a strictly risk averse individual prefers a portfolio with a lower standard deviation, ceteris paribus. Now we are ready to show that under the conditions postu lated above, individuals will choose to hold efficient portfolios. We first demonstrate that individuals' indifference curves in the mean standard deviation plane are positively sloped. Totally differentiat ing Vi with respect to E[rp] and u (rp) and setting the result equal to zero, we get
�
E[u' ( i) .z] E[ u ' (Wi)J
>
o,
(4.1 1 . 5)
96
Foundations for Financial Economics.
~
(T
!1 )
Indifference Curves when Asset Returns Are Normally Distributed
Figure
4.11.1:
which was to be shown. Finally, that individuals will choose to hold efficient portfolios is demonstrated in Figure 4 . 1 1 . 1 . The hyperbola in Figure 4 . 1 1 . 1 is the p ortfolio front ier. The three positively sloped curves are indif ference curves. The indifference curves further toward the northwest represent higher utility levels. This follows from the signs of rela tions (4.11 .2) and (4.1 1 .4) . Thus an individual will choose to hold a portfolio that is represented by the point of tangency between his indifference curve and the portfolio frontier. Then his strictly p osi tively sloped indifference curves imply that he will choose an efficient p ortfolio. Applying the above argument to all individuals, we know that all individuals hold efficient p ortfolios. Therefore, the market portfolio is an efficient portfolio, since it is a convex combination of individuals' p ortfolios. Furthermore, we have for any feasible portfolio (4.11 .6) and
(4.11 .7)
97
Two Fun d Separation and Linear Valuation
The equilibrium relations among expected rates of return on assets are as described in Figure 4.10.1. Relations (4. 1 1 .6) and (4. 1 1 .7) are known as the ZeroBeta Capital Asset Pricing Model which was developed by Black (1972) and Lintner (1969) . 4 . 1 2 . In carrying out the above analyses, we h ave assumed that there is no riskless asset. These next three sections are devoted to the case where there is a riskless asset. Suppose first that the expected rate of return on the minimum variance portfolio exceeds the riskless rate, i.e., r 1 < A/C. Let e be the tangent p ortfolio of Figure 3 . 1 8 . 1 . From (3.19.2) , we know that the rate of return on any feasible portfolio q can be expressed as:
(4. 12 . 1) with Cov(ie , Eqe )
=
E[eqe] = 0. We claim that j = 1 , 2, . . . , N
( 4.12 .2)
is a necessary and sufficient condition for { (i; )f= 1 , r 1} to exhibit two fund separation. \ We prove the necessity part first. Suppose that two fund sep aration holds. Along the line of arguments of Section 4.2, the two separating portfolios must be frontier portfolios and can be chosen to be portfolio e and the riskless asset. Therefore, when there ex ists a riskless asset and two fund separation holds, we h ave two fund monet ary separation. Arguments similar to those of Section 4.3 then show that (4.12.3) for all portfolios q . As r1 is nonstochastic and we c an always pick to be any risky asset j, it follows that
q
j = 1 , 2, . . , N, .
is implied by two fund separation when a riskless asset exists. The sufficiency part follows easily from the arguments of Sec tion 4.4 and the fact that (4.1 2 .2) implies (4. 12.3) . Thus, E[E";e l re] = 0 implies two fund separation when a riskless asset exists.
98
Founda tions for Financial Economics
Similar arguments also show that when rf
>
A/C
j = 1 , 2, . . . , N,
is a necessary and sufficient condition for two fund separation , where ( E;e• ) f= 1 are such that
4 . 1 3 . We shall assume throughout this section that two fund separation holds. When rf =j::. A/C and risky assets are in strictly positive supply, the tangent portfolios of Figures 3 . 1 8 . 1 and 3 .1 8 .2 must be the market portfolios of risky assets in equilibrium. There fore, in those two cases, the market portfolios of risky assets are on the portfolio frontier. Hence, using (3.19. 1 ) , we know
(4 . 1 3 . 1)
for any portfolio q in the market equilibrium. This is the tradi tional Capital A sset Pricing Model ( CAPM ) independently derived by Lintner ( 1 965 ) , Mossin ( 1 965) , and Sharpe ( 1 964) . When rf = A/C, the story is a little different. If this is an equilibrium, we claim that the riskless asset is in strictly positive supply and the risky assets are in zero net supply. To see this we recall that when two fund separation holds, individuals hold frontier portfolios. In addition, Section 3 . 1 8 says that when rf = A/C , an individual puts all his wealth into the riskless asset and holds a self financing p ortfolio. For markets to clear, it is then necessary that the riskless asset be in strictly positive supply and the risky assets be in zero net supply. As for equilibrium relations among asset returns, we cannot say much b eyond ( 3 . 1 9. 1 ) . 4 . 1 4 . I n Section 4 . 1 3, we were not able t o sign the risk premium of the market portfolio in the presence of two fund separation. We
Two Fun d Separation and Linear Valuation
99
will show in this section that when investors have strictly increasing utility functions, the risk premium of the market portfolio must be strictly positive when the risky assets are in strictly p ositive supply and (4.12.2) holds. We first claim that an investor will never choose to h old a port folio whose expected rate of return is strictly lower than the riskless rate when his utility function u(·) is strictly increasing and concave. Let ; be the random return of a portfolio chosen by u ( · ) such that
( 4.14.1) Note that
u( E [Wo (1 + r)]) < u( Wo (1 + r1 ) ) ,
( 4.14.2)
E [rq] = E [rzc( m.) l + /Jq m.(E [rm.]  E [rzc (m.)D
(4.14.3)
E[u( Wo ( 1 + r)) ]
�
where the first inequality follows from the Jensen's inequality and the second inequality follows from the strict monotonicity of u. Relation ( 4.14.2) contradicts the hypothesis that the individual chooses to hold the portfolio with a random rate of return ;. Given that (4. 12.2) holds and risky assets are in strictly positive supply, we know from Section 4.13 that r1 f. A/C. Suppose that r1 > A/C. Then no investor holds a strictly positive amount of the market portfolio. This is inconsistent with market clearing. Thus in equilibrium, it must be the case that r1 < A/C and the risk premium of the market portfolio is strictly positive. In this event, no individuals hold inefficient portfolios and the half line in the standard deviationexpected rate of return space composed of efficient frontier portfolios and the riskless asset is termed the Capital M arlcet Lin e in the literature. B efore we leave this section, we note that in the above proof when there exists a riskless asset, an investor will never choose to hold a portfolio having an expected rate of return strictly less than the riskless rate we only used the fact that an investor can always invest all his money in the riskless asset. Hence the conclusion holds even when borrowing is not allowed. We ask the reader to show in Exercise 4.6 that in such event, when investors choose to hold efficient frontier portfolios and the riskless asset is in strictly positive supply, (4. 11.6) and (4.1L7) are valid and E[rzc(m.)l � r ! · That is,
100
Foundations for Financial Economics
for any feasible portfolio and
( 4. 14.4 )
We also ask the reader to work out a similar conclusion when the riskless borrowing is allowed at a rate strictly higher than the riskless lending rate in Exercise 4.7. In th at case, we h ave
( 4.14.5 )
for any feasible portfolio and
( 4 . 1 4 .6 )
where rL and r11 denote the riskless lending rate and borrowing rate, respectively. The versions of the Capital Asset Pricing Model of ( 4.14.3 ) and ( 4 . 14.4) and ( 4.14.5 ) and ( 4.1 4.6 ) will be termed the constrained borrowing versions of the CAPM. The constrained bor rowing versions of the CAPM are the main subjects of discussion in Chapter 10 when econometrics issues in testing the CAPM are discussed. 4.15.
In the above analysis, we derived the traditional CAPM and the zerobeta CAPM by way of two fund separation. The risk premiums of risky assets are related to the risk premium of the mar ket portfolio in a linear fash ion. In this and the next section, we will discuss two simple situations where we can explicitly write down how the risk premium of the market portfolio is related to investors' optimal portfolio decisions. First, suppose that there exists a riskless asset with a rate of return r 1 and that the rates of return of risky assets are multivariate normally distributed. Let W; denote the optimally invested random time1 wealth for individual i. Section 1 . 18 shows that
( 4. 1 5 . 1 )
where
N
W; = wJ (l + ,, + L w;; (r;  r, )) . i= l
101
Two Fund Separation and Linear Valuation Using the definition of a covariance , (4. 1 5 . 1 ) can be written
as
( 4 . 1 5 . 2) Note that Wi and r; are bivariate normally distributed. We will employ the following mathematical result: Let Y and X be bivariate normally distributed . Then we have
Cov(g (X) , Y) = E [g' (X )] Cov (X , Y ) , provided that g is differentiable and satisfies some regularity condi tions. This mathematical result is called the Stein's lemma. Assum ing that Ui is twice differentiable and applying the Stein's lemma to (4.15.2 , we have
)
E [ u� ( Wi)] E[r1 
r1]
 E [ u�' ( Wi ) ] Cov (Wi , r;).
=
(4. 1 5 . 3
)
Defi ning the i�th investor's global absolute risk aversion
0i =
E[u?�Wi)] E[uHWi)]
,
dividing both sides of ( 4.15.3 by E[u� ( Wi )J , summing across i, and rearranging gives:
)
E[r; 
where
r1]
I
=
(L Oi 1 )  1 Cov( M, r; )
=
Wmo ( L Oi 1 ) 1 Cov (rm , r; ) , i= l
M=
i=l
I
L:: wi i =l
( 4. 1 5 . 4)
I
=
Wma ( 1 + rm ) ·
Note that 1('�[=1 9i1 )  1 is the harmonic mean of individuals' global absolute risk aversion. We can interpret Wma ( = t o;1 1 to be the aggregate relative risk aversion of the economy in equilibrium.
L;f
)
102
Found a tions for Financ ial Economics Relatio n (4 . 1 5.4) implies that
E[rm  r,]
=
I
(L o;1r 1 Wmo Var(rm ); i=l
(4 . 1 5 . 5 )
that i s , the risk premium on the market portfolio i s proportional to the aggregate relative risk aversion of the economy. Thus, the risk premium of the market portfolio is strictly positive when investors' utility functions are increasing and strictly concave. For example, when investors' utility functions are negative exponential , u i ( z) =  "t exp {  i z } , with 0,
a
a; > I
I
ci: o; 1 )  1 = CI: a; 1 )  I, i=1 i=1 a strictly positive constant . S ubstituting ( 4 .1 5 .4) into (4.1 5 .3 ) gives the familiar CAPM re lation ( 4 . 1 3 . 1 ) . However, in the present situation, we have a relation for the market risk premium ( 4 . 1 5 .5) . 4 . 16.
Now assume instead that utility functions are quadratic:
ui(z) = a;z  b;· z2
ai , b; > O.
a;/b;.
(4.16.1)
Note that individual i reaches satiation when z � We therefore assume that the rates of return of assets and the initial wealth are such that satiation will not be attained. Substituting ( 4 . 16 . 1 ) into ( 4 . 1 5 . 1) and rearranging gives
(�; 
)
E[ W;J E [r;
 r1] = Cov (Wi , r; ) ,
(4. 16.2)
where we have used the definit ion of covaria nce. Note that (� E[Wi ])  1 is equal to 0; , the global risk aversion of investor i ith a quadra tic utility fu nction. This result does not require the as sumption that asset returns are multiv ariate normally distrib uted. Summ ing (4. 16.2) over i gives
�
E[r;  r,] =
(I: :� I
i= 1
I
E [M]
)
1
Wmo Cov (rm , r; ) ,
( 4 . 16.3)
Two Fun d Separation and Linear Valuation where

103
I
"""' Wi . M=�
Relation ( 4.16.3 ) implies that
E[rm  r tJ =
(L :� I
i= l
I

i=l
E [M]
)1
W mo Var(rm) ·
( 4.16.4 )
Substituting ( 4.1 6.4 ) into ( 4.16.3 ) gives the CAPM relation ( 4 . 1 3 . 1 ) . 4 . 1 7 . I n the context of the CAPM, a risky asset's beta with respect to the market portfolio is a sufficient statistic for its contri bution to the the riskiness of an individual's portfolio. Risky assets whose payoffs are positively correlated with those of the market port folio have positive risk premiums. In such event, the higher the asset beta, the higher the risk premium. The intuition of this relationship can be understood as follows. Consider two assets A and B. Asset A and asset B have the same expected time1 payoffs. However, asset A's payoffs are positively correlated with the p ayoffs of the market portfolio, while asset B 's payoffs are negatively correlated with that of the market portfolio. That is, asset A has high p ayoffs when the economy is in relatively prosperous states, while asset B has high payoffs when the economy is in relatively poor states. One unit of the p ayoff is more valuable in a relatively poor state than in a relative abundant state. Therefore, asset B is more desirable, and its timeD price will be higher than that of asset A. Since assets A and B have the same expected payoffs , the expected rate of return of asset A will be higher than that for asset B . In other words, asset A's p ayoff structure is not as attractive as that of asset B. Therefore, it has to yield a higher expected rate of return than asset B to make itself as attractive as asset B in equilibrium. 4 . 1 8 . Recall from previous analyses that if the rates of return of risky assets satisfy ( 4.12 .1 ) and ( 4.1 2.2 ) , two fund separation implies that there exists a linear relation among expected returns of assets . In equilibrium, the coefficients of the linear relation are identified to
104
Foundations for Financial Economics
be related to the be t a's of asset returns with respect to the return on the market portfolio. Relations (4.12. 1 ) and (4.12.2) may be loosely interpreted as saying that the rate of return on an risky asset is generated by a "one factor model" plus a random noise whose conditional expectation given the factor is identically zero. Note however that the residuals from this "one factor model" are not uncorrelated across states. Also, the factor happens to be the rate of return on a portfolio. ln· _such event, the set of portfolios that will be chosen by some risk averse investor is spanned by the riskless asset and the factor. We h ave identified this set to be just the portfolio frontier ! Suppose that (4. 12.2) does not hold, but , instead, the l;e ' s are uncorrelated. Intuitively, the above mentioned spanning property may still be true approximately, as long as there are a large num ber of assets such that the le; 's can be diversified away by forming welldiversified portfolios. Then it may follow that a linear rela tion among expected asset returns will be valid approximately. The following analyses formalize this intuition and show that a linear re lation among expected asset returns will hold for most of the assets approximately if there is no arbitrage opportunity ( in the limit) and there are a large number of assets. 4 . 1 9 . We consider a sequence of economies with increasing num bers of assets. In the nth economy, there are n risky assets and a riskless asset. The rates of return on risky assets are generated by a K factor model:
i = 1 , 2 , . . . , n,
(4. 1 9.1)
where
and
E [fj] = 0, i = 1 , 2, . . . , n, E[fjlj ] = 0 if l =P j", (72
(fj )
:::;
ij2
i = 1 , 2 , . . . , n,
(4. 19.2) ( 4. 19.3) (4.19.4)
Two Fund Separation and Linear Valuation
105
.w here o 2 is a fixed strictly positive real number and f3'J�c are real num bers . We also assume that the 8r•s are rates of return on portfolios. Using matrix notation, we can write (4.19 .1) as
(4. 19.5) where a n is a n N X 1 vector of aj 's, a n is a n n X K matrix whose elements are {3'/J,. , j = 1 , 2, . , n , k = 1 , 2, , K, and 8 n is a Kvector of 8ts. To avoid degenerate cases, we will only be interested in economies with more than K risky assets. That is, an implicit assumption to be made henceforth is that n > K . .
4 .20.
.
...
We first note that if Vj
(4.20.1)
then there exists an exact linear relation among expected rates of return on assets in the nth economy, if one cannot create something out of nothing. This follows since the returns on risky assets are completely spanned by the K factors (portfolios) and the riskless asset. Formally, consider a portfolio of the K factors and the riskless asset, yj, with K
Y'/o = 1  L f3'J�c ,
Yn
�n
;1c = �'ilc •
lc = l
k
=
1 , 2,
.
.
. , K,
where Y'/o is the proportion invested in the riskless asset and Y'j1c is the proportion invested in the kth factor. The rate of return on this portfolio is
(1 
K
K
lc=l
lc= l
L f3'J�c ) rJ + L P'J�c 8i:.
Note that the factor components of the rate of return on Yj replicate those of asset j.
106
Foundations for Financial Economics
We claim that aj must be equal to (1  Ef=l ! · We shall show that if this is not the case, something can be created out of nothing. Suppose that
f3'J�c )r
a
'J
(1  E f=l
Pj.�.) r J , a
get
K
'J =· ( 1  L:; Pj.�:)r J .
( 4.20.2)
k= l
Substituting (4. 20.2) into (4. 19.5) and taking expectations, we (4.20.3)
where 1 n is a n n X 1 vector of ones. That is, there exists an exact linear relation among expected asset returns. When the fj 's are not zeros, the story is a little compli cated . Formally, in economy n, a portfolio of the n risky assets and the riskless asset is an arbitrage portfolio if it costs nothing. An ar bitrage opportunity ( in the limit ) is a sequence of arbitrage portfolios whose expected rates of return are bounded below away from zero, while their variances converge to zero. Roughly, an arbitrage opportunity is a costless portfolio in an economy with a large number of assets such that its expected rate of 4.21.
107
Two Fund Separation and Linear Valuation
return is bounded below away from zero while its variance is negli gible. That is, it is almost a free lunch. We wish to show that if there is no arbitrage opportunity, then a linear relation among expected asset returns will hold app roximately for most of the assets in a large economy. We first claim that,
a'J
F:l
(1

K
L P'J�e ) r
le =l
1
(4.21 . 1 )
for most of the asset in large economies, where F:l denotes approxi mately equal to. To see this, we fix £ > 0, however small . Let N(n) be the number of assets in the nth economy such that the absolute value of the difference between the two sides of (4 .2 1 . 1 ) is greater than £. Without loss of generality, assume that K
I a']  ( 1  L Pj�e ) rJ I
le =l
�
£,
j = 1, 2, . . . , N(n) .
(4.21 .2)
If we can show that there exists N < oo such that N(n) � N for all n, we are done. This follows since there will be at most N assets that satisfy (4. 2 1 . 2) for arbitrarily large n and £ can be arbitrarily small. We sh all proceed by contraposition. Suppose that there does not exist a finite N such that N(n) � N for all n. Then there must exist a subsequence of { n = K + 2, K + 3 , . . . , }, denoted by, {nl } , such that N(nl)  oo as nl  oo We construct a sequence o f arbitrage portfolios as follows. First, construct N(nl) arbitrage portfolios that have no factor risk as we did in Section 4.20 for risky assets j = 1, 2, . . . , N(nl) in economy nl . The rate of return for the jth arbitrage portfolio is .
laj' where
s�''
=
{
+1 1
(1 
K
L PjD ri i + sj' fj' ,
(4.21.3)
k =l
if a 'J'  ( 1 if aj'  (1
 Ef= 1 PjD 
Ef= 1 PjD
>
o; < o.
108
Foundations for Financial Economics
Next, form a portfolio of these arbitrage portfolios with a con stant weight, 1/N(nz ) , on each. The resulting portfolio is still an arbitrage portfolio with an expected rate of return
1 N( nz)
N ( nt)
?;
i aj1  ( 1 
K
£; .Bj� ) rJ i
� e
> O,
( 4.21 .4)
and a variance N (nt)
2 1 q ""' u 2 ( e""'' ) =N==2::(i � N ( nz ) ' nz:)
�
(4.21 .5)
where we have used (4.19.2) , (4.19.3) , and (4.19.4) . Since N(n1 ) + oo as l + oo , the variances of the sequence oi arbitrage portfolios converge to zero, while their expected rates of return are bounded below away from zero by e . Thus there exists an arbitrage opportunity, a contradiction. We hence conclude . t hat there must exists a finite N such that N ( n ) � N for all n . 4.22. Now using the fact proven above that for a given however small, there exist a t most N risky assets such that
i aj'  ( 1 
e' >
0,
K
L ,Bj� ) rJ I � e, k= l
and (4. 19.1 ) , we k now K
I E [rj ]  r 1  L ,BjA: ( E[Sf]  rJ ) i
� e
k=l
for all but at most N risky assets in any economy. Thus, for economies with the number of assets much larger than N, a linear relation among expected asset returns holds approximately for most of the assets. This relation is the Arbitrage Pricing Theory {APT} origi nated by Ross (1976).
Two Fun d Separation and Linear Valuation
109
4 . 2 3 . I n the above derivation of the APT, a linear relat i on among the expected rates of return on the risky assets holds approx imately for most of the assets when the economy is large. On the other hand, for any given asset, the deviation of its expected rate of return from the APT relation might be very large. For the APT to make predictions in an economy with a fi n ite number of assets, we would like to bound the deviation from the APT linear relation for any asset. To achieve this goal, we will make assumptions on the structure of the returns on risky assets that are somewhat different than those used in the above derivation of APT. Equilibrium rather than arbitrage arguments will be used in the following derivation. The resulting relationship among risky assets is sometimes called equilib rium APT. Suppose that there are N risky assets in the economy, indexed by j = 1 , 2 , . . . , N, and a riskless asset. These risky assets are in strictly positive supply. The rate of return on the riskless asset is r1 , and the rates of return on risky assets are generated by a Kfactor model like (4.19 . 1 ) :
i;
with
K
= a; +
L, �;1r.81r. + f'; , lr.=l
E[f';]
=
0
j = 1 , 2 , . . . , N,
and f'; �  1 .
( 4 . 23.1 )
( 4.23.2)
We also assume that the random variables
(f'l , . . . ' fN , 81 , . . . ' SK )
are independent and that 61r. are rates of return on portfolios. Note that in Section 4 . 19 we only required that f'; and !1 be uncorrelated if j f. l. Here a stronger independence assumption is used. Also, since we are not considering a sequence of economies, we do ncit need an assumption like (4 . 19 . 4) . Agents in the economy have utility functions that are increasing, strictly concave, and three times continuously differentiable, We also assume that agents' ArrowPratt measure of absolute risk aversion Ai (z) =  u�' (z) /u' (z) is bounded from above by a constant A for all > 0. i and that u'!' '
1 10
Foundations for Financial Econ omics
Recall from Section 4.20 that, when
a; =
K
(1

or equivalently,
l;
= 0,
L f3;�c)r J i lc= l
K
E [r;] = r1 + L f3;�c (E[8�c]  r 1 ) . lc=l
(4.23.3)
Our purpose here is to bound the deviation of E[ r;] from the right hand side of ( 4.23. 3 ) , in market equilibrium, when l; is not identically zero . 4.24.
Consider a p ortfolio of the K factors and the riskless asset having a rate of return r1
+
K
L f3;�c (8Jc  rJ) ·
lc= l
This p ortfolio is constructed in the same way as Section 4.20. Next form an arbitrage portfolio by investing one dollar in the above port folio and shorting one dollar 's worth of risky asset j. This arbitrage portfolio has a rate of return r = (1 
K
L f3; �c )r1  a;  l; . lc=l
(4.24.1)
Let Wio be individual i ' s initial wealth, and let Wi be individual i's time1 random wealth. Then a = 0 is the unique solution to the following problem: max a
E [ui (Wi + a r) ] ,
as Wi + a r is a feasible random wealth for individual i. The first order necessa �y condition for a = 0 to be an optimum is
111
Two Fund Separation and Linear Valuation
Using the definition of covariance, the fact that E [E';] (4.24. 1 ) , the abo v e relation can be written as
0, and
( 4.24.2)
Now we claim that, for all j such that E'; i 0, (1
K

L f3;�c )r1 lc=l

a;
0. The above argu ments can be applied to all j such that E'; i 0. Thus we have that (4.24.3) holds for all such j's. As a consequence of the above analysis, we can also conclude that if E'; i 0, then asset j is held in a strictly positive amount by all individuals. To see this, we note that ( 4.24.2) holds for all assets and all individuals. If E'; i 0, we must have ( 4.24.3 ) , which in turn implies that ( 4.24.4) must hold for all i. This is p ossible only if every individual holds a strictly positive amount of asset j, as E'; is independent of all the other random variables by assumption . 4 . 2 5 . Now fix risky asset j with E'; i 0. Let 01 ; be the dollar amount invested in risky asset j by individual i. From Section 4.24, we know that 01; > 0. Define
112
Foundations for Financial Economics
The L agrange remainder form of Taylor's expansion gi v es
E[ u� (Wi) f'; ] = E[ uHWt ) €;] + a;; E[ u �' ( e) f'J] = E [ u � (Wt) ]E[f'; ] + a ;; E[ u ? ( {) f] J = a;; E [ u; ( e) E;2 ] ,
(4.25 . 1)
II
where t h e second equality follows from the assumption that €; and all the other r andom variables are independent and where e is a r andom variable whose value lies between W; and W;* . By the assumption that f'; 2:  1 , we know > ..� 
Therefore ,
I E [ u �' ( e) f'J ] I
�
 .• w I
E[
 ,. ."
s up
z?.W;"  a ;;
We claim that sup
z?. W;"  a;;
l u�' (z) l
�
To see t his, we first note that since sup
z?. W;"  a ;;
�A
'"1]
lu�' (z) I €]] .
u' (Wt ) Ae.Aa; ; .
( 4.25.2)
( 4.25.3)
u �' (z)/uHz) � A ,
lu�' (z) l
sup z?.W;"  a ;;
uH z) � AuHW;"  a;;),
where the second inequality follows from the concavity of we observe that
ln(uHW/  a;;)) = ln uHW/ )  f � ln uHwn +  •) + = ln u;I (W;
�
w�
wt a; ; w·�
f
wt  a ;;

Aa;;.
( 4.25.4) u ;.
u?(z)/uHz) dz A dz
Next,
113
Two Fund Separation and Linear Valuation ' (  *) Ui' ( wi e A a I·J· ,  *  CJ.i j ) � U i wi Relations (4.25.4) and (4.25 .5) imply (4.25 .3). Now substituting ( 4.25.3) into ( 4.25.2) gives
Therefore,
I E[ uf ( e)£j] l � .AeAa;i E[uHWt)£j] � .Ae A a ; i E[ uHWt )]Var(£;) . that , since u f' � 0, by the law of iterative
Note also and the conditional Jensen's inequality, we have
E [uHWi )]
=
=
�
(4 .25.5)
( 4 .25 .6) exp ectations
E [uHWt + Cl.ij fj )] E[ E[ uHWt + CJ.i;l;)IWtJ E[ uHWt )] .
(4.25 .7)
Finally, subtituting (4 .25 . 1 ) , (4.25.6) , and (4.25.7) into (4.24.2) gives
Ia; or equivalently,
(1 
K
L P;A: ) rJ I � .Ae.Aa;iVar( f"; ) Cl.ij; A: = l
K
I E [ r;]  r1  L P;A: ( E[ 5A: ]  r/ ) 1 � .AeAa;iVar(f";) a.ii · A: = l
(4.25.8)
Let the total market value of asset j be denoted by S; . We claim that in equilibrium there exists an i such that Cl.ij � S ;/I, where I is the number of individuals in the economy. Suppose that this is not true, that is Cl.ij > S;/ I for all i. Then we must have
sj =
I
L Cl.ij i =l
>
S;
a contradiction . Since ( 4.25.8) holds for all i, we thus have K
I E [r; ]  r1  L P;A: ( E [5A:]  rJ) I A: = l
�
Ae.A 8i11Var(£;) S; /I .
(4.25 .9)
Foun dations for Financial Economics
1 14
Relation ( 4.25.9) gives an explicit bound of the dev iation of E[ r;] from the APT relation . The bound is small when B ;/ I, .A, or Var ( l;) is small, ceteris paribus . This bound can be improved by tightening the lower bound of €; , which readers are asked to do in Exercise 4.9.
Exercises
4 . 1 . Derive the zerobeta CAPM under the assumption that asset returns are multivariate normally distributed. Also derive the zerobeta CAPM under the assumption that investors' utility functions are quadratic. 4 .2 . Let the random time1 payoff of a risky security be y, and let B11 be its time 0 equilibrium price. Suppose that the CAPM holds and let the security's beta be denoted by {3,.. . Show that
B11
=
E [y] 1 + r1 + f3 11m (E [im] E[ y]  ,p• P 11m u ( Y ) 1 + r1
r1)
where and
Cov ( Y , im ) u(y)u(im ) · 4.3. Suppose that the CAPM holds. Let y and z be the random time 1 payoffs of two securities with time 0 prices B11 and B �� respec tively. Show that the equilibrium price for a security having a time 1 payoff y + z is equal to B11 + B� . 4.4. Consider an economy with N risky assets having independent and identically distributed rates of return. Show that there is one fund separation. 4.5. Suppose that there are five state of the nature denoted by Wn n = 1, 2, . . . , 5, each of which is of equal probability. Consider two risky assets with rates of returns iA and is as follows: Py m
=
115
Two Fund Separation and Linear Valua tion
4.6.
4.7.
4.8. 4.9.
w2 0.5 0.8
Wt 0.5 0.9
state TA rs
wa 0.7 0.4
W4 0.7 0.3
ws 0 .7 0.7
Suppose that assets A and B are the only two assets in the economy. Do we have one fund sep aration? Explain your answer in detail. If your answer is no, how could the numbers in the above table be rearranged to get one fund separation? Suppose that there is a riskless asset in strictly positive supply and investors prefer to hold efficient frontier portfolios. Borrow ing at the riskless rate 'I is prohibited . Show th at (4. 1 1 .6) and (4.1 1 .7) hold and E [r.rc (m )] � 'I · Suppose that investors in the economy would like to hold efficient frontier portfolios and that the riskless borrowing rate is strictly higher than the riskless lending rate, that is rb > 'L · Describe the viable positions of rb , TL , and E [rmup ] in equilibrium. Show that (4. 1 1 .6 and (4. 1 1 .7) hold and rb � E [i.rc ( m) l � ' L · Let K = 1 in ( 4.23.1 . Under the independence assumption made there, does two fund separation exist? In the context of Sections 4.234.25, assume instead that €; �  1/ S; , where Sj denotes the total market value of asset j. Show that
)
I E [r;]  'I .:..
)
K
l:: .B;�: (E [S�: ]  'I) I � .AeAf 1 Var(l; ) S; /1. 1: = 1
Remarks. For discussions on Kfund separation in general see Ross (1978) . Our treatment of two fund separation is taken from Litzen berger and Ramaswamy (1979) . Note that ( 4.3.3 ) is in general weaker than E [E'qp l rp , i.rc (p) ] = 0. Nielsen (1986) has shown that they are equi valent under certain reg ularity conditions . For the existence of a general equilibrium of the CAPM sort, we refer readers to Nielsen (1985 ) . The discussion in Section 4.14 is adapted from Merton ( 1982) . For a derivation of the
1 16
Founda tions for Financial Economics
Stein's lemma and its original application to finance, see Rubinstein ( 1976) . The derivation of the arbitrage pricing relation is different from Ross ( 1976). We made a simplifying assumption that the factors themselves are portfolios. Also, our definition for an arbitrage op portunity is slightly weaker. Huberman ( 1983) has a very nice proof of the APT when factors are not p ortfolios, to wh ich we refer in terested readers. For more recent developments of the AP'P, see Chamberlain and Rothschild ( 1983) , Chamberlain ( 1983) , and Reis man ( 1987a, 1987b, 1987c) . Connor (1984) is the first to discuss APT relation using equilibrium arguments. The discussions on the explicit b ound of the pricing deviation in finite economies are a com bination of Dybvig ( 1983) and Grinblatt and Titman (1983) . For a review of recent developments of APT, see Connor (1987) .
References
Black , F . 1972. Capital market equilibrium with restricted borrow ing. Journal of Business 45:444454. Chamberlain, G. 1983. Funds, factors, and diversification in arbi trage pricing models. Econ ometrica 50: 1 3051324. Chamberlain, G . , and M. Rothschild. 1983. Arbitrage , factor struc ture ' and meanvariance analysis on large asset markets. Econometrica 50: 128113 04. Connor, G. 1984. A unified beta pricing theory. Journal of Economic Theory 34:133 1 . Connor, G . 1987. Notes o n the arbitrage pricing theory. I n Fron tiers of Financial Theory. Edited by G . Constantinides and S . Bhattacharya. Rowman and Littlefield. Totowa, New Jersey. Dybvig, P. 1983. An explicit bound on individual assets' deviations from APT pricing in a finite economy. Journal of Fin ancial Eco nomics 1 2 :483496. G rinblat t, M . , and S . Titman. 1983. Factor pricing in a finite economy. Journal of Financial Economics 1 2:497507.
Two Fund Separation and Linear Valuation
117
Huberman, G . 1983. A simplified approach to arbitrage pricing the ory. Journal of Economic Theory 2 8 : 1 9831991 . Lintner, J . 1965. The valuation of risk assets and the selection of risky investments in stock portfolios and capital budgets. Review of Econ omics and Statistics 47:133 7 . Lintner, J . 1969. The aggregation o f investor's diverse j udgements and preferences in purely competitive markets. Journal of Fi nancial and Quantitative Analysis 4 :346382. Litzenberger, R., and K . Ramaswamy. 1979. On distributional re strictions for two fund separation. TIMS Studies in the Manage ment Science 1 1 : 99107. Merton, R. 1982. On the microeconomic theory of investment un der uncertainty. In Handbook of Mathematical Economics, Vol. J/:601669. Edited by K. Arrow and M. Intriligator. Mossin, J. 1 966. Equilibrium in a capital asset market. Econometrica 35 :768783. Nielsen , L. 1985. Risktaking and capital market equilibrium, Un published Ph.D. thesis, Harvard University. Nielse n, L. 1 986. Mutual fund separation , factor structure and ro bustness, unpublished manuscript, Department of Economics, University of Texas at Austin. Reisman, H. 1987a. A general approach to the arbitrage pricing the ory. Mimeo. Department of Finance, University of Minnesota. Reisman, H. 1987b. The arbitrage pricing theory with conditional information. Mimeo. Department of Finance, University of Min nesota. Reisman, H. 1987c. Intertemporal arbitrage pricing theory. Mimeo. Department of Finance, University of Minnesota. Ross, S. 1976. Arbitrage theory of capital asset pricing. Journal of Economic Th eory 13:34 1360. Ross, S. 1978. Mutual fund separation in financial theory: The sep aration distributions. Journal of Economic Theory 1 7 : 254286. Rubinstein, M. 197 4. An aggregation theorem for securities markets. Journal of Financial Economics 1:225244. Sharpe, W. 1964. Capital asset prices: A theory of capital market equilibrium under conditions of risk. Journal of Finance 19:425442 .
CHAP TER 5 ALL O C ATIVE E F F IC I E N C Y AND THE VALUAT I O N O F S TATE C O NTINGENT S E C U RITIES
5 . 1 . In Chapter 4 , uncertainty in the economy was character ized by distributions of returns on assets. However, it was noted in Chapter 1 that the primitive source of uncertainty is the uncertain states of nature , the collection of which is denoted by n with generic elements w . Recall that a state of nature is a complete description of a possible realization of the exogenous uncertain environment. V ery generally, an individual's primitive objects of preference are consumption in different states of nature. A state contingent consumption c l aim is a security that pays one unit of the consump tion good when one particular state of the world occurs and nothing otherwise. A state contingent claim is an elementary claim . Exist ing assets, however, may be viewed as complex bundles of elementary claims. For example, a riskless asset is a bundle of one of each state contingent claim. In a pure exchange single good economy, the ag gregate endowment of the consumption commodity is also a complex bundle of elementary claims. The first part of this chapter examines 119
120
Foundations for Financial Economics
the role of financial markets in allocating resources among individuals and the types of securities needed to achieve an efficient allocation. In allocationally efficient financial markets, security prices may be described in a simple way. We will show that when the alloca tion of state contingent claims is efficient and individuals have time additive stateinde p endent utility functions, prices in the econ omy are determined as if there were a single individual in the economy endowed with the aggregate endowment. This representative agent has a timeadd itive stateindependent utility function . The utility function of the representative agent will in general depend upon the distribution of the initial wealth of individuals. As a consequence, the prices in the economy will in general depend on the distribution of the initial wealth among individuals. This chapter also examines the necessary and sufficient conditions on indiyi�_uals' utility func tions for the prices in the economy to be determined independently of the distribution of initial wealth. Utility functions that exhibit this property are said to have the aggregation property . 5 . 2 . Consider a twoperiod pure exchange economy under un certainty with a single perishable consumption good in both periods. Individuals choose their consumption for today, say time 0, and state contingent claims on consumption for tomorrow, say time 1. There is uncertainty about which state of the world will occur at time 1. Assume that individuals h ave utility functions for the consumption good that are strictly increasing, strictly concave, and differentiable. Without loss of generality, the single consumption good is used as the numeraire throughout.
An allocation of state contingent consumption among in dividuals, denoted by { (c;0 , c;w , w E O); i = 1 , 2, . . . , !} , is feasible if 5.3.
I
L ew = Co and
i=l
121
Valuation of State Con tingen t Securities I
L i=l
c;..,
=
C.., Vw E n,
where c; o denotes individual i 's time0 consumption allocation, c,.., denotes individual i's time1 consumption allocation in state w , C0 denotes the aggregate time0 consumption available, and C.., denotes the aggregate time1 consumption available in state w . An allocation of state contingent consumption claims is said to be Pareto optimal or Pareto efficient if it is feasible and if there do not exist other allocations which are feasible and can strictly increase at least one individual's utility without decreasing the utilities of others. For example, an allocation that gives a single individual all the consumption available and others nothing is a Pareto optimal allo cation. This example illustrates th a t Pareto Of!timality is a weak criterion and does not provide a framework for addressing many so cial choice issues. 5 . 4. From the classical second welfare theorem (see, e . g . , Varian (1978)) , we know that corresponding to every Pareto optimal allo cation, there exist a set of nonnegative numbers, { �;}{= l• such that the same allocation can be achieved by a social planner maximizing a linear combination of individuals' utility functions using { �;}[=1 as weights, subject to resource constraints:
s.t. and
I
l: c;.., = Cw i= l
Vw E O
I
L c;o = Co , i= l
where 1fiw is the ith individual's subjective probability assessment about the o ccurrence of state w and u;.., (· , ·) is the ith individual's
122
Foundations for Financial Economics
utility function for timeD consumpt ion and for time1 statew con sumption . We will only be interested in Pareto optimal allocation s that correspon d to strictly positive weighting s \ , so all the >./s to appear are strictly positive . Forming the Lagrangian gives L max {c;0 ,c;w ; w Efl}f= 1
=
I
L >.; [ L 1riwUiw ( c;o , C;w )] w Efl
i=l
[
+ ¢o Co 
I
[
]
L c;o + L r/>w Cw wEfl i= l

I
L C;w ] . i=l
By the assumption that utility functions are strictly concave and the fact that the weights, {>. i }{= 1 , are strictly positive, the first order conditions for the above programming problem are necessary and sufficient for a global optimum. They are >.;
L 11"iwa u;w (cw , c;w)facw = ¢o
i
=
w Efl >. ;1r;w a Uiw ( c;o , c;w ) f ac;w = ¢w w E I
L Cjw = Cw i=l
and
Substituting relation vidual gives
1, . . . , I, 1l i
=
(5.4.1) 1, . . . , /.(5.4.2)
Vw E 1l
(5.4.3)
I
l: c;o = Co . i=l
(5.4.4)
(5.4.2) into relation (5.4.1) for the same indi
1r;wa u;w (cw , c;w ) fac,w Ew E n 1riw au;w (c;o , c;w )fac;o
=
¢w ¢o
w E 1l i
=
1, . . . , /.
(5.4.5)
It follows from (5.4.5) that a feasible allocation of state contin gent consumption is Pareto optimal if and only if, for each state, marginal rates of substitution between present consumption and fu ture state contingent consumption are equal across individuals.
123
Valuation of State Contingent Securities
5 . 5 . A Pareto optimal allocation can be attained in a com petitive economy if there exist a complete set of state contingent consumption claims . To see this, let rf>w denote the price at time 0 of state contingent claim that pays one unit of consumption at time 1 if and only if state w is the true state. An individual's problem is
{c;0 ,c;.,, w En}
max
s.t. c;o +
L
w En
1r
iwUiw ( c;o , C iw )
L r/>wC iw = eiO + L rf>weiw >
wEn
wEn
where e;o and e;w denote individual i's endowments for time0 con sumption and for time1 statew consumption, respectively. We as sume that an individual's endowment is such that his wealth at time 0 is strictly positive. Forming the Lagrangian gives max
{c;o ,c;., ; w En}
L
=
L 11"iw Uiw (c;o, ciw)
w En
[
+ 0; e;o  C iQ +
L rf>w ( eiw  C;w )
w En
].
The first order conditions, which are necessary and sufficient for an optimum are
L 11"iwa u;w (c;o , ciw )/ac;o = 0; ,
(5.5.1)
w En
11"iwa u;w (cio , ciw) faciw = O; rf>w
Vw E O,
(5.5.2)
and the budget constraint for which 0; is a shadow price. Note that 0; > 0, as an individual's utility functions are strictly increasing. Now substituting (5.5 1) into (5.5.2), we get .
11"iwa uiw (cio , c;w) /ac,w = r/>w Ew En 11"iwaUiw (c; o, c;w )/aciO
w E O.
(5.5.3)
In market equilibrium, (5.4.3) and (5.4.4) are satisfied. Now set r/>0 to be 1 and �i to be Oi1 as in Section 5.4. It follows that an allocation in
124
Founda tions for Financial Economics
a competitive economy with a complete set of state contingent claims (and a spot market for time0 consumption) satisfies (5.4. 1)(5.4 .4) and thus is Pareto optimal. Conversely, to achieve a Pareto optimal allocation corresponding to a competitive equilibrium allocation, the utility weight that the social planner assigns to individual i, �1 , is equal to fli 1 > 0. When we have a complete set of state contingent claim markets, we say that the markets are complete. We sometimes refer to those elementary claim prices to be state prices . For example, ¢ w is the state price for state w . I n actual securities markets, w e do not observe state contingent claims but rather a number of complex securities such as common stocks of firms. A complex security is a bundle of state contingent claims. The efficiency of an equilibrium in this case for ar bitrary preferences depe nds on whether or not the number of linearly independent securities equals the number of states. Suppose that we have an economy with one unit each of N (com plex) securities indexed by j = 1 , 2 , , N. Security j pays off x; w units of the consumption good at time 1 in state w E fl. We will sometimes simply use x; to denote the random time I p ayoffs of se curity j. Individuals have utility functions as in Sections 5.4 and 5.5, and their endowments are in the form of securities and time0 con sumption. We assume that individual i is endowed with a;; number of shares of security j and eiO unit of time0 consumption. An in dividual buys the time0 consumption good in the spot commodity market and buys future state contingent consumption through the securities markets. We will henceforth refer to this kind of economy as a securities markets economy . Letting O.ij and S; denote the number of shares of security j held by individual i in equilibrium and the price for security j at time 0, respectively, { ai ; i i = 1 , 2, . . . , N} is the solution to the following 5.6.
.
.
.
125
Valuation of State Con tingent Securities
program:
s.t.
CiQ
+
N
L a,;S; = eio + L ai;S; .
(5 .6 . 1 )
i =l
j
The first order conditions, which are necessary and sufficient for individual i's portfolio problem , are j = l , . . . , N,
(5.6.2)
where we have used the fact that
N
Ciw L aij:Z:jw · =
i= l
Each individual i will adjust his time0 consumption and security holdings such that (5.6.2) is valid. This does not , however , indicate that marginal rates of substitution between present consumption and state contingent consumption are identical across individuals as re quired for a Pareto optimal allocation.
If the number of linearly independent securities is equal to the number of states of nature, then markets are complete because any state contingent claim can be created by forming portfolios of existing securities. To illustrate this, let 1 0 1 denote the number of states of nature. Relation (5.6.2) may be written in matrix notation: 5. 7.
(
:Z:
:Z:
Jw 2w
l>
1
'
:Z: Nw ' 1
•
•
•
•
•
•
•
'
· '
·'
:Z:J w l n l :Z:2wl n l :Z: Nw j n j
)
'IT .
il u ;.., (c;0;c;..,101 )/il c ;.., 101
• w1n1 E
... e o w;..,r9u;.., (c;o ,c;.., ) / c3 c;o
}
(5.7 . 1)
126
Foundations for Financial Economics
If the number of linearly independent securities is equal to the number of states of nature and we include only linearly independent securities in (5.7 . 1 ), then the first matrix on the lefthandside of relation (5.7.1) is invertible. We can therefore invert it and solve for the ratios of marginal utilities between time1 statew and timeD consumption. Note that the matrix being inverted and the vector on the righthandside of (5.7.1) are both independent of individual indices; therefore: a uiw (cm , Ciw ) faciw w a uiw ( ci o , ciw )fa cio •w Ew En 11"i
1r ·

mw ,
.
a =
1 , 2 , . . . , I,
(5.7.2)
for some constants mw , w E 0. The relations in (5.7.2) are the nec essary and sufficient conditions for a feasible allocation to be Pareto optimal. Also, it follows from (5.5.3) that mw must be equal to ¢w , the state price for state w. Moreover , state contingent claims can be created by forming portfolios of the complex securities. For example, the state contingent claim that pays one unit of consumption in the first state w1 and nothing otherwise is formed by holding a; shares of security j with
(a1 , a2, · · · • a.N)
=
( 1 , 0, . . . , 0)
(
:Z: l w1 • :Z: 1 w 2 , :Z:2 w 1 ' :Z:2w2 '
• • • • • •
•
:Z: l w 1n1 :Z: 2 w lfll
•
)

l
( 5.7.3) for price state the be ust The cost of the above portfolio should j state 1 , which can be easily verified. From the previous analyses, it should be clear that securities markets in which the number of linearly independent securities is equal to the number of states are complete markets and a Pareto Optimal allocation of state contingent cl aims is achieved with indi viduals trading complex securities. :Z:Nw n :Z:Nw 2 '
• • •
•
:Z:N w 1n 1
A common feature of complete markets economies and securities markets economies is that the values of securities are addi tive. That is, if :z:w , Yw , and Zw are statedependent payoffs of three complex securities with Zw = :Z:w + Yw , then the value of z is equal to 5.8.
Valuation of State Con tingent Securities
127
the sum of the values of :z: and y . To see this, we note that when there exists a complete set of state contingent claims, a complex security having payoffs Zw can be viewed as a portfolio of state contingent claims consisting of Z w shares of the state contingent claim that pays off in state w . Thus the market value of that security must be equal to the cost of the portfolio that offers the identical state contingent payoffs , which is (5.8.1) r/Jw :Z:w . wEn If relation (5.8.1) were not satisfied , it would be possible to create something out of nothing, which is inconsistent with an economic equilibrium. For example, if the market value of the security were strictly greater than (5.8 . 1 ) , an individual could short sell the security and purchase Zw shares of state contingent claim for state w , for all w E 0. The payoff of this strategy is identically z e r o at time 1 , while its cost at time 0 is strictly negative! That is, something has been created from nothing. As a consequence, nonsatiated individuals will take unbounded positions in the strategy, and markets cannot clear. Relation (5.8.1) implies that security values are additive. That is, if Z w and Zw are statedependent payoffs for two complex securi ties, then the val u e for Zw = Z w + Yw is equal to the sum of values for :z: and z: ¢w Yw wEn = r/Jw:Z:w + r/Jw zw . w En wEn Next consider a securities markets economy. Let :z: , y, and z be three complex securities traded with Zw = :Z: w + Yw · Relation (5.6.2) implies that the value of z, denoted by s. , is au;w (c;o , Ciw )fac;w " Sz = L..., Zw 7r;w ""' L..J w en 7r ; w au;w ( c;o , c; w )fa c ;o wEn
L
L
L
L
a u;w (c;o , Ci w )fac;w :Z: w 7r iw ""' L..J w E n 7r ;w au; w ( c;o , c;w ) fac;o w En
" = L...,
" 'Kiw ""' au;w (c;o , Ciw)fac;w + L..., Yw L..J w en 7r ;w au; w ( c ;o, c; w )fa Cio w En = Sz + S11 ,
128
Foundations for Fin ancial Economics
where we have used S"' and 811 to denote the value of x ,.,' s and Yw 's, respectively. Hence, values are also additive. The additivity of values is a consequence of the requirement that in a competitive equilibrium someth ing cannot be created from nothing. If 811 is not equal to Sz + S11, say 811 > Sz + S11 , then one can buy complex securities Xw and Yw and short sell complex security Zw · The cost of this investment is equal to Sz + sll  s" , which is strictly negative. However, the time1 payoffs of this investment are uniformly zero. Hence something has been created from nothing, which is inconsistent with an economic equilibrium. 5 . 9 . An important application of the above idea is the Modigliani and Miller theorem  the values of firms in the same risk class are determined independently of their capital structure in a frictionless mark e t . Two firms are said to be in the same risk class if their time1 random payoffs are proportional across states of nature. To see this, consider two firms with time1 payoffs x1 and x2 , respectively. Assume that there exists an a > 0 such that axlw = X2w I Vw E n. Firm 1 is 100 percent equity financed and firm 2 is partly financed by, say, debt. We claim that the value of firm 2 is a times the value of firm 1 independent of the capital structure of firm 2. Let the time1 payoff of the equity of firm 2 be denoted by y and that of the debt be denoted z. We have
axlw
=
X2w
=
Yw + Zw Vw E n .
The value of firm 2 is just the sum of its equity value and its debt value. It then follows from the arguments of previous paragraphs that aS1 = S2 = S11 + 811 • Thus the values of firm 1 and firm 2 are proportional  independent of the capitalization of the firms. Here we note that the Modigliani and Miller theorem does not say that the value of a firm will be unchanged after a recapitalization. For example, when m arkets are incomplete and when a recapitaliza tion of a firm increases the number of linearly independent securities in the economy, the original equilibrium may be upset and . new equi librium prices will form . Under the new set of equilibrium prices, the value of the firm may be different. Even when the recapitalization does not increase the number of linearly independent securities, it
Valuation of State Contingent Securities
129
may involve buying back some of the existing ou t standing equ i t i es . This will cause individuals to rebalance their p ortfolio holdin gs. Un less there exists a unique equilibrium, the equilibrium prices after the recapitaliz ation may be different and the value of the firm may also be different. 5 . 10. One can see, by casual empiricism , that in actual s e curi ties markets the number of corporate securities is less than the num ber of the states of nature. Hence we would not in general be able to create a complete set of state contingent claims by forming portfolios of existing corporate securities. Thus, the markets for corporate se curities are unlikely to span all state contingent claims. Under some conditions , the corporate securities markets together with markets of options written on corporate securities are sufficient for reaching allocational efficiency for arbitrary preferences. A European call option written on a common stock is a financial security that gives its holder the right to purchase the underlying stock at a prespecified price on a prespecified date. Similarly, an European put option gives the right to sell the underlying as s et at a prespecified price on a prespecified date. The prespecified price is called the strike price or the exercise price; and the prespecified date is the expiration date or maturity date. Assume, for the time being, that there exists a portfolio of cor porate securities whose time1 payoffs, w E 0}, are strictly pos itive and separate states in that =/= Xw• if w =/= w 1 • This portfolio will henceforth be referred to as a state index portfolio . For ease of exposition, we assume that Xw; < Xwi if i < ;'. Since the holder of a call option is not required to exercise, he will do so if and only if the stock price at time 1 is greater than the exercise price. Now consider the payoff matrix of the state index portfolio and I 0 1 1 European call options written on it with exercise prices
x111

{x111 j
Founda tions for Financial Economics
130
Xw 1
0 0
0
Xw2 2 Xw  Xw 1
0
Xwa Xws  Xw l Xwa  Xw2
Xw 1 0 1 Xw 1 0 1  Xw 1 Xw l n t  Xw 2
0
0
Xw 1 0 1  Xw j n j  l
where the first row is the p ayoff structure of the state index portfolio and the nth row is the payoff structure of the call option with an exercise price Xw ,. _ 1 • This matrix is of full rank by the fact th at Xw 1 is nonzero. Hence the common stock and the 1 0 1  1 European call options are IO i linearly independent securities, and the markets are complete . Similar analysis applies t o a common stock and European put options with exercise prices Xw 2 , xw3 , • • • , xw 1 0 1 The payoff structure of an European stock call / put option with a nontrivial exercise price is a nonlinear function of the p ayoff of its underlying stock and hence is linearly independent of that of its underlying stock . When Xw l = 0, we simply replace the state index portfolio by an European put option with an exercise price equal to any of Xw 2 , , x w 101 in the above payoff matrix to make it of full rank. xw 3 , •
•
•
•
5 . 1 1 . The discussions of Section 5 . 1 0 on the optimality of an allocation appear to be very general. Individuals may have different probability beliefs about the states of nature and may have state dependent and nontimeadditive utility functions. To achieve a Pareto optimal allocation for arbitrary preferences and beliefs through securities markets with options, it suffices to have a state index port folio. The existence of such a portfolio amounts to saying that the realization of the portfolio's time1 payoff allows us to know the re alized payoff of any complex security. This is so, b ecause the realized payoff of the state index portfolio reveals the true state of nature. Equivalently, the existence of a state index portfolio implies that the time1 payoff of any complex security is a function only of that of the state index portfolio. This, however, is unlikely in actual securities markets.
131
Valuation of State Con tingen t Securities
In the following sections, we will restrict our attention to econ omies where individuals have homogeneous beliefs , {7rw i W E 0 } , and stateindependent timeadditive utility functions for lifetime con sumption Uiw ( c;o , ciw ) = Uio ( c;o ) + tt; ( c;w ) 1
where uiO ( · ) and u; ( · ) are increasing, strictly concave, and twice dif ferentiable. We will show that a Pareto optimal allocation may be reached without complete markets and without a state index portfo lio in this case. Fix a Pareto optimal allocation of time0 consumption, {em ; i == 1 , 2 , . . . , I} , and time1 consumption { ciw i W E O , i == 1 , 2, . . . , I} and let A ; , i = 1 , 2, . . . , I, be the set of strictly posi tive weights associated with the allocation. As individuals' utility functions are additive across time and states, relations ( 5 .4 . 1 ) and (5.4.2) become 5.12.
A; u�0 (ci0) = 4>o i = 1 , 2 , . . . , I, Ai11'w � (Ciw ) = 4>w w E O , i == 1 , 2 , . . . , I.
(5.12.1) ( 5 . 1 2.2)
As individuals' beliefs are homogeneous, (5. 1 2 . 1 ) and ( 5 . 1 2.2) imply ( 5 . 12 .3) and (5. 12.4)
respectively. Let two states w and w 1 be such that Ow > Ow ' · We claim that C iw > Ciw' for all i. That is, if the aggregate consumption in state w is strictly greater than the aggregate consumption in state ' w , then the optimal allocation of consumption in state w for an individual i is strictly greater than that in state w' . To see this, we note that since Cw > Ow' ' there must exist an individual k such that Clew > Clew' · By the strict concavity of utility functions , we have
A �eu� (c�e w )
Ciw' Vi. Hence, for given weights > 0, there exists a onetoone relation between the aggregate consumption in a state and the optimal allocation for an individual in t hat state. Also, since utility functions are state independent , this onetoone relation is state independent. In symb ols, there are realvalued fu nctions h such that we can write
..\i
( 5 . 12.5 )
where for brevity we have used Ci and C for individual i's time1 random consumption and for time1 random aggregate endowment, resp ectively. Note that !; is strictly increasing. S imilar analysis can be carried out for the optimal allocation for time0 aggregate consumption to show that there exist strictly increasing functions ho such that Cio = fiD(Co) . The existence of func tions fi (·) implies that a Pareto optimal allocation must have the characteristic that in states where the ag gregate consumption is the same, the allocations are identical . The functions liD 's and fi 's prescribe the Pareto optimal allocation of timeD and time1 aggregate consumption to different individuals. They are called the Pareto optim al sharing rules . In the following several sections, we will concentrate on the Pareto optimal sharing rules for time1 consumption and refer to them as simply the Pareto optimal sharing rules. 5 . 1 3 . The preceding analysis implies that spot market for time D consumption and elementary claims on time1 aggregate consump tion span the set of Pareto optimal allocations. Therefore, a compet itive equilibrium allocation in a securities markets economy where markets are complete with respect to the aggregate consumption states will be Pareto optimal. When corporate securities do not span all the time1 aggregate consumption states, call options written on the time1 aggregate consumption can be added in the manner dis cussed in Section 5 . 1 0 to complete the markets with respect to the time1 aggregate consumption states. In Section 5 . 12, the Pareto optimal sharing rules were shown to be strictly monotonically increasing. They can be nonlinear fu nc tions of the time1 aggregate consumption. When Pareto optimal
Valuation of State Contingen t Securities
133
sharing rules are linear, however, a competitive equilibrium alloca tion in a securities markets economy is always Pareto efficient if there exists a riskless asset . This is so, because from Section 5 .6, in a se curities markets economy, individuals are only endowed with time0 consumption and traded complex securities. To see this, consider a Pareto optimal allocation where individual consumption allocations are linearly related to the aggregate consumption: (5.13.1) I n a competitive securities markets economy, an individual can fi rst sell his endowments and then allocate his initial wealth according to the following. To achieve the constant term ai , individual i can purchase ai shares of the unit riskless discount b ond a t time 0. The _ market portfolio of existing securities pays exactly C . Hence the stochastic term b/J can be achieved by purchasing a bi fraction of the market p ortfolio. The remaining initial wealth will then be used for consumption at time 0. (Note that the existence of a riskless asset can easily be ensured by creating a riskless borrowing and lending opportunity.) The following section giv e s the necessary and sufficient condition on utility functions for Pareto optimal s har i ng rules t o be l i near for all s trictly positive weigh tings A.:s and for a l l distributions of aggregate consumption.
5. 14. Suppose that Pareto optimal sharing rules are linear for all weightings ,\ = (..\i){=1 : ( 5 . 14 . 1)
where 8 is the time1 aggregate consumption/endowment and where ci(..\) denotes the random time1 Pareto optimal consumption allo cation for individual i associated with the weightings ,\ . Note that the optimal sharing rules depend upon the weightings and thus the dependence of ai and bi on ,\. From (5 . 1 2.4) , we know ( 5 . 14.2) Differentiating (5.14.2) , with respect to 8 gives a condition that is necessary for a linear sharing rule to be optimal for a given set of
134
Foundations for Fin ancial Econ omics
weightings >. : v
i, k.
(5. 14.3)
Differentiating (5. 14.2) with respect to A i gives a condition that is necessary for a linear sharing rule to be Pareto optimal for any ar bitrary set of positive weights: uH c; (>. )) + A iU�1 (c; (>.) ) (a; ; (>. ) + b ; ; (>.)C) = .\�o: u� (c��: (>.) ) (a��: ; ( .\) + b�o:; (>.)C) Vi, k,
(5 .14.4)
where a��: ; (>.) and b�o:; (>.) denote aa ��: ( >. ) ja >. ; and ab��: (>. )ja >. ;, respec tively. Substituting (5 . 14.3) into (5. 14.4) and using (5. 14.1) gives u'. ( c; ( >.)) = A; ( >. ) + B; (>.) c; _ (>.) ,  u ' (c; (>.. ) )
�
(5. 14.5)
Since this is true for all distributions of aggregation consumption, fixing A and letting the distribution of C to vary gives a differential equation  u:(z)lu:'(z) = A,(A) + B,(A)z. Since the lefthand side is independent of A, the righthand side must also be so. ThusA �A) = A , and B;(A) = B, for some constants A, and B,. Now we claim that B, = Bk for all i, k. This follows from the second part of Exercise 5 . 6 . Therefore, we have proved that (5.14.5) with B, = B for all i is necessary for all Pareto optimal sharing rules to be linear. Note that (5.14.5) must be satisfied for arbitrary weights A. Therefore, (5.14.5) with B, = B for all i is equivalent to a differ ential equation:
(5. 14.6)
The lefthand side of (5 . 14.6) is the inverse of the ArrowPratt mea sure of absolute risk aversion and is termed the (ArrowPratt mea sure of absolute) risk tolerance, which we denote T; (z) . The deriva tive of T; (z) with respect to z is termed the cautiousness at z. Using the above terminology, (5. 14.6) says that utility functions exhibit linear risk tolerance with identical cautiousness B. Note that A; and B will be appropriately chosen so that the utility functions are increasing and concave. When B f. 0, (5.14.6) is equivalent to uHz) = p; (A; + Bz)  :li ,
(5.14.7)
135
Valuation of State Contingent Securities
for some ·strictly positive constant p; . When B is equivalent to
=
0, however, (5 . 1 4.6)
(5 . 14 . 8) for some strictly positive constant p; . We can interpret individual i's time preference parameter.
p;
to b e
5 . 15. Now we will prove that (5.14.6) is also sufficient for all the Pareto optimal sharing rules to be linear. We take cases. C ase 1: B =/= 0. Using (5.12.4) and (5.14.7) we have
(�::)
B
(A; + Be;) = (Ar.: + Bcr.:) 'Vi , k .
( 5 . 1 5 . 1)
Summing ( 5 . 1 5 . 1 ) over k and rearranging gives ( 5 . 15.2) which is clearly linear in 8. Case 2 : B (5. 14.8) imply . c,
) '  A'· ln ( A.aPa .
·

=
0. Relation (5. 12.3) and
c A; A; E�1 Ar.: ln ( A r.:Pr.: ) + ' I I Lk=l Ar.: Lk=l Ar.:
\../ " va,
which is also linear in 8. Hence (5. 14.6) is sufficient for Pareto optimal sharing rules to be linear for all weightings. Note that when utility functions for time1 consumption satisfy (5. 14.6) and when there exists a riskless security in zero net supply, an equilibrium in a securities markets economy involves two fun d separation i n that a n individual i n equilibrium holds a fraction o f the market portfolio and certain amount of the riskless asset. Readers can compare the results reported here and those of Sections 1 .27 and 1 .28. There we discussed necessary and sufficient conditions on utility functions for an individual's optimal portfolios, for all levels of initial wealth , to be composed of linear combinations of two funds. Those necessary and sufficient conditions are identical to (5. 14.6)
Foundations for Financial Economics
136
except that B is allowed to vary across different individuals. Here changing weights is equivalent to varying initial wealth as in Sections 1 .27 and 1 . 28. For individuals to hold the same risky fund, we also require here that Bi be equal for all i. F inally, the resource constraint that the sharing rules must satisfy necessitates that the separating risky fund is the market portfolio. As a consequence of the fact that a Pareto optimal alloca tion does not vary across states in which the aggregate consumption is the same , the state prices in complete markets have a special struc ture. This special structure enables us to value risky payoffs without knowing a complete set of state prices. F irst we shall fix some notation. Let Oo�; denote the subset of 0 such that C w = k if w E Oo�;, let cp(k) denote the price at time 0 of an elementary claim on aggregate time1 consumption that pays one unit of the consumption at time 1 if and only if C = k, and let 7r( k ) be the p robability that C = k. The timeadditivity and stateindependence of utility functions simplify (5.5.3) to be 5 . 16 .
Vw E O .
(5.16.1)
To rule out arbitrage, we must have cp( k ) =
L
wEO�
¢w
(5.16.2)
where the second equality follows from (5.1 2 . 5) and the third equality follows from the definition of 7r(k) . Now let z be the random payoff of a complex security. Its value
Valuation of State Contingent Securities
137
at time zero, denoted by Sz , is Sz =
L tPw Xw = L L
wEn
k
w EO k
tPw Xw
uH fi ( k)) � 1 � 1r wXw uiO ( Ci Q ) wEn k k = L t/J (k) L Xw w EOk k = I: ,p(k) E[xi6 = k ] , � = �
( 5.16.3 )
11"��)
k
(
)
(
)
where the third equality follows from 5.12.5 and 5.16.1 , the fourth equality follows from 5.16.2 , and the fifth equality follows from the definition of conditional expectation, and E[ ·] denotes the expecta tion with respect to 1rw 's. From 5.16.3 , the price of any complex security can be com puted using ust its expected payoffs conditional on aggregate con sumption states and the prices for the elementary claims on aggregate consumption.
(
(
j
)
)
5 . 1 7 . Recall from Section 5 . 10 that state prices can b e com puted from prices of put and call options on a state index p ortfolio. It then follows that elementary claim prices for the aggregate consump tion states can be computed from put and call option prices on the time1 aggregate consumption. Or equivalently, elementary claims on aggregate consumption states can be manufa c tured by forming portfolios of put and call options written on time1 aggregate con sumption. Under certain conditions on the possible values of time1 aggre gate consumption, we will show in Section 5 . 1 8 that an elementary claim on an aggregate consumption state can be constructed by a butterfly sp rea d of call options . When the time1 aggregate con sumption has a continuous distribution and when the price for a call option· on aggregate consumption is twice differentiable with respect to its exercise price, the price for an elementary claim on aggregate consumption states is related to the second derivative of the option's price with respect to its exercise price.
Foundations for Financial Economics
138
1 2 3
C=L
x (2) 0 0 1
x ( 1) 0 1 2
x(O)
0=1 G= 2 G=3
.
.
.
L
L1
L2
Table 5 . 18 . 1 : Payoffs for Call Options on the Ag gregate Consumption 5 . 18. Initially, suppose that time1 aggregate consumption has possible values: 1, 2, . . . , L. Denote the vector of payoffs of an Euro pean call option on aggregate consumption with one period to ma turity and an exercise price of k as its time0 price will be denoted b y p ( k) . For calls with exercise prices o f 0, 1 , and . 2 units of consumption, the statecontingent payoffs k) are as shown in Table 5.18.1. Note that as t h e exercise price o f a call option i s increased from k to k + 1, two changes in the payoff vector occur: ( 1 ) the payoff in the set of states with 6 = k + 1 becomes zero , and ( 2 ) the payoffs in all states with 6 � k + 2 are reduced by the change in the exercise price. Therefore, in this example, k) k + 1 ) gives a payoff of one unit of consumption in every state with 0 � k + 1, and k + 1) k + 2 ) gives a payoff of one unit consumption in every state for which 0 � k + 2. A security having a payoff of one unit consumption if and only if 0 = 1 may be constructed as [x )  x ( l ) ]  [x ( 1 )  x ( 2 ) ] , since this combination of calls would have a payoff vector of
x(lc) ; x(
x(

x(
x(

x(
(O
1 1 1
0 1 1
1 0 0
1
1
0
An elementary claim for any given level of time1 aggregate consumption can be constructed in a similar manner. Given the call
Valuation of State Contingent Securities
139
prices, p ( k ) , prices of elementary claims on aggregate consumption must be those computed from the replicating portfolio of call options. The portfolio yielding one unit of consumption if and only if time1 aggregate consumption is k consists of buying one call with an exercise price k  1, buying one call with an exercise price k + 1 , and selling two calls with an exercise price k, which i s a butterfly spread. For example, if L = 3 and the prices of calls are p ( O ) = 1 .7, p ( 1 ) = 0.8, and p(2 ) = 0 . 1 , then the respective prices for elementary claims on aggregate consumption are t/>( 1) = 0 .2, t/> ( 2) = 0.6, and t/> ( 3) = 0 . 1 . Note that p ( 3 ) = 0 because L = 3. The price of a one period riskless discount bond paying one unit of consumption in all states would be t/>( 1) + t/> ( 2) + t/> ( 3) = 0.9. The riskless rate of interest would be 1 1 .1 % ; i.e. 1 / .9  1 = 0 . 1 1 1 . 5 . 1 9 . In general, i f the step size between p otential levels of aggregate consumption is !:::. , then , letting k be a possible value of the time1 aggregate consumption, x ( k )  x ( k+t:::. ) has a payoff vector with zeroes for aggregate consumption C � k, and !:::. for all levels of aggregate consumption greater than or equal to k + !:::. . Therefore , the portfolio of call options that produces a payment of one unit of consumption if the aggregate consumption is k and zero otherwise IS:
1
!:::.
( x ( k  !:::. )  x ( k ))  ( x ( k )  x ( k + !:::. ) ) ,
where the coefficient of x ( k ) in this expression is the numb er of calls with exercise price of k that should be held in the portfolio. Since the portfolio of calls gives a payoff of 1 if C = k in period 1, the cost of the call portfolio is t/> ( k) . Now note the following: With step size of !:::. , t/> ( k ) divided by the step size may be written as t/> ( k)  [p(k + !:::. )  p ( k )]  [ p ( k )  p ( k  !:::. ) ] !:::. t:::,. 2
(5 . 19 . 1 )
Suppose that the time1 aggregate consumption has a continuous distribution and that the prices for call options on the time1 ag gregate consumption are twice differentiable with respect to their
140
Foundations for Financial Economics
exercise prices. Then w; � approaches zero in (5.19.1 ) , we get ¢ (k) dk
[p (k + � )  p(k)]  [p (k)  p(k  �)] �2 Ll t O a 2p(k) = ak2 .
=
lim

(5.19.2)
From differential calculus, the finiteness of a 2 p(k) jak 2 implies that the elementary claim price for any particular level of the time1 aggregate consumption is zero. We thus interpret (5.19.2) to be the pricing density for elementary claims on aggregate consumption states in that a complex security that pays one unit of consumption if and only if the time1 aggregate consumption lies in a subset A of the real line has a price:
{ ¢ (k)· dk = { a p(k) dk. }A a k 2 }A dk 2
Moreover, as a generalization of (5 . 16.3) , the price for a complex security having payoffs x is (5. 19.3) Hence (5 . 19.1) gives the pricing function for an elementary claim on C maturing in one period in the discrete cw;e, and (5. 19.2) gives 8 2p(Jc) the pr1cmg d ens1ty w h en c h as a contmuous d'1s t r1'b u t'10n and a .1c 2 exists and is finite. Note from (5. 19.2) that the positivity of elemen tary claim prices implies strict convexity of call option's price in its exercise price . ·
·
·
·
5 . 2 0 . The economies considered in this chapter are two pe riod economies. By the strict monotonicity of their utility functions, individuals will consume all their wealth at time 1. As the single consumption good is the numeraire, by the market clearing condi tion in an economic equilibrium, the aggregate time1 wealth is the aggregate time1 consumption which in turn is the aggregate time1 endowment . Note that the aggregate wealth is just the market
Valuation of State Contingent Securities
141
portfolio that we discussed in detail in Chapter 4. Thus all the con clusions above about time1 aggregate consumption apply to the market portfolio. For instance, ¢(k) can be computed from prices for put and call options written on the market portfolio maturing at time 1. Also, ( 5 . 16.3) can be written as Sz
=
L ¢(k) E[i i M = k] , k
where, as usual, we have used M to denote the time1 value of the market portfolio. Note that the equivalence of time1 aggregate consumption and aggregate wealth holds true only in a two period economy. When the economy lasts for more than two periods, say times 0 , 1 , and 2 , time1 aggregate consumption is only a fraction of the aggregate wealth at that time. The aggregate wealth at time 1 is composed not only of the aggregate time1 consumption / endowment but also of the value of the time2 aggregate consumption / endowment. We will come back to this point later in Chapter 7 . 5 . 2 1 . Now let {¢w i W E n} be the state prices in a competitive economy with complete markets where individuals h ave homogeneous beliefs and timeadditive, stateindependent utility functions that are increasing, strictly concave , and differentiable. We claim that { ¢w i w E 0} will be the state prices in an otherwise identical economy except that there is only one representative agent who is endowed with the aggregate end owmen t (Co , Cw ; w E 0). We will prove our claim through construction  construct a representative agent, endow him with (Co , Cw ; w E 0) , and show that equilibrium prices for state contingent claims are {¢w ; w E 0}. 5 . 2 2 . Consider the competitive complete markets economy equi librium prices of elementary claims { ¢w ; w E 0} and equilibrium al location { cio , Ciw i W E n, i = 1, . . . , I} in Section 5.5. Let Ai be the inverse of the Lagrangian multiplier of individual i's maximization problem, that is, Ai = o; 1 , where we recall that the 9/s are strictly
Foundations for Financial Economics
142
positive. Now define functions ua and u 1 as follows:
s.t.
I
L zi = z,
( 5.22.1)
i=l
and
s.t.
I
L zi = z .
( 5.22.2)
i=l It can be verified that uo and u1 are increasing and strictly concave. An immediate consequence of the definitions of ua and u1 is that I
dcio , , u0 ( Co ) = " L, .Xi ui a ( Cia ) dC0 i=l =
I
L ��o0 = 1,
( 5 .22.3)
i =l
and ( 5. 22.4)
where we have used the resources constraints I
L ciO = Co i =l
and
I
L Ciw = Cw i=l
Vw E n
Valuation of State Contingent Securities
143
to conclude that I
and
" dcio � dCo = 1 i=l
(5.22.5)
" dciw = 1 Vw E O . � dCw i=l
( 5 .22.6)

I

5 . 2 3 . Now consider an economy with a representative agent, whose timeadditive and stateindependent utility functions are u o and UI , whose probability beliefs are { 1rw j W E 0 } , and whose en dowments of time0 and time1 consumptions are (Co , Cw ; w E 0}. We claim that the prices for elementary consumption claims in this economy, using the single consumption good as the numeraire, must be { tPw i W E 0} . O bserve that in this single individual economy, for the markets to clear, the prices must be set so that the representative individual's optimal consumption choice is to hold his endowment . Therefore, using the time0 consumption good as the numeraire, the state price for state w must be equal to the representative individual's marginal rate of substitution between time0 consumption and time1 statew consumption, which is 1rw uH Cw ) ( 5.23 . 1 ) ub (Co) · Substituting ( 5 .22.3 ) and ( 5 .22.4 ) into ( 5 .23.1 ) gives 1rw U� (Cw ) ub (Co) w = 1r tP w = tPw · 1rw
( 5.23 .2 )
Thus, for the market to clear, it is necessary that the price for state w contingent claim be tPw · It is also straightforward to show that { tPw i W E 0} are indeed equilibrium prices in the economy with the representative agent.
144
Foundations for Fin ancial Economics
5.24. The utility functions for the representative indiv idual we constructed in Section 5 .22 depend on the weightings �; . The representative individual's utility functions will be affected when we change the relative magnitudes of those weightin gs. Recall from Sec tion 5.5 that the �i is just the inverse of individual i's Lagrangian multiplier in his maximization and depends upon his initial endow ment . Therefore, the relative magnitudes of the �; 's are determined by the distribution of initial endowments of individuals. As a conse quence, the prices in the economy will be affected by the distribution of initial endowments across individuals. We say that an economy with heterogeneous agents satisfies the aggreg ation p roperty if the equilibrium prices are determined inde pendently of the distribution of initial endowments. We will demon strate below that a sufficient condition for the aggregation property is that utility functions of individuals exhibit linear risk tolerance with identical cautiousness and have the same time preferences. Note that this condition is a little stronger than ( 5.14.6) , the necessary and suf ficient condition for all Pareto optimal sharing rules to be linear. For ( 5. 14.6) to hold, it is not necessary that individuals' time preference parameters be identical. The solutions of ( 5.14.6) include power fu nctions
and negative exponential functions:
where we understand the power function to be ln (A; + Bz) when B = l.
5 . 2 5 . Suppose first that individuals' utility functions are power functions with B =j:. 1 and the same time preference p :
145
Valuation of State Contingent Securities
Fix a set of weightings (Ai ) . Define uo and u 1 as in ( 5 . 22 .1 ) and (5.22.2) . We can compute uo and u1 explicitly. The first order conditions for (5.22.1) are 1
(Ai + B Yi )s = () I
Ai
i = 1, 2, . , 1, .
.
L Yi = y , i=l
{5.25. 1 ) (5.25.2)
where () is the Lagrangian multiplier for the resource constraint . Now raising (5.25.1) to the B power and summing over i gives I
I
L Ai + By = (o)B 2: >f . i=l
( 5 . 25 .3)
Solving for () from (5.25.3) and substituting it into (5.25 . 1) gives
Summing the above relation over i gives
uo ( Y)
=
=
I
1  "' A i (A;
B1� i=l
+ B yi ) l  Ji
I I 1 "k (L _ _ ( f A; + By) l  Ji . ) L >. B  1 i=l i=l
(5.25.4)
Similarly, for u1 we have
u l ( Y)
=B =
I
>.i (At + B y;) 1  "k � 1PL i=l
�
I
I
>.f ) k ( L A; + B y) l  Ji . B 1 P (L i=l i=l
(5.25.5)
Foundations for Financial Economics
146
From (5 .23.2) , the prices for e i ementary securities are tP
w
'll'w P (L[= l Ai + BCw )  Ji  ('E[=1 Ai + B Co )  i; '

Vw E 0
(5.25.6) ,
which are independent of the weightingsl Prices are determined in dependently of the distribution of initial endowments. We leave verification of the aggregation property when B = 1 for readers in Exercise 5 .4 .
5.26. When the utility functions are negative exponential, we can prove a bit more. In the above construction of a representative individual, we assumed that individuals have homogeneous beliefs and time preferences. The representative individual naturally inher its these beliefs and time preferences. In the negative exponential case, we can allow heterogeneous beliefs and time preferences in the construction of the representative agent , and his beliefs and time preference will be c omposites of individuals' beliefs and time prefer ences. Consider the following maximization: m �x
{!l;., ;wEO,t=l,2,
. . .
,I}
I � � � �i Pi �
1= .
1
y·
'll' iw ( Ai exp { � } ) Ai w Eu ,....
I
s.t. L Yiw
i=l
=
(5.26.1)
Yw Vw E n.
The fi rst order necessary and sufficient conditions are
 Yiw �iPi 'triw exp { � } = Ow Vw E 0, i = 1 , 2 , . . . , I (5.26.2) I
L Yiw = Yw i=l
Vw E n,
(5.26 . 3)
where Ow is the Lagrangian multiplier for the resource constraint in state w. The first order conditions imply (5.26.4)
147
Valuation of State Contingent Securities
where f1 is the product sign. It then follows from (5.26.2) an d (5.26.4) that the maximand of (5.26 . 1 ) is
We interpret (5.26.6)
i=l
to be the time preference o f the representative individual, I
A;
IT (11"iw) E�=l A k
i=l
Vw E
n
(5.26.7)
to be his probability b eliefs, and (5.26 . 8) to be his utility function for time1 consumption . Here let us note that (5.26 .7) may not be probabilities as they may not sum to one. That constitutes no problem, as we can always normalize to get prob abilities. Taking Pi = 1 and 11"i w = 1 in (5.26 .5) , we get (5.26 .9) The prices for elementary claims in an economy with this representa tive individual and with aggregate endowments (Co , Cw ; w E 0) are,
Vw E O ,
(5.26 .10)
148
Foundations for Financial Economics
which are independent of the weightings. We will ask you to verify in Exercise 5.5 that ( 5 .26. 10} gives prices for elementary claims in a complete markets economy with individuals having negative expo nential utility functions. Hence negative exponential utility functions are sufficient for aggregation.
Exercises
5 . 1 . Verify that the uo and u1 defined in (5.22 . 1 ) and (5.22.2) are strictly increasing and concave. 5.2. Solve (5. 14.6) to verify that the solutions include power func tions and negative exponential functions. 5.3. Verify that negative exponential utility functions imply linear P areto optimal sharing rules. 5 .4. Derive equilibrium prices for elementary claims in a complete markets economy when individuals have log utility functions. 5.5. Verify that ( 5.26.10) gives equilibrium prices for elementary claims in a complete markets economy where individuals have nega tive exponential utility functions with different coefficients of the ArrowPratt measure of absolute risk aversion. 5.6. Fix a Pareto optimal allocation and the sharing rules for this allocation , k Let T; (z) = uHz)/u"(z) be the risk tolerance function for individual i. Show that
The cautiousness for individual i at z, denoted by u; (z) , is Tf (z) . Show that for the fi 's to be linear, it is necessary and sufficient that That is , individuals' cautiousness must be identical along the optimal schedule, although their cautiousness may not be a con stant across individuals.
Valuation of State Con tingent Securities
149
5.7. Let /iO and ft be Pareto optimal sharing rules for individual i for a set of weightings l. Define u�(z) = l:u�0(/io(z)) and uHz) = liu�(f.(z)) . Show that tto and tti are strictly increasing and concave and can be utility functions for a representative agent as discussed in Section 5.22 and Section 5.23. 5.8. Suppose that all the individuals' utility functions exhibit de creasing absolute risk aversion. Show that the representative agent's utility functions as defined in Section 5.22 also exhibit decreasing absolute risk aversion .
Remarks. The general framework of this chapter i s due t o Ar row ( 1964) and Debreu (1959) . For more discussion on the issues related to recapitalization and the Modigliani and Miller theorem, see Litzenberger and Sosin ( 1977) . The discussions on the aHoca tional role of stock options follow Breeden and Litzenberger ( 1978) and Ross (1976) . The construction of a representative agent is taken from Breeden and Litzenberger ( 1978) , Constantinides ( 1982) , and Wilson ( 1968) . The discussions on the aggregation property of prices and the Pareto optimal sharing rules are freely adapted from Amer shi and Stoeckenius ( 1983) , Rubinstein (1974) , and Wilson ( 1 968 ). For an extensive review of the literature on equilibrium under un certainty, see Radner ( 1982) . Exercise 5.8 is taken from Kraus and Litzenberger ( 1983) .
150
Foundations for Financial Economics References
Amershi, A., and J . Stoeckenius. 1983. The theory of syndicates and linear sharing rules. Econometrica 5 1 : 1 4071416. Arrow , K. 1964. The role of securities in the optimal allocation of riskbearing. Review of Economic Studies 31: 9196. Breeden, D . , and R. Litzenberger. 1978. State contingent prices implicit in option prices, Journal of Business 5 1 :621651 . Constantinides, G . 1982. lntertemporal asset pricing with heteroge neous consumers and without demand aggregation. Journal of Business 5 5 : 253267. Debreu, G. 1959. Theory of Value . Yale University Press, New Haven . Kraus, A . , and R. Litzenberger. 1983. On the distributional condi tions for a consumptionoriented three moment CAPM. Journ al of Finance 3 8 : 1 3 8 11391 . Litzenberger, R., and H . Sosin. 1977. The theory o f recapitaliza tions and the evidence of dual purpose funds. Journal of Finance 3 2 : 14331455. Modigliani, F., and M. Miller. 1958. The cost of capital, corpo rate fin ance and the theory of corporation finance. American Economic Review 48:261297. Radner, R. 1982. Equilibrium under uncertainty. In Handbook of Mathematical Economics, Vol II:9231006. Edited by K. Arrow and M. lntriligator. NorthHolland Publishing Co, New York . Ross , S . 1976. Options and efficiency. Quarterly Journal of Eco nomics 90:7589. Rubinstein, M. 1974. An aggregation theorem for securities markets. Journal of Financial Economics 1:225244. Vari an, H. 1978. Microeconomic An alysis. W . W . Norton & Com pany, Inc . , New York . Wilson, R. 1968 . The theory of syndicates, Econometrica 36:1 19131.
C H A PT E R 6 VALUATIO N O F C O M PLEX S E C U RITIES AND O P T I O N S W I T H P REFERE N C E RESTRI C T I O N S
In Chapter 4, we made assu mptions on the return distri butions in order to derive linear valuation relations . In this chapter, we will first discuss valuation principles for complex securities in the framework of Chapter 5 without special assumptions on either the return distributions or individuals' utility functions. We will then de rive explicit valuation expressions for risky assets under preference and distribution restrictions. In particular, the price of a European call option written on a stock when individuals' utility functions ex hibit constant relative risk aversion and the option's underlying asset has a payoff structure that is j ointly lognormally distributed with the aggregate consumption is explicitly computed. We also apply the pricing formula for a European call option to study the pricing of risky corporate debt. In the last section of this chapter, we derive a pricing relation similar to the CAPM for a particular class of risky assets. The pricing relations derived in this chapter provide additional 6.1.
15 1
Foundations for Financial Economics
152
testable propositions concerning the pricing of complex securities such as common stocks and options. Some of these propositions will be empirically examined in Chapter 10.
6.2. Assume that individuals have homogeneous beliefs 1f'w and utility functions that are timeadditive and stateindependent, de noted by uio and ui , and assume that these utility functions are increasing, strictly concave, and differentiable. There are N + 1 securities traded, indexed by j = 0 , 1 , . . , N. Individuals' time0 endowments are units of timeD consumption good and shares of traded securities. Security j is represented by its state dependent payoff structure Xj w · The 0th security is a riskless discount bond with XOw = 1 for all W E 0 . We assume that t h e equilibrium allocation is Pareto optimal. We recall from Chapter 5 that under this condition, a representative agent with increasing and strictly concave utility functions uo and u1 c an be constructed , and the price of a primitive state contingent claim can be expressed as .
1f'w u � (Cw ) u� (Co )
Vw E O,
(6.2.1)
where ¢w is the state price for state w . A complex security may be viewed as a portfolio o f elementary state contingent claims. Thus the price for security j is
S;
=
L tPw Xjw ·
w En
(6.2.2)
Substituting (6.2 . 1 ) into ( 6.2.2) for ¢w gives (6.2.3) where we have used i; to denote the random time1 payoff of se curity j. For the case of a riskless unit discount bond  a complex security that pays one unit of consumption at time 1 in all states, we have so
=
E
[ l
u� (c) · ub(Co )
(6.2.4)
Valua tion of Complex
Se c u r ities
with Preference Restrictions
153
As So is the price of a unit discount bond, the riskless interest rate Tf
IS
rf
1
=   1.
( 6 . 2 .5)
So
By the strict monotonicity of the utility functions, So > 0. This implies that r1 >  1 Substituting (6.2.5) into (6.2.4) gives .
[ l
u� (C) 1 __ = E 1 + r1 u� ( Co)
·
(6 . 2.6)
Dividing both sides of (6.2.3) by S; and using the definition of covariance, we can write E\F;  'I ] = 
( [:�(�;)])' E
Cov ( ;; , u� (O)juh (Oo)) ,
( 6 .2 . 7)
where :;i = x ; / S;  1 is the rate of return of security j. Substituting (6.2.6) into (6.2.7) gives an equilibrium relation for the risk premiums on securities: (6.2.8) By the fact that Ut is strictly concave, the risk premium of a secu rity is positive if and only if its random payoff at time 1 is positively correlated with the time1 aggregate consumption . Note that in a twoperiod (period 0 and period 1) economy, by the strict mono tonicity of utility functions, the time1 aggregate consumption a is equal to the time1 aggregate endowment, which in turn is equal to time1 aggregate wealth M. Therefore, (6.2.8) can be written as (6 . 2 . 9) That is, the risk premium of a security is positive if and only if its time1 random payoff is p ositively correlated with the time1 aggregate wealth. The intuition behind this result is the same as that of the Capital Asset Pricing Model. One unit of consumption in a state where the aggregate resource is abundant is less valuable than one unit of consumption in a state where the aggregate resource
154
Foundations for Financial Economics
is scarce. Therefore, a security that pays more in states where the aggregate consumption/wealth is low is more valuable than a security that pays more in states where the aggregate consumption/wealth is high, ceteris paribus. As a result, the price for the former will be higher than that for the latter, and the rate of return on the former will be lower than that on the latter. The market portfolio is a portfolio of traded securities. Thus its rate of return rm must also satisfy (6. 2.9) : (6.2. 10) Relation (6.2 .10) implies that the risk premium on the market portfo lio must be strictly positive, as r I >  1 and u� is strictly decreasing so that Cov (rm , u� (M )) is strictly negative. (Readers should com pare this with the result of Section 4 . 14.) Substituting (6 2 10) into (6.2.9) gives .
] ]  Cov (i; , u� ( M )) E[ E [ r; rm  rI .  rI Cov(im, uHM)) 
.
( 6 .2. 1 1 )
I n equilibrium, the risk premium of security j is proportional t o that of the market portfolio. The proportionality is equal to the ratio of the covariance of r; and u� ( M ) and the covariance of r m and u� ( M ) . 6 . 3 . We will now specialize the pricing relation of (6.2 . 1 1 ) by considering a class of utility functions. Assume that individuals' utility functions for time1 consumption are power functions:
(6.3.1 ) and that there is a riskless asset. Note that B is assumed to be con stant across individuals. We also assume that the u/s are increasing and strictly concave over the relevant region. Recall from Chapter 5 that the Pareto optimal sharing rules for time! consumption are linear in this case and can be attained if there is a riskless asse t and if all assets are traded. Therefore, by the assumption that individ uals are endowed only with holdings of shares of traded securities
155
Valuation of Complex Securities with Preference Restrictions
and time0 consumption good in a securities markets economy, the equilibrium allocation is Pareto optimal. Moreover, Chapter 5 also shows that there exists a representative agent whose utility function for time! consumption is a power function with the same B : u1 (z
where A
=
'2:{=1 A; .
I
)=8 
1
1
(A + B z ) 1  s ,
(6 . 3 . 2 )
Hence (6 . 2 . 1 1 ) becomes (6.3.3}
Note that when B =  1 , individuals' utility functions are quadratic and (6 . 3 . 3 } becomes the familiar CAPM relation. When B =  1 /2, the representative agent's utility function for time! consumption is a cubic function: 2 I 3 U t ( z) =  3 ( A 2 z) ·
The marginal utility in this case is
so U t is increasing and strictly concave for (6. 3 . 3 } becomes
z
o
=
o
if x; ?: k , if z; < k .
The time1 payoff of this strategy i s nonnegative and i s strictly pos itive with a strictly positive probability, since there is a strictly pos itive probability that z; < k. Therefore, its initial cost must be strictly positive to prevent something being created from nothing. That is, we must have p; ( S; , k)  S; + kl ( 1 + ri)
>
0,
which is equivalent to P; (S; , k)
>
S;  kl (1 + ri ) .
(6 . 5 . 2)
Lastly, since a holder of the option only has the right and not the obligation to buy a share of its underlying security at the exercise price, the price of the option must be nonnegative. Moreover, by assumption there exists a strictly positive probability that the option will be exercised. Therefore, P; (S; , k) > 0. This observation together with ( 6. 5 . 2 ) gives p; (S; , k)
>
max [ S;  kl (1 + r I ) , 0 ] .
which was to be shown. The intuition behind this inequality is as follows. The present value of an obligation to buy a share of security i at time 1 at a price k is S;  k I ( 1 + r I). When there exists a strictly positive probability that x; will be strictly less than k, the option not to buy has a strictly positive value. Thus the call option must be worth strictly more than S;  kl ( 1 + r I) . On the other hand, there
158
Fo un dations for Fin an cial Economics
exists a strictly positive probability that the option will be exercised. H ence , p; (S; , k) > 0. 6 . 6 . In Section 5 . 1 9 , by assuming that an option price is a twice differentiable fun ction of its exercise price, we showed that an option price is a convex function of its exercise price . This property, as it turns out, holds more generally. We want to show that
(6 .6 1 ) .
where k = o.k + ( 1  o.) k and o. E (0, 1 ) . Consider the strategy of buying o. shares of the call option with an exercise price k and (1  o. ) shares of the call option with an exercise price k , and short selling a share of call option with an exercise price k. Without loss of generality, assume that k > k. The time1 payoff of this strat egy
lB
{
0 o.(x;  k) > o 1  o.) (k  x;)
�
> o
if if if if
ri; � k , k < x; � k:, k:
x;
< :c,. � k, > k,
which is nonnegative. Thus,
o.p; ( S; , k) + ( 1  o.)p; (S; I k)  p; (S; , k) � o,
which is j ust ( 6.6.1 ) . When there is a. strictly positive probability that x; E (k, kj , the wealt inequality becomes a strict inequ ality. You are asked to prove in Exercise 6.2 that p; (S; , k) is a de creasing function of k. Hence Pi ( S; , k) is a decreasing and con vex fun ction of its exercise price. In a frictionless market, buying an option on two shares of se curity j with an exercise 2k should be equal to buying two options with an exercise price k on security j. To see this, we simply note that the random payoffs of the former are identical to those of the latter . 6 .7 . An option on a positively weighted portfolio of securities with an exercise price k is less valuable than a positively weighted
Val uation of Complex Securities with Preference Restrictions
159
por tfolio of options with equal exercise prices k. Consider a p ositively weighted portfolio of securities with weights a; , j = 1, . . . , N, where a; denotes the p ortfolio weight on security j. Note that N
=
L: o; i=l
o; 2: 0.
1
The time0 cost and time1 random payoff of this portfolio are
s•
:=
and
i*
=
N
L o;S;
j= l N
l.: o; i; ,
i=l
respectively. Let p * (S • , k) b e the price of a European call w ritten on the portfolio of securities with an exercise price k that expires at time 1 . The time1 random payoff of this option is
max f
N
2: o;i;  k, OJ .
j= l
Since max [ z, D) is a convex function of z , by the Jensen's inequality we have N
N
max f2: o; i;  k, O] � 2: o; max[x;  k , OJ . j= l
j =l
Note that the righthand side of the inequality is the time1 ran dom payoff of a portfolio of call options on individual securities with identical exercise prices k. Thus N
p* (S\ k) � L o ; p;(S;, k) , j =l
where the inequality is strict if and only if there exist some j and j' such that x; < k < xi' with a strictly positive probability. Suppose
Foundations for Financial Economics
160
that all securities have payoffs such that all individual options with an exercise price k will be optimally exercised simultaneously. Then a positively weighted portfolio of options on individual securities with an exercise price k will be worth just as much as an option on a portfolio of securities with the same exercise price and using the same weights. Suppose on the other hand that some options on individual securities will not be optimally exercised simultaneously. A p ortfolio of options, unlike an option on a portfolio of securities, gives its holder an "option" to exercise different options individually. Thus a positively weighted portfolio of call options is strictly more valuable than a call option on a portfolio of securities with the same weights. 6 . 8 . The holder of a European put option has the right to sell its underlying security at the exercise price on the maturity date. Let P; ( S; , k) be the time0 price of a European put written on security j with exercise price k and maturity date 1 , when the current price of security j is S; . The time1 payoff of this put option is
{
k 0

x;
>
0 if xi < k, if i; 2:: k .
The price o f a European put can be computed from the prices of its underlying security and its European call counterpart through a relation called putcall parity . We claim th at P; ( S; , k ) = k/(1 + r1)  S; +p; ( S; , k ) . To see this, consider the following strategy: lend k/( 1 + rJ) at the riskless rate, short sell one share of security j, and buy one share of the European call with exercise price k. The time1 payoff of this strategy is if x; < k, if k :::; i; ,
which is clearly the payoff of a European put with exercise price k . To rule out arbitrage opportunities, two packages of financial assets having the same payoffs must sell for the same price. Hence our assertion follows. It is easy to see that P; (S; , k) is an increasing function of k. Also, using arguments similar to those of Section 6.5, we can get
Valuation of Complex Securities with Preference Restrictions
161
You will be asked to verify these two relations in Exercise 6 .3 . Given putcall p arity and the facts that t h e price o f a call i s a decreasing function of its exercise price and that the price of a put is an increasing function of its exercise price, we have
when p; (S; , k ) and P; (S; , k) are differentiable functions of k. More over, P; (S; , k ) is also a convex function of k by putcall parity. 6 . 9 . In the previous two sections, we have seen that a call price is a decreasing and convex function of its exercise price. We will show in this section that, fixing a return distribution on the underlying as set, the price of a call option is an increasing and convex function of the price of its underlying asset. Readers are cautioned to note that this relation is a comparative statics result and is not an arbitrage relation, since different stock prices can not prevail contemporane ously. Consider p; (S; , k) . Assume the distribution of x;/S; is invariant with respect to changes in S; . For example, if we increase S; to Sj , then x; changes to Sj x; /S; . We claim first that .P; (S; , k ) is increasing in S i and is strictly so if the probability that x; > k is strictly positive. By the assumption that return distribution is held fixed, we h ave
p; ( Sj , k )
=
2::
s� � Pi(S;, kS; /Sj ) s s� � p; ( S; , k ) 2:: P; (S; S
(6.9.1)
1
1
,
k) .
(6.9.2)
Relation (6 .9 . 1 ) follows since, given that the return distribution is invariant to changes in S; , the time1 payoff of a call on securityj with an exercise price k when security;"s price is Sj is equivalent to the time1 payoff of Sj / S; shares of call options with an exercise price kSi /Sj when security j's price is S; . Note that (6 .9.1) amounts to saying that the function Pi (S; , k) is homogeneous of degree one in Si and k. The first inequality of (6.9.2) follows since the call option
162
Foundations for Financial Economics
price is a decreasing function of its exercise price. Finally, the second inequality of (6.9.2) follows because Sj /S; > 1 by assumption and p; ( S; , k ) 2: 0. The second inequ ality of (6.9.2) is strict if there is a strictly positive probability that x; > k , as then P; (S; , k ) > 0. Next we want to show that p; ( S; , k ) is a convex function of S; . Let S; =: aS; + ( 1  a ) S;, where a E (0, 1 ) . To prove convexity, we must show that A
ap; ( S; , k ) + ( 1  a) p; ( Sj , k )
2:
P; ( S;, k ) .
From t h e fact that p; ( S; , k ) i s a convex function of k for all possible S; , 'YP; (1 , kl ) + (1  y )p; ( l , kz ) 2: p; ( 1 , ks)
Vy
E
(0, 1) ,
( 6.9.3 )
where k3 = yk1 + (1  y ) k2 . Now take 1 = a S; /S; , k1 = k/S;, and k 2 = k/Sj . Multiplying both sides of (6 .9.3 ) by S; and recalling from ( 6.9 . 1 ) that p; ( S; , k ) is homogeneous of degree one in S; and k,
Using the definitions of S; , y, k1 , k2 , and ks , this inequality can be written as
ap; ( S; , k) + ( 1  a) p; ( Sj , k) ;::: P; ( S;, k ).
(6.9.4)
The weak inequality of (6.9.4 ) will be strict if that of ( 6.9.3 ) is strict. From Section 6 .6 , we know that ( 6.9.3 ) is a strict inequality if there exists a strictly positive probability that x; lies between kl and k2 or equivalently between k/S; and k/Sj . Figure 6.9.1 illustrates the general shape that a call option price should have as a function of its underlying security price and its exercise price, while holding constant the return distribution of its underlying asset . Note that we have used ( 6.5 . 1 ) in F igure 6 .9. 1 . 6 . 10 . In this section, a pricing formula for a European call is derived under conditions on individuals' preferences and on the joint
163
Valuation of Complex Securities with Preference Restrictions
0
k
S J·
Figure 6 . 9 . 1 : Call Option Price as a Function of Its Underlying Stock Price distribution of the time1 aggregate consumption and the payoffs of the option's underlying asset. Consider a twoperiod securities markets economy. Individ uals' utility functions are timeadditive extended p ower functions with identical cautiousness as in (6.3 . 1 ) . Moreover, assume that '2:{= 1 Ai = 0. Recall that in a securities markets economy, an in dividual is endowed with time0 consumption and shares of traded securities. From Sections 5.14 and 5 . 1 5 , we know that the equilib rium allocation will be Pareto optimal, since optimal sharing rules are linear. Thus, a representative agent can be constructed with power utility functions:
where
p
is the time preference parameter. Then (6.2.3) implies that (6 . 1 0 . 1)
164
Foundations for Financial Economics
We will further assume that x; and C are bivariate lognormally distributed, that is, In x; and In C are bivariate normally distributed with means (fi.; , fi.c) and variancecovariance matrix:
where K. is the correlation coefficient of In x; and In C . This assu·xnp tion implies that ln(x; /S; ) and ln p (C /Co)  B are bivariate normally distributed with means ( 1' ; , 1' c ) = (fi.;  In S; ,  B fi.c + ln p + B i n Go) and variancecovariance matrix
G iven the above distributional assumption, (6.10.1) can be writ ten as
!+ oo 1 +oo (ez  k/S;) e11 f(z, y ) dzdy, (6.10.2) oo density function for z = ln(x; /S; ) and y = where f ( z, y ) is the joint p; (S; , k) = S;
ln{lc/Sj)
ln p ( C /C0 )  B . Relation (6. 10.2) may be rewritten as the difference between two integrals u
!+oo oo 1+oo ez+ f (z , y ) dzdy k J + oo j +oo e11 f(z , y ) dzdy.
P; (S; , k ) = S;
ln{lc/Sj)
 oo
( 6. 10.3)
ln{lc/Sj)
In the next section the two integrals are evaluated and the following relations are obtained:
Valuation of Complex Securities with Preference Restrictions and
=
/_: loo ez + 11 f (z , y )dzdy
( el'j +J.'e+ (u] + 2�taj u. + u�) / 2 ) N(
a + j #L + O' j
K.O'.,
+
u; ) ,
165
(6. 10.5)
where N( ) is the distribution function of a standard normal random variable : ·
_
2 1 n (z ) = _ e z / 2
and
..j2i
is the standard normal density function. By setting a = ln(k/S; ) , (6. 1 0.4) may b e used t o evaluate the second integral o n the right hand side of (6. 10.3) , and (6.10.5) may be used to evaluate the first integral on the right side of (6.10.3) by setting a = ln (k/S; ) . It is easily verified that (6. 10.6) and that E
(!:p (�) B]
�
,•;+P
=
N(Z�c
=
k (1 +. rI )  1 n ( Zk )
0,
= ( 1 + r,) 2kN(Z�e)
169
(6. 13.2) > >
0' 0.
(6 . 13.3) (6 . 1 3 .4)
Relations (6 . 1 3 . 1 ) and (6. 13.2) are confirmations of the general discussions in Sections 6.6 and 6.9. Note that in computing (6. 13.2), we have assumed that the distribution o f ln(z;/S;) i s unchanged when we vary S; . Relation (6 . 13 .3) says that the higher the variance of ln (z;/ S;) , the more valuable the option is. This is so because an option holder does not have an obligation to exercise  he only has the right to do so. Hence, a higher u; allows a higher upside potential for an option. Finally, the higher the riskless interest rate, the lower the present value of the exercise price in the event of exercising at time 1, thus the more valuable the option is. From (6. 1 3 . 1 ) , we can also get (6 . 1 3 .5) That is, p; (S; , k) is a convex function of the exercise price, which is a general property of call options proved earlier using an arbitrage argument. From Section 5 . 19 , we can interpret the righthand side of (6. 13.5) to be the price density for a security that pays one unit of consumption at time 1 if and only if the payoff of security j is equal to k at that time. This price density is always strictly positive. 6 . 1 4 . The option pricing formula derived above can be applied to study the pricing of risky corporate debt . Consider the economy in Section 6 . 10. Suppose that firm j has one share of common stock and a discount bond with face value k outstanding, with prices S; and D; , respectively. The discount bond matures at time 1. The total time1 earning of firm j is x; , which is assumed to be joint lognormally distributed with C , the time1 aggregate consumption, with parameters as in Section 6. 10. The present value of x; is the total value of the firm at time 0 and is denoted by V; . Note that
1 70
Foundations for Financial Economics
V; = S; + D; . The time1 random payoff of the discount bond is min [z; , k] . When the firm is solvent at time 1, that is, when x; � k, the bond holders receive the face value of the discount bond; otherwise, the firm goes bankrupt, and the bond holders take over the firm and get i; . One way to compute D; is to use (6.2.3) : (6 .14.1) We can , however, use the BlackScholes option pricing formula to compute D; in a direct and straightforward way. Note that the time1 payoff to the equity holders is max[z;  k , O ] . When the firm is solvent at time 1, the equity holders pay off the bond holders and get the residual value of the firm, which is Xj  k; otherwise, the bond holders take over the firm , and the equity holders get zero . Therefore , the equity holders are holding a European call on the total value of the firm with an exercise price k maturing at time 1. The value of the equity is si
=
V; N(Z�e + Uj )  (1 + Tf )1kN ( Z�e) (6.14.2) ln(V;/k) + ln(1 + rf ) ! . u; where Z�e = 2 Uj From the comparative statics of Section 6 . 1 3 , we know that, ceteris paribus , the value of the equity decreases as the face value of the b ond increases, and increases as u; increases, that is, ·as the total time1 earning of the firm becomes more volatile. When the total value of the fi rm V; is fi xed, the increase in u; shifts value from the debt holders to the equity holders. The value of the discount bond is D; = V;  S; (6. 14.3) = V; ( 1  N(Z�e + u; ) ) + ( 1 + r 1) 1 kN(Z�e)· " It increases as its face value increases and decreases as the earnings of the fi rm become more volatile. We give two interpretations of a risky debt. The time1 random payoff of the debt can be written as _
min[z; , k]
=
=
x; 
max[z;  k , O] k  max[k  i; , O] .
(6.14 .4) (6. 14.5)
Valuation of Complex Securities with Preference Restrictions
171
Using (6. 14.4) , the bond holders can be v iewed as holding the fi r m while selling a European call option on the value of the fi rm with an exercise price equal to the face value of the debt to the equity holders. We used this interpretation to compute (6. 14.3) . On the other hand, (6.14.5) implies that the bond holders are holding a riskless discount bond with a face value k, while at the same time selling a European put option on the value of the firm with an exercise price k. In the event that the fi r m's time1 earnings are strictly less than k, the equity holders will sell the firm to the bond holders at a price k. Since x; is lognormally distributed and there, therefore, exists a strictly positive probability that the fi.rm will default on the debt, the put option of (6.14.5) is not worthless, and the debt is risky. Therefore, the discount bond will sell at a price strictly less than k /(1 + r 1) and has a strictly positive risk premium. 6 . 1 5. Not all securities . in strictly positive supply can have payoffs that are jointly lognormally distributed with time1 aggre gate consumption, since the sum of lognormally distributed random variables is not lognormally distributed. Under the conditions of Sec tion 6.10, however, we can always use (6.10. 10) and (6. 10. 13) to price European call options on the aggregate consumption/wealth . Recall from Section 5.19 that these option prices can, in turn, be used to price any complex securities. Let Pc ( k) denote the price of a European call on aggregate con sumption with an exercise price k maturing at time 1. Then
(6 . 1 5 . 1 ) where (6.15.2) (6 .15.3) and (6.15.4) is the present value of the time1 aggregate consumption.
172
Foundations for Financial Economics
The price for one unit of consumption paid in states where the time 1 aggregate consumption is equal to k is (6.15.5)
Since the probability that time1 aggregate consumption will be equal to any fixed k is formally equal to zero, �c(k) is equal to zero. (Mathematically, dk is treated as zero .) The pricing den s ity q,c ( k ) = � c (k)/dk, is strictly positive, however. We can do a comparative statics analysis of q, c ( k ) . The elas ticity of q,c ( k ) with respect to an increase in the value of Sc is ana lyzed holding the distribution of (0 /Be) constant thereby implying a proportional change in 0. The resulting increase in probabilities of "high" levels of 0 and decrease in probability of "low" levels of 0 increase and decrease, respectively, their elementary claim prices. This elasticity is a In q, c ( k ) (6.15.6)  fie . a ln Sc Note that Z�c is a strictly decreasing functio n of k and .
k+0
lim Z�c =
oo ,
k+oo
lim Z�c =
 oo .
Therefo re, (6. 1 5.6) is positiv e for high k and negativ e for low k. The elastici ty of q,c ( k ) with respect to the standard deviati on
•
(6. 15.7)
This elasticity will b e positive for very large and very small k and will be negative for k near E (O) , because an increase in variance increases the probability of extreme observations relative to the probability of central observations. The elasticity of the elementary claim price with respect to the riskless bond price, ( 1 + r1 )  I , is (6 . 1 5 . 8)
Thus , an increase in the riskless bond price lowers the prices of claims that pay off when the level of 0 is high and raise the prices of claims
Valuation of Complex Securities with Preference Restrictions
that pay off when the level of C is low. However, since (1 + Tf )
173
1 :::::
J0+oo cf>c (k)dk, an increase in (1 + r1t 1 must increase the average
elementary claim price. From (6. 10.6) and (6 . 10.8), a change in the bond price may be associated with either a change in the expected growth rate of aggregate consumption and/or a change in relative risk aversion . Both these possibilities would provide an intuitive explanation for the resulting impact on elementary claim prices
6.16. We analyzed the comparative statics of cf>c (k) with re spect to the parameters of the economy in the previous section. We can also extract some information on the structure of cf>c ( k) for dif ferent levels of k. The elasticity of cf>c (k) with respect to the level of consumption on which it is contingent is
3 ln cf>c (k) zk = 3 ln k fTc
 1
•
(6.16.1)
For levels of k far below E (C) , the elasticity will be positive be cause the probability density for C increases as k increases. For k well above E (C) , the elasticity will be negative due to the combined effects of the decreasing probability density of the level of consump tion on which the elementary claim is contingent and of decreasing marginal utility of consumption. 6.17. Previous analyses demonstrated that lognormally dis tributed aggregate time1 consumption and constant relative risk aversion utility functions for the representative agent are sufficient conditions for the BlackScholes option pricing formula to price Eu ropean options on aggregate consumption correctly. Given that time1 aggregate consumption is lognormally distributed, a constant rel ative risk aversion utility function for time1 consumption for the representative agent turns out to be also necessary for the Black Scholes formula to price European options correctly. Rec all that the pricing density of an elementary claim on aggregate consump tion divided by the probability density of occurrence of that level of aggregate consumption is the marginal rate of substitution between present consumption and future consumption for the representative
174
Foundations for Financial Econ omics
agent. The elasticity of this ratio with respect to aggrega te consump tion is constant if and only if the represen tative agent's utility func tion for time1 consump tion exhibits constant relative risk aversion . The probabil ity density of a given level of aggregat e consump tion k , given Sc , is (6.17.1) Therefore, the price density of a n elementary claim o n aggregate consurqption divided by the probability density of the occurrence of that level of aggregate consumption is tf>c ( k ) 1rc (k) =
[
r1) 1n (Sc /k) + ( p ln ( H r1)) ( p+ln ( H rt)  an exp p ln(aH 2 &� � (1 + r1 )
The e lasticity o f (6.1 7.2) with respect t o k is a ln (t/>c (k)/,... c (k)) = 8 ln k
]
( 6.17.2) (6 . 1 7 .3)
which is a constant. Exercise 6.5 will ask the reader to show that the negative of the above elasticity is the coefficient of relative risk aversion for the representative agent's utility function for time 1 con sumption. Thus, using the BlackScholes formula to price options on aggregate consumption is implicitly assuming that individuals' util ity functions for time1 consumption aggregate to a constant relative risk aversion utility function.
Exercises
6 . 1 . Derive a pricing relation similar to (6.3.3) when individuals have log utility functions.
Valuation of Complex Securities with Preference Restrictions
175
6.2. Show that the price of a European call option is a decreasing function of its exercise price, and find the conditions under which it is a strictly decreasing function of its exercise price. 6.3. Let P; (S; , k) be the price of a European put on security j with an exercise price k maturing at time 1 . Show that P; is an increasing function of k and that, when the probability that x; < k lies strictly between 0 and 1 ,
P; ( S; , k)
>
max [ k / ( 1 +
r1)  S; , O] .
6.4. Verify relations ( 6.13.1 ) to (6.13.5 ) . 6.5. Prove that ( 6.17.3 ) is the relative risk aversion of the represen tative agent's utility function for time1 consumption. 6 .6 . We derived the BlackScholes option pricing formula by assum ing that the representative agent's utility functions for time0 as well as for time1 consumption are of constant relative risk aversion . Derive a similar pricing formula by assumin g only that the representative agent 's utility function for time1 consump tion exhibits constant relative risk aversion .
Remarks . The discussion on skewness preferences follows Kraus and Litzenberger ( 1976, 1983 ) . Discussions in Sections 6.56.9 are taken from Merton ( 1973 ) , in which readers can also find a host of related subjects. Our derivation of the BlackScholes option pricing formula is from Rubinstein ( 1976 ) . Unlike the original BlackScholes deriva tion, Rubinstein's derivation uses equilibrium arguments. Black and Scholes ( 1973 ) use arbitrage arguments and derive their formula in a continuous time economy. Merton ( 1973 ) formalizes and extends the BlackScholes results. Cox, Ross, and Rubinstein ( 1979 ) also use arbitrage arguments to derive an option pricing formula in a discrete time economy by assuming that risky stock prices follow a binomial random walk . This subject is covered in Chapter 8. They show that their formula converges to the BlackScholes formula when the
176
Foundations for Financial Economics
trading intervals shrink to zero and when appropriate limits of their price system are taken. The application of the BlackScholes op tion pricing formula to the pricing of corporate risky debt is adapted from Merton ( 1974) . The discussions in Sections 6 . 1 56. 18 are taken from Breeden and Litzenberger ( 1978) . An option pricing formula under the assumptions that individuals have negative exponential utility functions and that asset returns are multivariate normally distributed is derived by Brennan ( 1979) . For a review of the recent developments in option pricing theory and its applications see Cox and Huang ( 1987) .
References
Black , F . , and M. Scholes. 1973. The pricing of options and corpo rate liabilities. Journal of Political Economy 8 1 :637654. Breeden, D . , and R. Litzenberger. 1978 . Prices of statecontingent claims implicit in option prices. Journ al of Business 5 1 :6216 5 1 . Brennan, M. 1979. The pricing of contingent claims in discrete time models. Journal of Fin ance 34:5368. Cox, J . , and C . Huang. 1987. Option pricing theory and its applica tions. In Frontiers of Financial Theory. Edited by G . Constan tinides and S. Bhattacharya. Rowman and Littlefield. Totowa, New Jersey. Cox, J ., S. Ross, and M. Rubinstein. 1979. Option pricing: A sim plified approach. Journal of Fin ancial Economics 7 :229263 . Kraus, A . , and R. Litzenberger. 1976. Skewness preference and the valuation of risk assets. Journal of Fin ance 3 1 : 10851 100. Kraus, A., and R. Litzenberger. 1983. On the distributional condi tions for a consumptionoriented three moment CAPM. Journ al of Finance 3 8 : 1 38 1  1391. Merton, R. 1973 . Theory of rational option pric ing. Bell Journal of Economics and Management Science 4: 141183.
Valuation of Complex Securities with Preference Restrictions
177
Merton, R. 1974 . On the pricing of corporate debt: The risk struc ture of interest rates. Journal of Finance 29:449470. Rubinstein , M. 1976. The valuation of uncertain income streams and the pricing of options. Bell Journal of Economics 1:407425.
178
Foundations for Financial Economics References
Black , F . , and M. Scholes. 1973. The pricing of options and corpo rate liabilities. Journal of Political Economy 8 1 :637654. Breeden, D . , and R. Litzenberger. 1978 . Prices of statecontingent claims implicit in option prices. Journ al of Business 5 1 :6216 5 1 . Brennan, M . 1979. The pricing of contingent claims i n discrete time models. Journal of Finance 34:5368 . Cox, J . , and C. Huang. 1987. Option pricing theory and its applica tions. In Frontiers of Financial Theory . Edited by G . Constan tinides and S . Bhattacharya. Rowman and Littlefield. Totowa, New Jersey. Cox, J., S. Ross, and M. Rubinstein. 1979. Option pricing: A sim plified approach. Journal of Financial Economics 7:229263. Kraus, A., and R. Litzenberger. 1976. Skewness preference and the valuation of risk assets. Journal of Finance 3 1 : 10851100. Kraus, A., and R. Litzenberger. 198 3. On the distributional condi tions for a consumptionoriented three moment CAPM. Journal of Finance 38:13811391 . Merton, R 1973 . Theory of rational option pricing . Bell Journ al of Economics and Managemen t Science 4 : 14118 3 . Merton, R . 1974. On t h e pricing of corporate debt: The risk struc t:ure of interest rates. Journal of Finance 29:449470. Rubinstein , M. 1976 . The valuation of uncertain income streams and the pricing of options. Bell Journal of Economics 7:407425.
CHAP TER 7 MULTIPERI O D S E C U RITIES MARKETS I: EQ UILIBRIUM VALUAT I O N
In Chapters 5 and 6 , we discussed allocational efficiency and general pricing principles in two period economies. Most of the results there can be generalized in a straightforward manner to economies with more than two periods. In twoperiod economies with a single consumption good, individuals trade only at time 0. In a multiperiod economy with reconvening markets, there exists the possibility of trade after time 0. Therefore, individuals' expectations of future prices become an indispensable part of an economic equilib rium . This leads to the notion of a rational expectations equilibriu m, in which individuals' ex ante expectations about future prices are ful filled ex post . One of the most important features of a multiperiod economy is that the number of securities needed to complete the markets is in general much smaller than the number of states, since individuals can revise their po rtfolio compositions at each trading date. Th at is, all the state contingent cl aims can be manufactured dynamically by trading in a limited number of longlived complex 7.1.
179
HlO
Foundations for Financial Economics
securities  securities that are available for trading at every trading date. Thus , the equilibrium allocation will be Pareto optimal. In allocationally efficient securities markets, we can then follow the ar guments in Chapter 5 to construct a representative agent for pricing purposes. The sign of the risk premium of a risky security depends upon the covariability of its return with the aggregate consumption process. Only in very special cases can the aggregate wealth process, or the market portfolio, play the role of the aggregate consumption process. These are the subjects to which we now turn. 'T . 2 . Consider a multi period pure exchange economy under un certainty that has T + 1 trading dates indexed by t = 0, 1, , T. There is a single p erishable consumption good available for con sumption at each trading date. Any possible complete history of the exogenous uncertain environment from time 0 io time T is a state of nature and is denoted by w . We as before use 0 to denote the collection of all possible states of nature. In a twop eriod econ omy with time 0 and time 1, individuals at time 0 only know that the true state of nature is one of n, which will be revealed at time 1 . In a multiperiod economy, we will assume that the true state of nature is partially revealed to individuals over time and is completely revealed at the final date of the economy, T. The revelation of the true state of nature over time is best represented by an e v e nt tree , which we will term an info rmation structure . Figure 7.2 . 1 illustrates an informa tion structure when there are five possible states and three trading dates. The five possible states are denoted by w1 , . . . 1 ws . At time 0, the only information is that the true state is one of {w1 , . . . , w s } . At time 1, individuals learn that the true state of nature is either one of {w1 , w2 , ws } or one of {w" , w5 } . At time 2, the true state of nature is revealed. For example, suppose that the true state of nature is w s . Then at tim e 1, it will be learned that the true state i s one among the first three possible states and is neither w4 nor ws. At time 2, state ws iskn._own to be the true state. Before we proceed, we first fix some terminology and notation. An e ve n t is a subset of n. For example, {w1 , w2 , ws } is an event. In Figure 7 .2.1, this event simply says that the true state is one member of {w1 1 wz , ws } . Two events are said to be disjoint if they have an .
.
.
Multiperiod Securities Markets I: Equilibrium Valuation I = 0
Figure 1 . 2 . 1 :
I= I
I
•
2
181
An Information Structure
empty intersection, that is, if we know the true state is in one event, it cannot be in another. A p artitio n of n is a collection of events { A 1 , . . . , An} such that the union of these events is equal to n and the pairwise intersections of these events are empty. A given partition is said to be finer than another if any event in the l atter p artition is a union of some events in the former. In Figure 7.2 . 1 , the event n forms a partition at time 0. At tim e 1 , the two events {w1 , w2 , w3} and {w4 , w5 } form a partition. At time 2, the five individual states form a partition. The information revelation represented by an event tree can be described by a family of partitions of n, indexed by time, that become finer and finer. We will use F = {Ji; t = 0, 1, . . . , T} to denote the common info rmation structure with which individuals are endowed, where each J; is a partition of n with the property that 1t is finer than 111 if t 2:: s . We will always assume that 10 = {0} and 1T is the partition generated by all the individual states. This simply means that individuals know at time 0 that the true state is in n and they learn the true state by time T . A timeevent co ntingent claim i s a security that pays one unit of consumption at a trading date t ? 1 in an event ae E Jt and nothing
Foundations for Financial Economics
182
otherwise. There are I individuals in the economy indexed by i = 1, 2, . . . , /. To simplify notation, we assume that individuals have homogeneous beliefs about the possible states of nature denoted by 71'"111 , w E 0. As usual, we assume also that 71'"111 > O , 'Vw E n. These probabilities in duce probabilities about events. For an event at E 1t , the probability of its occurrence is ( 7 . 2 . 1) 7r a1 ::=: 71'"w . wEa1
L
Individual i has preferences for timeevent contingent claims that are represented by timeadditive and state independent von Neumann Morgenstern utility functions that are increasing, strictly concave, and differentiable : Uio ( zo) +
T
L L 7r a, uit ( za, ) . t = l a1ET1
(7.2.2).
Implicit in the assumption that an individual 's preferences are de fined on the timeevent contingent claims is the requirement that h e can not consume different amounts of consumption across t h e states that comprise the event at E 1t at time t. This is a natural infor mational constraint  as an individual at time t cannot distinguish among different states in any at E 1,, his consumption choice cannot therefore depend upon different states in an event in 1, . We also assume that an individual has a nonnegative endowment of time0 consumption and timeevent contingent claims denoted by {eb , e�, , at E 1t; t = 1 , . . . , T} .
To avoid degeneracy, we assume that there exists a t and a n event . at E 1i such that e�, > 0. Before we leave this section, we remark that the conclusions of Sections 7 . 3 to 7 .14 are valid when individuals have heterogeneous probability beliefs, as long as their probability beliefs all assign a strictly positive probability to each state w E 0 . Readers can draw on the analogy between the discussions of Sections 5.4 to 5 . 10 and those of Sections 7.3 to 7 .14. The former do not depend upon individuals
Multiperiod Securities Markets I: Equilibrium Valuation
183
having homogeneous probability b eliefs. We leave it for interested readers to show that the latter do not either. 'T
3 An allocation of time0 consumption and the timeevent contingent claims is said to be Pareto optimal if there does not exist another allocation that increases at least one individual's expected utility without decreasing any other individual's expected utility. As in Chapter 5, a Pareto optimal allocation .
.
{ ch , c�, a t E J; , t = 1 ,
. . .
, T, i = 1 , 2,
. . .
, /}
is the solution to the following maximization problem for a set of positive weightings ( � i ) [=1 :
{
..
; Ifax
•o•• , '." e E .7e . t= l , . . . ,T,t = l , 2 , . . . , 1
}t � i (uio ( � ) + t L 7ra, Uit (z�, ) ) t=l aeE.Te i=l I
I
= L eh
s.t. I >�
(7.3 . 1)
i=l
i=l I
I
I: z!, = L: e�, 'Vat E J; 'Vt . i=l
i=l
We shaU b e interested only in Pareto optimal allocations that cor respond to strictly positive �/s. The first order conditions that are necessary and sufficient for ( c� , c�, , a t E .1, , t = 1, . . . , T) [= 1 to be a solution to (7.3 .1) with strictly positive � ·s are
�iu � o ( ch ) = a. (at) = a . '+'a•
if t < s and
a8 � at.
(7.5.3)
We claim that the prices defined in (7 .5.1) and (7 .5 .3) and the com plete markets competitive equilibrium allocation { ch, c�, ; at E Jt , Vt = 1 , . . . , T} are a rational expectations equilibrium in which trading t ak es place only at time Individuals have rational expectations in the sense that they believe that prices will evolve according to (7.5 . 1 ) a n d (7.5.3). To prove our claim, it suffices to show that at any time t and in any event at E Jt, {c� . , aa E 1., a8 � a, , s 2: t } is a solution to the program:
0.
(7 .5.4) That is, at each trading date t, given his expectations of future prices, an individual is content to hold his competitive equilibrium alloca tion of the timeevent contingent claims. First note that { c� , a8 E fa , a8 � a t } is feasible for (7.5.4) , because the budget constraint is satisfied . Next, from the Lagrangian theory, as utility functions are concave, it is sufficient to demonstrate the existence of a constant ��. , the Lagrangian multiplier for the budget constraint of (7.5.4) , such that •
(7.5 .5)
Multiperiod Securities Markets I: Equilibrium Valuation
for all a8 E fs, a8 c at , and solution to (7.5.4) . Now let
s
>
t, for { c � • • a 8
t/>a, i Ia, = 'Yi , 
11"at
E
!'a, a8 �
at }
187
to be a (7.5 .6)
where 'Yi is the Lagrangian multiplier for the program (7 .3.5) . Substi tuting (7.5 .6) into (7.3.6) and (7.3.7) , and using (7. 5 .2) and (7.5.3) , w e immediately get (7.5 .5) . That i s , there exists a constant 1!t , as defined in (7.5 .6) , such that (7.5.5.) is satisfied. We thus conclude that individual i will be content to hold his allocation in a competi tive complete markets equilibrium at any time t and in any event at even when he has the opportunity to retrade. Therefore, there ex ists a rational expectations equilibrium with complete markets such that the equilibrium allocation is identical to t.hat in a complete mar ket competitive equilibrium. Note that in this rational expectations equilibrium, trading only occurs at time 0 . 1 . 6 . An individual's optimal consumption and portfolio deci sions in a multiperiod economy are simple when he has rational ex p ectations and when there exists a complete set of timeevent contin gent claims markets at time 0. This is so because he only has to solve (7.3 .5) at time 0. After time 0, there is no need to adjust portfolio holdings of the timeevent contingent claims; and, as a consequence, there is no value in markets reopening after time 0 . When there does not exist a complete set of the timeevent contingent securities, however, there may be value for the markets to remain open after time O, and an individual's optimal consumption and portfolio prob lem becomes more complicated . The following several sections are devoted to an analysis of multiperiod securities markets which lack a complete set of timeevent contingent claims. The most important lesson to be learned is that, with reconvening markets, the markets can b e completed by dynamically managing a portfolio of longlived securities, in number far fewer than the number of timeevents. 1.1. Before we proceed, we will first revise some definitions which were used before in twoperiod economies. A complex security
188
Foundations for Financial Economics
is
composed of units of time0 consumption good and a bundle of timeevent contingent claims and is represented by x = { xo , Xae i at E .1i , t = 1, . . . , T } , where xo and Xa, are the dividends paid at time 0 and at time t in event a t . respectively, in units of the consumption good. A long lived security is a complex security that is available for trading at all trading dates. A real world example of a long lived security is a common stock traded on a stock exchange. A multiperiod securities markets economy or simply a securities marke ts economy is composed of a finite number of longlived securities in strictly positive supply represented by their dividend processes x; , j = 0, 1 , . . . , N, and a .finite number of individuals indexed by i = 1 , 2 , . . . , [. Assume that the dividend process x; is nonnegative and is strictly positive at some time t in some event at E ft . Instead of being endowed with the time0 consumption good and timeevent contingent claims, individual i is endowed with shares of the long lived securities. Let 0� (0) denote the numbers of shares of security j with which individual i is endowed. To avoid degeneracy, assume that 9} (0) � 0 and, for every i , there exists some j such that li} (o) > 0. We also assume without loss of generality that the total number of shares of each security is one. Thus I
l: ii} (o) i=l
=
1 Vj = 0, 1, . . . , N.
As the total supply of a longlived security is one share, the number of shares held by an individual can be interpreted as the proportion of the total supply held. The exdividend price of security j at time t is a random variable denoted by S; (t) . It is n atural to require that S; (T ) = 0, for all J·. Being a random variable, realizations of S; (t) depend on the state of nature. So, formally, a realization of S; (t) can be denoted by .S; (w , t ) . As individuals at time t cannot distinguish among states of nature in an event at E 1t , a natural informational constraint of S; (t) is that its realizations not vary across states in an event at E Jt . In mathematical terminology, this property i s usually termed the measurability of S; ( t ) with respect to Jt. Individuals are assumed to have rational expectations in that they agree on the mappings S; (w , t) , j = 0, 1, . . . , N.
Multiperiod Securities Markets I: Equilibrium Valuation
189
A process is a collection of random variables indexed by t . For example, S; is a process. If S; (t) is measurable with respect to Jt for all t, then we say that the price process S; is adapted to F . Using this terminology, a longlived security is represented by a dividend process adapted to F. As there does not exist a complete set of the timeevent con tingent securities, individuals may not be able to trade to a Pareto optimal allocation at time 0. As a consequence, there may exist in centives for them to trade after time 0. Therefore, we must have a way of describing the manner in which individuals' portfolio compo sitions are adjusted over time. To this end, we introduce the notion of a trading strategy. A trading strategy 0 is an N + !dimensional process 0=
{O;(t) ; j = 0, 1, 2, . . . , N; t =
1, . . . , 1'} ,
where O; (t) is the number of shares of security j held from time t  1 to t before trading takes place at time t . Therefore, the values of O; ( t ) are determined at time t  1, and a natural informational constraint is that O; (t) be measurable with respect to 1t  l · In mathematical terminology, the process 0; is then said to b e predictable with respect to F, or simply predictable when there is no ambiguity about the information structure F. Note that a predictable process is certainly adapted. We shall denote ( Oo (t) , . . . , ON (t))T by O (t) , where as before T denotes the transpose. A consumption plan c is a process adapted to F :
c = {c(t) ; t
=
0 , 1 , . . . , T} ,
where c(t) is the random consumption at time t, in units of the single consumption good. A trading strategy 0 is said to be admissible if it is predictable and there exists a consumption plan c such that, V t ,
O (t + l)T S(t) O (t)T (S(t) + X(t))  c(t) , =
(7.7. 1 )
where S(t) = (So(t) , . . . , SN (t))T , and X(t) = ( x0 ( t ) , . . . , :�w (t))T . Note that in (7.7. 1 ) , we have used the convention that when t = T ,
190
Foundations for Financial Economics
the lefthand side is 0, since S; (T) = 0 for all j. Thus, at t = T , (7.7.1) i s just 8(T) T X(T) = c(T) . (7. 7.2) Relation (7.7. 1) is a natural budget constraint. The lefthand side of (7.7.1) is the value of the portfolio at time t after trading and consumption, while the righthand side is the value of the portfolio at time t before trading plus dividends received net of consumption . At the final date of the economy, consumption is equal to the total value of the dividends, which is natural because of the assumption that utility functions are strictly increasing. The consumption plan of (7.7.1) is said to be financed by the trading strategy 8. The space of admissible strategies will be denoted by H. Individual i's problem is to maximize his expected utility of con sumption by choosing a consumption plan financed by an admissible strategy. Formally, individual i's problem is
T.M uro (c(O)) + E
[t,
u11 ( c(t ))
s.t. 8(0) = iJi (o) and { c( t)} is financed by 8.
]
(7.7 .3)
A rational expectations equilibrium for the securities markets is composed of I pairs of admissible strategies and consumption plans {8 i , c i ; i = 1 , 2, . . . , I} and price processes for the longlived securities { S (t) ; t = 0, 1 , . . . , T} such that 8 i solves (7.7.3) , c i is financed by 8 i , and markets clear at all times: I
L 8} (t) = 1 i=l
for all t = 1, 2, . , T and all j = 0, . . , N. In a securities markets economy, the source of consumption is the dividends paid by longlived securities, in units of the consumption good. Therefore, m equilibrium, Walras law implies that, for all t = 0, 1 , . : . , T, .
.
.
I
L i= l
c i (t) =
N
L x; (t) . i =O
(7.7.4)
Multiperiod Securities Markets I: Equilibrium
Valu a tion
191
7 .8. In the above definition o f a rational expectations equi librium, each individual maximizes his lifetime expected utility of consumption at time 0. As markets are open and individuals are able to retrade at each trading date, for the above definition to make sense , it must be an equilibrium property that at each trading time t , there does not exist another consumption plan financed by a trading strategy admissible from t onward that strictly improves the welfare of any individual. We prove this property by contraposition. The idea is that if there exists such an improvement at time t � 1 , then the expected utility of lifetime consumption computed at time 0 can be strictly improved upon. This is a contradiction . Formally, suppose that there exists a trading d ate t with 0 < t < T , an event a1 E � . a trading strategy 8, and a consumption plan c starting at time t such that, for every state in the event a, and V k = 1 , 2 , . , T  t , .
.
B( t + l)T S( t ) = Bi ( t ) T (S( t) + X( t ))  c(t) , B(t + k + l)T S( t + k) = B( t + k)T (S( t + k) + X ( t + k))  c(t + k),
(7.8. 1 )
and T
LL
""a. (at ) us. ( ca. )
>
T
LL
""a. (a e ) us, ( c � . ) .
(7 .8.2)
Define a trading strategy and a consumption plan starting at time 0: if s � t; i f 8 > t and i n the events if 8 > t and i n the events if 8 < t; if 8 � t and in the events if 8 � t and in the events
a, !l a, ; a, � at ;
a, !l a t ; a, � a, .
It is easily verified that fJ is an admissible trading strategy and finances the consumption plan c. We claim that the consumption plan c yields a strictly higher expected utility at time 0 than does ci . Denoting the conditional expectation at time t by E[·l �], relation
192
Foundations for Financial Economics t
1 =0
•
I
t
=
2
y
2 3
4
5
Figure 7 . 8 . 1 :
Conditional Expectation of a Random
Variable
(7.8.2) implies that E
[t.
1 [t.
u;, (C(s) ) \ Ji � E
u;, (o' ( s)) \ Ji
1,
(7.8.3)
with the strict inequality holding in the event a t . Here we note that the conditional expectation at each time t > 0 is a random variable measurable with respect to 'ft. For example, consider the information structure depicted in Figure 7.2.1 and redrawn here in Figure 7 . 8 . 1 . The numbers at time t = 2 are possible realizations o f a random variable y. Assume that each branch is equally likely. Then the expectation of y conditional on 1i is a random variable: if the true state is a membe r of {w1 , w2 , w3 } ; if the true state i s a membe r of {wo& , ws } .
That i s, at time 1 , if the upper node is realize d, the co�dit ional s, the expec tation of y is equal to 2; while if the lower node reahze are condit ional expect ation of y is equal to 4 . 5 . These two numbe rs record ed in Figure 7.8.1 on the nodes at time 1 .
Multiperiod Securities Markets I: Eq uilibri um Taking expectation E
[t. [�
on
Valu ation
both sides of ( 7 .8.3 ) gives
u;, ( l(s
))
] [t. .!(s))] . ] [� ] >
E
193
u;, (
(7 .8 .4)
u;.( ci ( s ) ) .
(7.8.5)
By the definition of c, we know that t 1
E
u;. (c( s ))
=
t 1
E
Adding (7.8.4) and (7.8.5) yields (7.8.6) This contradicts the fact that ci is a solution to (7.7.3) . Therefore , in a rational expectations equilibrium, there is no incentive for indi viduals to deviate from the strategies chosen at time 0. 7.9. Optimal consumption policies and trading strategies in a multiperiod securities markets economy are traditionally character ized by dynamic programmin g . This procedure starts at time T  1 with one period remainin g. For notational simplicity, we will drop individual indices in this section and Section 7.10 and use Ut ( · ) to denote the utility function for some individual at time t. Naturally, Ut ( · ) is increasing, strictly concave, and differentiable. The problem faced by the individual at time T  1 is max
{ c(T:1),9( T )}
UT1 (c(T  1 )) + E[uT (c(T)) i 1T1]
s.t. c(T  1) + 8(T) TS (T  1)
= B(T 1) T (S (T  1) + X (T  1)) and c(T) = B(T)T X (T),
(7.9. 1)
where B(T 1 ) is the vector o f shares of longlived securities that are held from T  2 to T  1 and is 1T 2 measurable. We note that 
194
Foundations for Financial Economics
c (T  1) and O(T) are random variables measurable with respect to 1T  1 • Denote the maximand of (7.9.1) by V ( O (T  1 ) ; 1T t ) , which we term the indirect utility function at time T  1 . Embodied in this notation is the fact that, given O (T  1 ) and 1T b the individual's maximization problem is welldefined and thus so is V ( O ( T 1 ) ; 1Tt ) . Given 1T t 1 the individual knows at which "node" along the event tree representing F he is at time T  1 and thus has "correct" expec tations about the X (T) to be realized at time T . Knowing O ( T  1 ) , the portfolio that the individual holds from time T  2 t o time T  1 , a. I lows him to know the value of his portfolio and to meet his budget constraint at time T  1 . Now go t o time T  2 . The problem facing the individual is
{
T
I.:
}
max 'ILT2 (c(T  2) ) + E[ Ut ( c ( t ) ) I 1T 2 ] (c(t)) ;,:�_ 2 T t= 1 (B(t))f=T1 s.t. c(t  1) + O (t) T S (t  1) = O (t  1 ) T ( S (t  1 ) + X (t  1 ) ) , t = T  1 , T, and c(T) = O ( T ) T X (T) .
(7.9.2)
We similarly define the maximand of (7.9.2) to b e V ( O ( T  2 ) ; 1T 2 ) , the indirect utility function at time T  2. Using the arguments of Section 7 .8, it is easily seen that the optimal consumption plan and trading strategy chosen at time T  2 must remain optimal at time T  1 . This implies that once decisions about c( T  2 ) and O ( T  1 ) are made, the maximal expected utility at time T  1 (for future consumption) conditional on 1T  1 is V (O (T  1 ) ; 1T t) · Thus (7.9.2) is equivalent to max
{ c( T2),e(T1)}
U T2 (c(T  2) ) + E[V (O (T  1) ; 1T t ) l 1r 2 ]
s.t. c(T  2) + O (T  1) T S ( T  2) = O (T  2) T ( S (T  2) + X (T  2) ) .
In general, at any time t
=
(7.9.3)
0 , 1 , . . , T  1 , the individual solves .
Multiperiod Securities Markets I: Equilibrium Valuation
195
the following program: max
{ c(t) ,IJ(t+ I)}
Ut (c(t)) + E[V (O( t + 1 ) ; Jt + I ) ! Jt ]
s.t. c(t) + O (t + 1 ) T S (t)
=
O(t) T (S (t) + X(t) ) ,
(7.9.4)
where V (O(t + 1 ) ; Ji+ I ) is the indirect utility function at time t + 1 . That is, the dynamic problem the individual faces is equivalent to a sequence of twoperiod problems: At every time t < T , the individual chooses a consumption c(t) and a portfolio strategy O(t + 1) to maximize his consumption preferences for c(t) and his preferences for the random wealth at time t + 1. Note that the random wealth at time t + 1 can be computed if we know O (t + 1) and S(t + 1). Moreover, a n individual has rational expectations, therefore, h e has the correct expectation about S (t + 1) given Jf . This explains the dependence of V on O (t + 1) and 'ft. The preferences for the time t + 1 random wealth are represented by the indirect utility function V (O(t + 1) ; Ji+t) , which represents the maximal expected utility for future consumption conditional on Ji+ t when the random wealth at time t + 1 is equal to O(t + 1) T (S(t + 1 ) + X (t + 1) ) . 'T .10. Now we derive a property of an optimal consumption plan which will be useful later in this chapter. S uppose throughout this section that (c, O) is optimal . As mentioned in Section 7.9 , the wealth of the individual at time t is
W (t)
:=
O (t) T (S(t) + X(t) ) .
(7. 10. 1 )
I t i s clear from (7.9 .1) and (7.9.2) that, at any time t , the role o f O(t) is to allow the individual to satisfy his or her budget constraint:
c(t) + O (t + 1) T S (t)
=
O (t) T ( S(t )
+
X (t)) = W (t) .
(7. 1 0 .2)
Therefore, the indirect utility function at time t can equally well be written a s V (W (t) ; Ji ) . Then (7.9.4) implies that, along the optimal solution,
V (W (t ) ; Ji ) = U t (c(t ) ) + E[V ( W ( t + 1 ) ; 'fi + I ) I Jt] .
( 7 . 10 . 3)
196
Foundations for Financial Economics
Exercise 7.3 asks the reader to show that V (W (t) ; Ji ) is increasing and strictly concave in W (t) . We shall assume that V is differentiable in W (t) . Differentiating both sides of (7. 10.3) with respect to W (t) while noting (7.10.2) gives
Vw (W (t) ; Ji ) =
u
� ( c (t ) ) ,
(7. 10.4)
where Vw denotes the partial derivative of V with respect to W (t) . That is, along the optimal solution and at any time t , the marginal utility of wealth must be equal to the marginal utility of consumption  this is the socalled envelope con dition . By strict concavity, u� and Vw are strictly decreasing. Thus, there exists a function g, dependi n g on Ji , such that (7. 10.5) c( t ) = g (W ( t ) ; 1t)
and g is strictly increasing in W (t ) . The dependence of g on Jt sig nifies the possibility that c ( t) is not a deterministic function of W (t) and may well depend also on the events in 'ft. Relation (7.10.5) for malizes the intuitive notion that the higher the wealth an individual has at any time, the higher the consumption chosen at that time will be, ceteris paribus. 7 1 1 In the complete market rational expectations equilibrium discussed in Section 7 .5, the number of timeevent contingent claims available for trading at time 0 is equal to .
.
T
2: # (Jt ) , t= l
where # ( Jt ) i s the numb�r o f disjoint �vents i n Jt . This number is equal to the total number of nodes, excluding the node at zero, in the event tree that represents the information structure F . With these timeevent contingent claims, the equilibrium allocation is P areto optimal. In the securities markets of Sections 7.6 and 7.7, there does not exist a complete set of timeevent contingent securities. Thus, the markets are not complete at time 0, and individuals may not achieve
Multiperiod Securities Markets I: Equilibrium Valuation t
=
0
t= I
t
=
2
197
( �) ( �) w3(� ) w4 ( i ) WI
w2
Figure 7.1 2 . 1 : An Example of a Securities Marlcet
their optimal allocation of timeevent contingent consumption by simply trading once at time 0. However, the opportunity to trade af ter time 0 may allow individuals to reach a Pareto optimal allocation even without a complete set of the timeevent contingent securities. Helped by the discussion of Section 5.7 about market completion by complex securities in a twoperiod economy, our intuition suggests that markets can be completed by dynamic trading in longlived securities if, at every node on the event tree representing the infor mation structure, the markets are complete for the subsequent nodes. In a twoperiod economy, market completion by complex securities can be ensured by havin g complex securities, equal in number to the number of states of nature, that have linearly independent time1 random payoffs. In a multiperiod economy, however, the market completion of a twoperiod subeconomy will depend upon the en dogenous prices of the longlived securities. This is a major dis tinction b etween a twoperiod economy and a multiperiod economy. This point will be made clear by the example discussed in the next section. 7.12.
Consider the event tree in Figure 7 . 1 2 . 1 . We will not
198
Foundations for Financial Economics
specify probabilities for different states of nature except that each state is of a strictly positive probability. The following analysis is valid for all probability specifications that assign a strictly positive probability to every state. Excluding the node at time 0 , there are seven nodes. That is, in a complete markets competitive equilibrium, we need seven timeevent contingent securities. Suppose now that in a securities markets economy, there are three longlived securities that do not pay dividends until time 2. Security 0 pays one unit of consumption at time 2 in all states and has unit price before time 2 . Securities j = 1, 2 are risky, and their prices at each timeevent before time 2 and their dividends paid at time 2 are listed along the event tree. For example, if the true state is one of {wi , w2 , w3} at time 1, the price for security 1 is 5 and that for security 2 is 3 . The prices and dividends specifi ed are certainly adapted to the information structure represented by the event tree. One question immediately comes to mind. Are the prices spec ified for the longlived securities reasonable? For example, do they admit arbitrage opportunities? We will have a complete answer to this question in Chapter 8 . For now, we just say that these prices do not admit any arbitrage opportunities and withhold the reasoning until Chapter 8 . We claim that any timeevent contingent claim can b e created by dynamically managing a portfolio of the three longlived secu rities. We shall demonstrate how to design a trading strategy to replicate the payoff of the timeevent contingent security that pays one unit of consumption at time 2 if and only if W I i s the true state. Other timeevent contingent claims can be replicated by analogy. We use the idea of dynamic programming and work backwards. Suppose that the true state is one of {w 1 1 w2 , w3 } . Then we will be at the upper node at time 1 . The task to be accomplished is to design a portfolio having a payoff equal to 1 at time 2 if WI is the true state and nothing otherwise. Let :z: be the number of shares of the riskless asset held, and let Yi be the number of shares of security j = 1 , 2 held to create a payoff of 1 at time 2 if and only if WI is the true
199
Multiperiod Securities Markets I: Equilibrium Valuation
state. Then ( x , y1 1 Y2) must satisfy the following: X+
3 y l + 2 y2 = 1 , X + 4 yl + 3 y2 = 0, X + 8y 1 + 4 Y2 = 0.
( 7.12 . 1)
The unique solution of ( 7.12.1) is ( x , Y1 1 Y 2) = ( 8 / 3 , 1 / 3 , 4/3).
( 7.12.2)
That is, buying 8/3 shares of the riskless asset, 1 / 3 shares of security 1, and short selling 4 / 3 shares of security 2 at time 1 if the upper node is reached creates a payoff of 1 at time 2 if w1 is the true state and nothing otherwise . If, on the other hand, the true state is either w4 or ws , at time 1, we will be at the lower node. Since the p ayoff we are replicating is the timeevent contingent security that pays off only at time 2 in state w1 1 we do nothing at the lower node. Thus, the payoff at time 2 will be zero in states w4 and ws . We go back one period to time 0. From the previous calculation, we learn that to replicate the timeevent contingent claim paying off at time 2 in state W i t at time 1 we need to follow the strategy specified in ( 7 .12.2) at the upper node and take no position in the securities at the lower node. To implement this strategy at time 1 , we need 1 / 3 units of the consumption good at the upper node and 0 units of the consumption good at the lower node. Thus the question at time 0 is what p ortfolio to purchase in order to have, at time 1 , a payoff of 1/3 at the upper node and 0 at the lower node. Again let ( x, y 1 1 y 2 ) denote the number of shares of longlived securities. We must solve X+
5 y l + 3 y2 = 1 / 3, X + 7 Yl + 4. 5y2 = 0.
( 7.12.3)
There exist many solutions to ( 7.12.3) . One of the solutions is
(7.12.4) The cost of this strategy at time 0 is
200
Foundations for Financial Economics
One can also verify that all solutions of (7.12.3) have the same cost. Now we are ready to state the trading strategy that replicates the timeevent contingent claim under consideration. At time 0, purchase 7/6 shares of the riskless security and short sell 1/6 shares of security 1. This costs 1/6 units of the consumption good . At time 1 , if we are at the upper node, the portfolio constructed at time 0 will have a value equal to 1/3 units of the consumption good. We adjust our portfolio composition according to (7.12.2), which costs exactly 1/3. The portfolio of (7.12.2) has a payoff of 1 at time 2 if and only if w1 is the true state. If the lower node is realized at time 1 , the portfolio purchased at time 0 is worthless and we close out our position and do nothing. The dynamic strategy specified in (7.12.2) and (7.12.4) replicates the timeevent contingent claim that pays off at time 2 in state w1 . 7 . 1 3 . In the example considered in the previous section, the number of timeevent contingent claims is equal to seven  the total number of nodes in the event tree of Figure 7.12.1, excluding the node at time 0. However, markets can be completed in a securi ties markets economy by dynamically trading only three longlived securities. Note that the three longlived securities of Section 7.11 dynamically complete the markets; because, for every node o n the event tree of Figure 7 .12.1, the prices of the three securities at the subsequent nodes form a matrix whose rank is equal to the number of subsequent nodes. That is, markets are complete for every one period subeconomy. For example, if the true state is w 1 , we will be at the upper node at time 1. The prices for the three longlived securities at the subsequent three nodes form a 3 X 3 matrix
which is of rank 3. Certainly, dynamic completion o f markets depends o n the price processes of the three longlived securities. Consider a trivial ex ample in which all longlived securities have proportional dividend processes. Their prices at every node will simply be proportional,
Multiperiod Securities Markets I: Equilibrium Valuation
201
and dynamic trading does not provide the investors with more span ning than does trading once at time 0. On the other hand, for the markets to be dynamically complete , the number of longlived secu rities traded must be no less than the maximum number of branches leaving each node on the event tree representing the information structure. In the example presented in Figure 7 . 1 2 . 1 , this number is three. Suppose that the number of traded longlived securities is strictly less than three. If the upper node is realized at time 1, the spanning property fails for the subsequent nodes, because the num ber of securities is strictly less than the number of states that will p ossibly be realized at time 2. Therefore, the number of longlived securities being no less than the maximum number of branches leaving each node is a necessary but not sufficient condition for the dynamic completion of markets. It is a sufficient conditi on in a. gener i c sense, however. This is the subject to which we now turn. In the rational expectations complete markets equilibrium of Section 7.5, the arguments of Section 5.8 can be used to show that the exdividend price of a complex security or a longlived security x at time t in event at E Jt is 7 .14.
Sz ( a t , t) =
T
L L
l= t+ l
Ao E T0
••
s;; .t
t/>a. ( at ) xa.
(7. 14. 1 )
(7. 14.2) =
where the third equality follows from the defi n ition of Sz (att1 , t + 1 ) . Recall that 7ra. ( at ) is the conditional probability of event a., at time t given that the true state is in at . Let Bz (t) denote the random exdividend price for the complex security x at time t . Then ( 7 . 14.2)
Foundations for Financial Economics
202
and ( 7 . 1 4 . 3) can be rewritten as
[ LT [ u�t+ul(� (cc;i((tt+) ) (x (t +l + S (t + t l x (t) Bz (t) =
E
=E
s=t + l
'
u�8 ( ci ( s )) u'. ( ci ( t )) x ( s ) ! Jt t!
1))
1)
( 7 . 14.4)
1)) ! Jt ,( 7.14.5)
where is the random dividend paid by security x at time .t. A securities markets rational expectations equilibrium is said to implement a complete market rational expectations equilibrium if the following three conditions are met: First, the prices for the long lived securities in the securities markets equilibrium are the same as those in the rational expectations complete markets equilibrium determined according to ( 7.14.4) and ( 7.14.5). Second, the markets are dynamically complete in the securities markets rational expecta tions equilibrium. Third, the equilibrium allocations under the two market structures are identical. Now let N be the maximum number of branches leaving each node of the event tree representing the information structure , and let X denote the collection of all the possible longlived securities. The generic sense of sufficiency referred to at the end of Section 7 .13. is as follows: Fix a complete market rational expectations equilibrium. If we randomly pick N longlived securities from X and assign their prices at each node using the timeevent prices of the complete mar kets rational expectations equilibrium, then, with probability one, the N longlived securities dynamically complete the markets. In this probability one event , it is easily verified that there exists a se curities markets rational expectations equilibrium having the same equilibrium allocation as the complete markets rational expectations equilibrium. That is , with probability one, a complete markets ra tional expectations equilibrium can be implemented in a securities markets rational expectations equilibrium by randomly selecting N longlived securities. Conversely, it is also easily seen that every se curities markets rational expectations equilibrium with dynamically complete markets corresponds to a complete markets rational ex pectations equilibrium, as the complete set of timeevent contingent claims can be manufactured by dynamic trading of the longlived securities.
Multiperiod Securities Markets I: Equilibrium Valuation
203
The proof for the above generic implementability is rather tech nical, and we refer interested readers to Kreps ( 1 982) for complete details. 7.15.
From the previous discussion, the number of longlived securities required to complete the markets dynamically in a secu rities markets equilibrium is no larger than the number of distinct timeevents and may be much smaller. The minimum number of longlived securities needed is determined completely by the way information is revealed over time or by the temporal resolution of uncertainty . If uncertainty is resolved gradually, that is, there are few branches leaving each node and thus the am ount of informa tion to be revealed the next period is small, the minimum number of longlived securities needed is small. On the other hand , if, at some node , the amount of information to be revealed is large, the mini mum number of longlived securities needed is large. Intuitively, at each node, for every possible independent source of uncertainty to be resolved the next period, there should be a security or a portfolio of securities having payoffs contingent upon that u ncertainty. Here we note that the ideas discussed in Sections 5. 1 D5. 19 can be generalized in a straightforward manner to multiperiod economies. For example, since we have assumed in this chapter that individuals have timeadditive stateindependent utility functions and h omoge neous b eliefs, it can be shown that an individual's Pareto optimal al location of timet consumption is a strictly increasing function of the timet aggregate consumption/endowment. Thus as long as securi ties markets are dynamically complete with respect to the aggregate consumption "events" , a Pareto optimal allocation will be achieved. Other generalizations are left for the readers. Henceforth, we will only consider securities markets economies with dynamically complete markets unless we specify otherwise. 'T . 16. Recall from Sections 5.21 and 5.23 that in a twoperiod competitive equilibrium with complete markets , when individuals have homogeneous beliefs and stateindependent and timeadditive
204
Founda tions for Financial Economics
utility functions, a representative agent with the same probability be liefs and with a stateindependent and timeadditive utility function can be constructed to support equilibrium prices by consuming the aggregate endowment. In our securities markets economy, the same conclusion holds. That is, a representative agent can be constructed in equilibr ium so that equilibrium price processes can be supported by this representative agent endowed with all the longlived securi ties. The optimal trading strategy of this representative agent is to hold the aggregate supply of longlived securities throughout with out trading, and the optimal consumption plan is simply to consume the aggregate dividends paid at each date and in each event . The representative agent's utility function can be constructed as follows. F irst we note that, as markets are dynamically complete in equilibrium, we can compute the complete set of timeevent con tingent claim prices, denoted by ¢a1 • The equilibrium consumption plan of individual i, must be a solution to (7 .3.5) . Let "'/i be the Lagrangian multiplier for the budget constraint. Putting = '"fi\ we define
ci , U
t(
z)
=
.>.i
I
L .>.i Uit(zi)
max (ll;)f=l i=l s.t.
I
(7.16 . 1)
L Zi = z, i=l
for all = 0, 1 , . . . , T . Using the same arguments as in Section 5 .22, it is straightforward to show that equilibrium prices i n the securities markets economy are supported by an agent with the utility functions defined in (7.16.1 ) , probability beliefs 7r a1 1 and endowment of the total supply of longlived securities. Employing ( 7 . 1 4. 5 ) , the price process for a longlived security j can then be written as
t
S; (t  1 ) where
=
E
[
uH C t) (z; (t) S; (t)) l 1t t , + Utl (Ct  d 1
I
N
i=l
j=O
Ct L c i (t) = L x;(t) =
l
( 7 . 16.2 )
Multiperiod Securities Markets
1:
205
Equilibrium Valua tion
is the random aggregate consumption at time t . Defining ( + S; (t) r;t = x; t) S; ( t  1 ) _
_
 1,
the rate of return for holding one share of security j from time t  1 to time t , (7.16.2) implies that
1=E = Covt 1
[
_ l ) [
uH 8t) ( 1 + r;t) 1i1 �_ u 1 (Gt t ) I
(
_
]
uHCt ) 1 uHC t)  l .rt:r 1 J , , + r;t + E I Yt  1 E [ 1 + v;t u�_ 1 (Gt 1 ) u�_ 1 (Gt1)
(7. 16.3) where Cov t ( ·, ·) denotes the conditional covariance operator given the information at time t . To facilitate our analysis, w e will assume that t h e 0th long lived security is riskless in that it does not pay dividends until time T and its price at time t < T is
where r!• is the riskless interest rate from time 8  1 to time 8 . Note that r/t can be stochastic, provided that the value of TJt is known at time t  1 . The d ividend paid at time T by security 0 is
Then ( 7. 16.2 ) implies
(7.16.4) Now we can proceed as in Section 6.2 by substituting ( 7. 16.4 ) into ( 7. 16.3 ) to get
E [ r;t l .1i1]  r ft =  ( 1 + r1t) Covt1 (r;, . u� (Ct)/u�_ 1 (CtJ )) .
( 7.16.5).
206
Foundations for Financial Economics
In equilibrium, the risk premium of security j at time t  1 is positive if its resale value plus its dividend at time t is positively correlated with the time t aggregate consumption, and vice versa. The intuition for (7. 16.5) is the same as for (6.2.8) . Note that in Section 6.2, we used the observation that in a two period economy, time1 aggregate consumption is equal to time1 aggregate wealth to substitute aggregate wealth for aggregate con sumption in a relation like (7. 16.5) . In a multiperiod securities mar kets economy, there are more than two d ates in the economy. Hence, the above observation is only valid at the final date of the economy. At any time before time T , aggregate consumption is only a fraction of the aggregate wealth. The aggregate wealth at any time t is com posed of the aggregate consumption or the aggregate dividends at that time and the exdividend values of all the longlived securities. As long as the aggregate exdividend value of long·lived securities is not zero, aggregate consumption will not be equal to aggregate wealth. Even though aggregate consumption is not equal to aggregate wealth , there exists a functional relation between the two. Con sider the representative agent in the dynamic programming context of Sections 7.9 and 7 . 10. Relation (7 . 1 0.5) implies that there exists a function g ( · ; Jt) such that
where Mt denotes aggregate wealth at time into (7 .16.5) gives
E[i;t l lt1]  rft
t.
(7. 16.6)
Substituting (7. 16.6)
=  ( 1 + rft ) Cov tl (i;t. u� (g (Mt ; Jt ) ) / u�  l (g(Mt  l i Jt _ l )) . (7 . 1 6.7) . As g may not be a deterministic function of just M and t, the sign of the covariance of (7 .16. 7) may not be determined by just the sign of Cov t  l ( i;t . Mt ) . Hence, there may not be a marketbased pricing relation . Under a set of very restrictive con d i t i on s , howe ver, the relationship between Ct and Mt is deterministic  independent of the events in Jt . As a consequence, a marketbased asset pricing relation holds. This is the subject to be covered in Section 7.22.
Multiperiod Securities Markets I: Equilibrium Valuation
207
7 . 1 7 . Explicit pricing formulae can be derived by making as sumptions on individuals' utility functions, as in Section 6.3, and uti lizing ( 7.16.5 ) . For example, suppose that the representative agent's utility functions are quadratic
( 7 . 1 7 .1 ) where p represents the time preference. As in Section 4.16, we note that individual i reaches satiation at time t when z � at /bt . We therefore assume that aggregate consumption at all times is such that satiation will not be attained. Substituting ( 7.17.1 ) into ( 7.16.5 ) , we get E[r;e l 1t t] =

 rIt
( 1 + rI t )Cov tl
(
;,t. ,
P t ( a t  b t Ct ) P t  1 (at  1  bt  1Ct d N
)
•
( 7. 17.2 )
In the arguments in Sections 4 . 16 and 6.2, we also utilized the fact that in a two period securities markets economy there exists a portfolio of securities whose time1 payoff is perfectly correlated with that of the aggregate consumption/ wealth, namely the market portfolio itself. Therefore, the risk premia of risky securities can be expressed in terms o f their betas with respect to the market portfolio and the risk premium on the market portfolio. In a multiperiod se curities markets economy, however, there is no guarantee that there exists a p ortfolio of securities whose p ayoff is perfectly correlated with the aggregate consumption process. We therefore choose a dy namically rebalanced portfolio of securities whose rates of return over time are most highly correlated with the aggregate consumption pro cess. Denote the rate of return of this portfolio at time t by ict . As ict is the rate of return on a feasible portfolio, it must also satisfy ( 7.17.2 ) : (7 .17.3 )
203
Foundations for Fin ancial Economics
Substituting
(7. 17.3 )
into
(7.17.2)
gives
(7.17.4) Now define the
consumption beta
at time
t1
of security j to be
(7. 17.5 ) where Vart ( · ) is the c onditional variance at time t . Dividing
Vart 1 ( Ct)
into the numerator and the denominator of the righthand side of
(7. 17.4)
gives
EfY;tl 1tt]  rft = This is the
henceforth
�;c,_1 (E[rct l h 1]  TJt) · fJCC t  1
(7.17.6)
Consumption Capital Asset Pricing Model , abbreviated CCAPM. Other pricing relations for d ifferent utility
as
functions can be derived. We leave them t o readers. 7.18.
As a historical note, a consumptionbased asset pricing
model of the form (7.16.5) was first derived by Rubinstein
( 1 976)
in a
discrete time single agent economy. Later, Breeden and Litzenberger
( 1978)
d erived a similar relation in a complete markets competitive
equilibrium with many heterogeneous individuals. shows that t h e
CCAPM
relation
(7.17.6 )
Breeden
(1979)
is valid in general in a con
tinuous time securities markets economy where local changes of the prices of longlived securities and the optimal individual consump tion are smalL We will give a heuristic derivation of Breeden's result. Recall from c alculus that
(7.18.1 ) if
M• ( t )
is small, where
209
Multiperiod Securities Markets I: Equilibriu m Valuation
(7.14.5)
Using
':t'
and the arguments used in deriving
E['��'i t I .rt  !J. J  rf t
=

( 1 + rf t ) Covt .0.
(
r;t ,
1
(7.16.5) , we h ave
u�t(ci (t))
"it  !J.
(. c' ( t
_
!J. ) )
)
,
(7. 18 .2) where r;t and rft now denote t h e rate of return on s ecuri ty j and t h e riskless rate from time t  !J. t o time t, respectively. Substitu ting ( 7. 18. 1) into (7.18.2) gives E[r;t l 1ttJ.]  rf t ( 1 + rft ) cOV t  !J.
� 
=

(
r;t ,
u�t(ci(t  tJ. ) ) + uft ( c'(t  !J. )) !J. ci (t)
"it  !J. ( c'' ( t  !J. ) ) u�� (c' (t  tJ. ) ) i ( 1 + Y'Jt ) 1 i( Covt!J. ( r; t � !J.c ( t ) ) , tJ. ) ) uit !J. ( c t I
_
where the second equality follows b ecause t 
!J.. Dividing both sides of
(7. 1 8.3)
and summing across i gives
by
ci( t  tJ. ) is known
)
(7. 18.3)
at time
 u�'t (c1 (ttJ.) ) /u�t!J. (c'(t tJ.))
(7. 18.4) where
tJ.Ct
==
I
L !J.Ci (t) = Ct  Ct  !J. i =l
Ct = C t !J. +
!J. Ct and
that Ct  !J. is known at time t  !J. . Let the rate of return from time
a n d the equality follows from t h e facts t h a t
t

!J.
to t of a p ortfolio of securities whose rate of return is most
h i ghly correlated with the aggregate consumption be denoted by ict · This rate of return must also satisfy
(7.18.4) :
(7. 1 8.5)
2 10
Foundations for Fin ancial Econ omics
Substituting get
(7.18.5)
into (7.18.4) and dividing by Var� ( Ct) , we (7. 18.6)
That is, the CCAPM holds approximately over the short time inter val [ t  � . t] . Note that in the above derivation , we require neither that markets be (dynamically) complete nor that utility functions be quadratic. Breeden (1979) shows that (7. 18.6) holds exactly in a continu ous time economy. His derivation uses stochastic calculus, which is outside the scope of this book. Interested readers should consult his paper for details.
·
7 . 1 9 . In Section 7. 16, we argued that in a multiperiod econ omy the aggregate consumption is not equal to the aggregate wealth except at the final date of the economy. Therefore, asset pricing relations depend upon the aggregate consumption process except at T  1. At any time t � T  1 , however, there should be some relation between consumption and wealth. If this relation is nonstochastic and onetoone, then consumption at any time can be expressed as a nonstochastic function of wealth. As a consequence, we can have a pricing relation that gives the risk premia of risky securities in terms of the stochastic properties of the market portfolio. In the remaining sections of this chapter, we will formalize the ideas in the previous paragraph and give sufficient conditions under which there exists a onetoone nonstochastic relation between con sumption and wealth at each trading date. As the reader will see, these sufficient conditions essentially make a multiperiod problem into a sequence of almost disconnected single p eriod problems. The only connections between adjacent periods are the wealth dynamics. Thus, there exists a nonstochastic relation between consumption and wealth when the multiperiod economy essentially lacks any intertem poral flavor. As an interim result, we will also derive an approximate multi beta asset pricing relation in Section 7 .21.
7.20.
To avoid technical complexity, w e will assume that there
Multiperiod Securities Ma.rkets I: Equilibrium Val uation
211
is a single representative individual in the multiperiod securities markets economy with utility functions ue (z) for time t consump tion, which are increasing, strictly concave, and differentiable. Let Y ( t) denote an Mvector of random variables that are observable at time t. A possible realization of this vector of random variables from time 0 to time t is a state of nature, or an w in the state space fl . Thus, a n observation of the realization o f Y from time 0 t o any time t tells us which states are possible and which states are not. When Y (t) can only take a finite number of values, as we will assume , all the possible realiz ations of Y from time 0 t o any time t generate a parti tion, Jt , of the state space. The sequence { Jt ; t = O , 1 , . . . , T} is then an information structure. We assume that the single individual is endowed with the information structure constructed above. We will call the vector Y the vector of state variables and Y ( t) the vector of state variables a.t time t. We also assume that Y (O) is a vector of known constants, so that J"o is just { n} . To begin with, we will put some mild structure o n Y. Fix t < T. Let z be a random variable, whose value depends upon { Y ( t + 1 ) , . . . , Y (T) } . We assume that E [ z i Jt J = E [ z i Y (t) J .
(7.2 0 .1)
That is, the conditional expectation of z given the information con veyed by { Y (O ) , . . . , Y ( t ) } is equivalent to the expectation at time t of z conditional only upon Y (t) . That is, the historical realizations of Y strictly before time t are irrelevant for the conditional expectation of a random variable whose value depends only upon the possible realizations of Y after t. This kind of memorylessness property is known as the Markov property. Assume that the dividend paid by longlived security j at time t depends only on Y ( t) and t. That is, x; ( t) = x; ( Y ( t) , t ) . The total supply of any longlived security is one share . The single individual is endowed with the total supply of longlived securities. The aggregate consumption at time t is thus C(Y (t) , t) =
N
L x; (Y ( t) , t) .
i =O
212
Foundations for Fin ancial Economics
In this single individual economy, it is straightforward to identify a rational expectations equilibrium for the securities markets. Define S; (Y(t) , t)
:=
=
E E
[8�1� [ �� t s
u � (C (Y ( s) , s) ) (Y ( s ) , s ) I Ji u a c (Y(t) , t) ) x;
]
u� (C (Y (s ) , s) ) ( Y ( ) ) Y (t) , s ,s I u a c (Y (t) , t) ) x;
]
(7.20 .2)
where the second equality follows from the Markov property. You are asked in Exercise 7.6 to verify that the prices defined in (7.20.2) , the consumption plan C (Y (t) , t) , and trading strategy O;(t) 1 for = all j comprise a rational expectations equilibrium for the securities markets. 7. 2 1 . In the rational expectations equilibrium established in Section 7.20, (7.16.5) is certainly valid. Thus , a consumptionbased asset pricing relation holds approximately in the framework of Sec tion 7.18. With a Markov structure , however, another linear asset pricing relation also holds approximately. This relation has more than one "beta." Formally, we assume that aggregate consumption , C (Y (t) , t ) , is a differentiable function of Y (t) and t. We will denote the partial derivatives of C (Y (t) , t) with respect to Y (t) and t by Cy (t) and C1 (t) , respectively. Note the Cy is an Mdimensional vector. Cal culus implies that
C (t) and uac (t))
�
C (t  � ) + Cy (t  � ) T � Y (t) + c,_6 (t  � ) � t ,
( 7. 21 . 1 )
�
(7.21 .2)
uac (t  � ) ) + u� (C(t  � )) � C (t) ,
when � C(t) , � Y(t), and � t are small, where � Y (t) ::: Y (t )  Y (t  � ) , � C (t) ::: C(t)  C(t  � ) ,
Multiperiod Securities Markets I: Equilibrium Valuation
and we have suppressed t h e dependence of O(Y (t) , t) stituting (7 .21.2) into (7 .21 . 1) gives
uao( t ) )
s=::�
(
on
213
Y (t) . Sub
uHO ( t  t.)) + u� ( O (t  t.)) o y ( t  t.) T t. Y (t)

)
+ Ot t. (t  t.)t.t . (7.21 .3) Now substituting (7 .21.3) into (7.18.2) and using the Markov property gives
E [r;t ! Y (t  t. )]  rft �' ( C (t  t.)) Cy (t t.) T ( 1 + rJc) u Cov,_ 6 (r;t , .t. Y (t)) , s=::� u�_ 6 (C(t t.))

_
_
Writing (7.21.4) in matrix notation gives
Vj. (7.21 .4)
(7.21 .5)
where 1 is an N vector of 1 's , it is an N vector of asset returns from time t  t. to t, and V z11 (t  t.) is the N X M matrix of the variancecovariance matrix of it and Y (t), conditional on Y ( t  .t. ) . Next, let iz,. t denote the rate of return, from time t  t. to t , on a portfolio whose rates of return over time are most highly correlated with !l. b the kth state variable. We denote (rz1 t , . . . rzM t ) T by izt · Each r11kt satisfies (7.2 1 .4), and in matrix notation we h ave
,
(7.21 .6)

where Vzy( t t.) denotes the M X M variancecovariance matrix of r11t a.nd Y (t) , conditional on Y (t  .t. ) . Substituting (7.21 .6) into
214
Foundations for Financial Economics
(7.2 1 . 5 ) gives
E [rt j Y (t  � ) J  "Jtl R:� Vz11 (t  � ) V..,11 (t  � )  1 ( E[r... t i Y (t  � ) J  rft l ) = V z11 (t  � ) V1111 (t  �)  1 (V..,11 (t  � ) V1111 (t  � )  1 )  1
( E [r..,t JY(t  � )]  rft l)
(7.21 .7)
=: B z11 (t  �)B..,11 (t  � )  1 ( E[r ..,t J Y(t  � ) J  rft l ) ,
where V11 11 (t  � ) denotes the M X M variancecovariance matrix of Y (t) conditional on Y (t  � ) and V11 11 (t  � )  1 denotes the inverse of V11 11 (t  � ) . Note that the jth row of B z 11 (t  � ) and the the kth row of B ...11 (t  � ) contain the "multiple regression coefficients" of r; t and r..,kt on Y (t) conditional on Y (t  � ) , respectively. That is, as an approximation, expected rates of return on risky assets are related to the rates of returns on M portfolios in a linear way. Merton ( 1973) derives an exact multibeta asset pricing relation in a continuous time economy. Relation (7.21 .7) is an approximation. 7.22.
The wealth of the single individual at time t is denoted by W (t ) . It is equal to the sum of the exdividend values of the longlived securities and the total dividends, at time t: N
W (t) = �)S; (Y (t) , t) + :z:; (Y(t) , t)) . j =O
(7.22 . 1)
The righthand side of (7.22. 1) depends only on Y (t) and t. Thus we can write W (t) = W (Y (t) , t) . Defining
'l (Y (t) , t) we can write
C (Y (t) , t)
_
=
C (Y (t ) , t) W (Y (t) , t) '
= r1 (Y(t) , t)W (t) .
(7.22.2)
That is, at each time t, the equilibrium consumption is a stochas tic fraction of the wealth, which depends on the realization of the
215
Multiperiod Securities Markets I: Equilibrium Valuation
state variables. Note that (7.22.2) is just a special case of (7.10.5). Substituting (7.22.2) into (7.16.5) gives
E [r;t l 1t 1]  rtt
(
)
uaq (Y (t) , t)W ( t))  (1 + rtt )Cov t  1 r;t , � _ . u (q (Y(t  1) , t  1) W (t  1)) ' (7.22. 3 ) . s o the wealth W (t) does appear i n the pricing relation . Unfortu nately, the dependence of q (t) on Y (t) may be nontrivial . Thus the covariance term on the righthand side of (7.22.3) may depend not only on W (t) but also on some part of Y (t) , and the covariability between the rate of return on a security and that on the market p ortfolio alone cannot determine the sign of the risk premium. Now we will provide a su fficient condition for consumption at each time to be a nonstochastic function of wealth so that the co variability between r;t and W (t) alone determines the sign of the risk premium for security j at time t  1. Let us assume that {Y ( 1 ) , Y (2) , . . . , Y (T)} is a sequence of independent random vari ables, that is, any partial observation of the sequence has no effect on the probabilities of its future realizations. More formally, let z b e a random variable whose value depends upon {Y (t + 1), . , Y (T) } . Then =
..
E[zi JtJ = E[z] . Using this independence assumption, (7.20.2) becomes S; (t)
_
=
=
[�
u� (C (Y( s ) , s ) )
E ��� a C (Y (t) , t) ) :z:; (Y ( s ) , s ) I Jt u E[ l:�· =H l u� (C(Y( s) , s )) z; ( Y ( s ) , s ) J uac(Y (t) , t) )
l
(7.22.4)
Note that the numerator of (7.22.4) depends only on the calendar time t. The denominator, however, is a strictly decreasing function of C(t), because utility functions are strictly �oncave. Therefore, w_ e can write S; (t) = S; (C (t) , t) . Moreover, we know that S; (C(t) , t) I S
216
Foundations for Financial Economics
a strictly increasing function of C (t) . Rewrite (7.22.1) as W (t)
N
=
_L)si ( Y ( t ), t ) + xi (Y (t) , t) ) j=O
=
=
N
N
L si ( C(t) , t) + L: xi ( Y (t) , t) j=O i =O N
L Sj (C (t) , t) + C( Y (t) , t) .
j=O
It is easily seen that the wealth at time t can be written as W ( t) = W ( C (t ) , t) and is a strictly increasing function of C (t ) . As a con sequence, there exists a function g such that g (W (t) , t) = C(t) and g is strictly increasing in W (t). Substituting C(i) = g(W(t) , i) into (7. 16.5) and using the Markov property gives E [iit i Y (t  1 ) ]  r!t =

 (
( 1 + r J t ) Co v t 1
uHg (W (t) , t)) rjt 1 u _ . ( g (W (t 1 ) , t � _
_
1))
)
(7.22.5). ·

Risky security } has a positive risk premium at time t 1 if and only if its rate of return at time t is positively correlated with W (t), the aggregate wealth at t, and vice versa. We have thus recovered a single period result by making an assumption strong enough to make a multiperiod problem into, essentially, a sequence of disconnected single period problems in the sense that the realization of the state variables before time t conveys no information about the stochastic properties of Y after t. Moreover, using a local approximation argu ment on (7 .22.5) as in Section 7 .21, we will get a singleb eta linear asset pricing relation with the rate of return on the market portfolio as the pivotal variable. We leave this for the reader in Exercise 7 .10.
Multiperiod Securities Markets I: Equilibrium Valuation
217
Exercises 7 . 1 . Let the consumption plan Show that
c
be financed by a trading strategy 6 .
t
O (t)T S(t) = O( o)T 8 (0 ) + L O (s )T [S(s)  S(s  1 ) ] +
s=l
t 1
t 1
L O (s)TX(s)  � c(s) . s =O
s=O
7.2. Show that if ( 7.5.3) is violated, there exists an arbitrage oppor tunity in a rational expectations competitive economy.
7 .3. Show that the indirect utility function of Section 7 .10 is increas ing
and strictly concave in wealth. 7 4 Consider a multi period securities markets economy with dynam ically complete markets. Show that the indirect utility function, as a function of wealth for an individual with timeadditive util ity functions; 1 pt z l b ' ut ( z ) = 1b has the following form, .
.
V ( Wt t ) ; Jt )
=
l ( t ; Jt ) W(t ) 1 b +
h(t ; Jt ) ,
where I and h are possibly stochastic functions depending on events in Jt . 7.5. Show in the context of Exercise 7.4 that when U t ( z) = pt ln z, the indirect utility function has the form V (W(t ) ; Jt)
=
I( t) l n W( t) + h( t; :Ft) ,
where we note that I is a· deterministic function of time.
7.6. In Figure 7 .12.1, we listed prices for two longlived securities along the event tree. Besides these two securities, there also ex ists a riskless security as described in Section 7 . 1 2. Show that the prices of these three longlived securities do not admit arbitrage opportunities. That is, there do not exist trading strategies that
2 Hl
7.7.
7.8. 7.9. 7 .10.
Foundations for Financial Economics
cost nothing at time 0 and finance consumption plans that are nonnegative and strictly positive at at least one node on the event tree. In the context of Section 7 . 1 1 , determin e the price of a complex security that pays $5 at time 1 if and only if the upper node is realized and pays $ 1 0 at time 2 if and only if Ws is the true state. Also, describe the dynamic strategy that manufact ures this comple x securit y. S how that the utility functions defined in (7. 16 . 1) are utility fun ctions for a representative agent in the sense of Chapter 5. Verify that the prices defined in (7.20.2) can be supported in a rational expectations equilibrium for the securities markets considered in Section 7.20. Use a local approximation argument on (7.22.5) as in Section 7.21 to get a singlebeta linear assel; pricing relation with the rate of return on the market portfolio as the pivotal variable.
Remarks. Sections 7.2 to 7.5 are freely adapted from Arrow (1964) and Debreu ( 1959). Note that if individuals do not have rational ex pectations and if markets remain open after time 0, there may exist speculation . A good reference for this is Svensson (1981) . When util ity functions are not timeadditive, there exist situations where there are incentives to trade after time 0 in a complete markets competitive economy. Interested readers should consult Donaldson , Rossman and Seiden (1980) and Donaldson and Selden ( 1981). The discussions on the issues relating to dynamically completing markets are borrowed from Kreps ( 1982). Our heuristic derivation of the CCAPM in con tinuous time is from Bhattacharya (1980) . Interested readers should consult Breeden's (1979) derivation. Recently, Duffie and Zam (1987) rederived the CCAPM more rigorously. There now exists a growing literature on the existence of an equilibrium when there are an in finite number of states of nature. Interested readers can consult MasColell (1986) , MasColell and Richard (1987), and Zam ( 1987), for example. Duffie and Huang (1985) discuss dynamic spanning when there are infinitely many states of nature in continuous time economies. The Markov economy constructed in Section 7.19 is adapted from Lucas (1978) and Prescott and Mehra (1980) , although the method
Multiperiod Securities Markets I: Equilibrium Valuation
219
of analysis is different. For generaliz ations of Lucas' model in ex change as well as production economies, see Brock ( 1982) , Cox, In gersoll, and Ross ( 1985) and Huang ( 1987) . Constantinides (1980, 1982 ) gives various conditions under which a single market beta asset pricing model is valid in exchange as well as production economies. Chamberlain ( 1987) derives a single market beta asset pricing model under conditions more general than those in Section 7.22. Hansen and Richard ( 1987) discuss the role of conditioning information in dynamic asset pricin g models. For a recent overview of asset pricing theories, see Constantinides (1987) . In many places in this and earlier chapters, we assert without proof that Cov (f(X), Y) and Cov(X, Y) have opposite signs when {() is pos itive and monotone decre a s ing. This is i n general not t rue. It is true, however, w hen
X = a + Y + e with Y and e independent. For example, (4 .24.4) on p .
111 i s true.
References
Arrow, K. 1 964. The role of securities in the optimal allocation of riskbearing. Review of Economic Studies 3 1 :9196. Bhattacharya, S. 198 1 . Notes on multiperiod valuation and the pric ing of options. Journal of Fin ance 36:163180. Breeden, D. 1979. An intertemporal capital pricing model with stochastic investment opportunities . Journal of Fin ancial Eco nomics 7 : 265296. Breeden, D . , and R. Litzenberger. 1978 . State contingent prices implicit in option prices, Journal of Business 5 1 :62165 1 . Brock, W . 1982. Asset prices in a production economy. In The Economics of Un certainty and Information. Edited by J . McCall. University of Chicago Press. Chicago. Chamberlain , G. 1987. Asset pricing in multiperiod securities mar kets. Mimeo. Department of Economics, University of Wisconsin Madison. Constantinides, G . 1980. Admissible uncertainty in the intertempo ral asset pricing model. Journal of Fin ancial Economics 8 : 7186. Constantinides, G. 1982. lntertemporal asset pricing with heteroge neous consumers and without demand aggregation. Journal of Business 5 5 :253267. Constantinides, G. 1987. Theory of valuation : Overview and recent
220
Foundations for Financial Economics
developments. In Frontiers of Financial Theory. Edited by G. Constantinides and S. Bhattacharya. Rowman and Littlefield. Totowa, New Jersey. Cox, J., J. Ingersoll, and S. Ross. 1985. An intertemporal general equilibrium model of asset prices. Econometrica 53:363384. Debreu, G. 1959. Theory of Value . Yale University Press, New Haven and London. Donaldson, J . , M. Rossman , and L. Selden. 1980. On the need to re vise dynamic consumption strategies in the presence of unchang ing ArrowDebreu preferences. Mimeo. Columbia University. Donaldson , J . , and L. Selden. 1 98 1 . ArrowDebreu preferences and the reopening of contingent claims markets. Economic Letters G :2092 16. Duffie, D., and C . Huang. 1 985. Implementing ArrowDebreu equi libria by continuous trading of few longlived securities. Econo metrica. 5 3: 1 337 1356. Duffie, D., and W. Zam. 1987. The consumptionbased capital asset pricing model. GSB Research Paper #922, Stanford University. Hansen, L . , and S. Richard. 1987. The role of conditioning informa tion in deducing testable restrictions implied by dynamic asset pricing models. Econometrica 5 5 :587614. Huang, C . 1987. An intertemporal general equilibrium asset pricing model: The case of diffusion information . Econometrica 5 5 : 1 1 7142. Kreps, D. 1982. Multiperiod securities and the efficient allocation of risk : A comment on the BlackScholes option pricing model. In The Economics of Uncertainty and Information. Edited by J . McCall. University of Chicago Press. Chicago. Lucas, R. 1978. Asset prices in an exchange economy. Econometrica 46: 14261446. MasColell, A. 1986. The price equilibrium existence problem in topological vector lattices. Econometrica 54:10391053. MasColell, A., and S . Richard. 1987. A new approach to the exis tence of equilibria in vector lattices. Mimeo. Economics Depart ment, Harvard University. Merton, R. 1973. An intertemporal capital asset pricing model. Econometrica 4 1 :867887.
Multiperiod Securities Markets I: Equilibrium Valuation
221
Prescott, E . , and R. Mehra. 1980. Recursive co mpetit ive equi librium: The case of homogeneous households. Econometrica 4 8 :13651379. Radner, R. 1972. Existence of equilibrium of plans, prices and price expectations in a sequence of markets. Econometrica 40:289303 . Svensson , L . 198 1 . Efficiency and speculation in a model with price contingent contracts. Econometrica 49: 1 3 11 5 1 . Rubinstein , M . 1976. The valuation of uncertain income streams and the pricing of options. Bell Journal of Economics '1:407425. Zam, W. 1985 . Competitive equilibria in production economies with an infinite dimensional commodity space. Mimeo. Department of Mathematics, State University of New York at Buffalo.
CHA P T ER 8 M U LT I P E RI O D S E C URITIES MARKETS I I : VALUAT I O N B Y A R B I TRAGE
8 . 1 . In Chapter 7 , we discussed equilibrium asset valuations in multiperiod economies . In this chapter, we will focus our attention on issues related to the pricing of derivative securities. Derivative securities are longlived securities whose payoffs depend on the price processes of other longlived securities. For example, a call option written on a common stock is a derivative security. In many situa tions, pricing of derivative securities does not rely on equilibrium ar guments  a derivative security may be priced by arbitrage arguments in the sense that if securities markets do not admit arbitrage oppor tunities, the price of the derivative security can be determined com pletely by the price processes of other longlived securities. When we price by arbitrage, we normally take the price processes of a set of longlived securities as given and price derivative assets. There is one question that naturally needs to be resolved, namely how do we know that the price processes we take as given do not admit any arbitrage opportunities? 223
224
Foundations for Financial Economics
The necessary and sufficient condition for price processes not to admit arbitrage opportunities is that they are related to martingales, to be defined, through a normalization and change of probability. This martingale connection also provides a direct way to compute price processes of derivative assets that can be priced by arbitrage. As an application, we use this method to price a call option when the price process of the under lying stock follows a binomial random walk. 8 . 2 . In this chapter we consider a multiperiod securities markets economy with a single perishable consumption good . We fix, until further notice , a state space 0 with a finite number of states of nature and an information structure F = { Jt ; t = 0, 1, . . . , T } . As in Chapter 7, we assume that the true state will be learned by time T and JQ = {0}. There are N + 1 longlived securities traded, indexed by j = 0, 1, . . . , N. As in Chapter 7, longlived security j is characterized by its dividend process x ; = { x; (t ) ; t = 0, 1, 2 , . . . , T } , where x; (t ) i s the random dividend paid at time t, in units of the consumption good. Dividend processes are adapted to F. To simplify the analyses that follow, we assume that the 0t h longlived security does not pay dividends until time T, at which time it pays one unit of the consumption good in all states. The 0th longlived security is just a Tperiod discount bond with a face value equal to one unit of the consumption good. We depart a little from the notation and setup used in Chap ter 7. The exdividend price of the 0th security at time t is denoted by B (t ) ; while the exdividend price of security j � 1 at time t is denoted by S; (t ) . As the price processes are exdividend , S; (T ) = 0 and B(T ) = 0. Naturally, B (t ) , and S; (t ) are random variables mea surable with respect to Jt. Since only relative prices are determined in an economic equilibrium, we assume without loss of generality that prices of longlived securities are in units of the single con sumption good. That is , the spot price of the consumption good is unity throughout. Individuals in the economy, indexed by i = 1, 2, . . . , I, h ave timeadditive von NeumannMorgenstern utility functions Uit ( · ) that
Multiperiod Securities Markets
II:
Valuation by Arbitrage
225
are increasing, strictly concave, and differentiable. For ease of expo sition, we assume that u�t (O) = oo. This will ensure that individuals always choose strictly positive consumption , which is a plausible eco nomic condition. Individual i is endowed with a probability belief 1r i = {1r� ; w E 0} about the states of nature and with shares of longlived securities (i.ii (O) , Oi (o)). Since B (t ) and S (t ) are random variables and depend on the states of nature, so they are functions of w and t. We assume that individuals have rational expectations in that they agree on the mappings B (w, t ) and S; (w , t ) . We assume that, for every i , 1r i assigns a strictly positive probability to every state. A trading strategy is denoted by {a ( t ) , O ( t ) = (O; ( t )) f= 1 } , where a(t) and O;( t ) are the random number of shares of the 0th and the ;'th security, respectively, held from time t  1 to time t, before trading occurs at time t. A trading strategy is a predictable process. A consumption plan c = {c(t ) ; t = 0, 1, . . . , T} is a process adapted to F, where c(t ) denotes the random units of timet consumption. A trading strategy ( a, 0) is said to be admissible if there exists a consumption plan c such that, Vt,
a(t + l) B ( t ) + O ( t + If S(t) = a (t ) B ( t ) + O ( t f (S (t ) + X( t ))  c (t) ,
(8.2.1) where S (t) = (St (t ) , . . . , SN (t ) ) T , X (t) = (z1 (t ) , . . . , xN (t ))T , and O (t ) = (Ot ( t ) , . . . , ON (t))T . Note that in (8.2.1) we have used the convention that the lefthand side is identically equal to 0 when t = T. Thus, at t = T, (8.2.1) becomes
a (T) + O (T)T X(T) = c (T) .
(8.2.2)
The consumption plan of (8 . 2.1) is then said to be financed by (a, 0 ). Alternatively, a consumption plan is said to be m arketed i f it is .fi · nanced by an admissible trading strategy. Note that the definition of a marketed consumption plan does not depend upon any of the individuals' probability beliefs 1ri . They will certainly agree whether a consumption plan is marketed and whether a consumption plan is financed by an admissible trading strategy.
226
Foundations for Fin ancial Economics
8 . 3 . AB we mentioned in Section 8 . 1 , one of the major purposes of this chapter is to price derivative securities while taking a price system ( B, S) as given. Therefore, our first order of business is to give conditions on (B, S) such that it forms a "reasonable" price system. As the weaker our requirement of reasonableness is, the stronger the implications of our requirement will be, we would like to impose as weak a condition on (B, S) as possible . Ideally, the requirement should be a necessary condition for any equilibrium price system so that its implications would apply to all equilibrium price systems. As individuals in the economy strictly prefer more to less, it fol lows that a necessary condition for a price system to be an equilib rium price system is that it cannot admit the possibility of something being created from nothing, or admit the existence of arbitrage op portunities . This necessary condition of an equilibrium price system will be our requirement for a reasonable price system. Formally, an arbitrage opportunity is a consumption plan c fi nanced by an admissible trading strategy (a, 0) with the following properties: c is nonnegative and there at least exists a t and an event at E Jt such that c ( at. t ) is strictly positive, and a(O) B (O) + O (o) T (S (O) + X(O)) � 0, where c ( at , t) denotes consumption at time t in event at . That is, an arbitrage opportunity is a consumption plan that is nonnegative always and strictly positive in at least one event, and has a nonpositive initial cost. A price system that admits arbitrage opportunities can never be an equilibrium price system, because nonsatiated individuals will take unbounded positions in an arbitrage opportunity so that markets cannot clear. Here we remark that the definition of an arbitrage opportunity does not involve any of the individuals' possibly diverse probability beliefs 1r i 's. We will show in Sections 8 .4 and 8.5 that when no arbitrage opportunities are present , a longlived security price process and its accumulated dividends, in units of the 0th longlived security, have the property that at any time t, the conditional expectation of their sum at any future date must be equal to the sum of their values at time t, where the expectation is taken under a (pseudo) probability 1r• = {1r� ; w E 0 } not necessarily equal to any of the individuals' endowed beliefs 1r i 's. Equivalently, the price of a long lived security plus its accumulated dividends, in units of the 0th
227
Multiperiod Securities Markets II: Valuation by Arbitrage
longlived security, must form a martinga le under a probability ,.. • . Like all of the ,.. i •s, this probability ,.. • also assigns a strictly positive probability to every state w E 0. Since the absence of arbitrage opportunities is a necessary condition for an economic equilibrium, it then follows that every equilibrium price system must have this martinga le p roperty. It turns out that this martingale property is also sufficient for a price system not to admit any arbitrage opportunities. Before we leave this section , we will give a formal definition of a martingale. A process Y = {Y(t) ; t = 0, 1 , . . . , T} is a martingale adapted to F under a probability 1r if E[Y ( s) I Jt ] = Y (t)
where E[· I Jt] is the expectation under martingales will be adapted to F .
Vs � t ,
1r
conditional on Jt . All the
8 .4. We will prove first that the no arbitrage condition implies the martingale property. Suppose that (B, S ) admits no arbitrage opportunities. Note that an individual i in this economy solves the following problem:
T
max Ei [L Ui t (c(t))] (a,ll) t =O s.t. c is financed by and
( a (O) , O( O )) =
( a , 0) (ai (O) , Oi (O ) ) ,
( 8 .4 . 1 )
where Ei [·J denotes the expectation under ,.. i , As there are only a fi nite number of timeevents, there are no arbitrage opportunities, and individuals only choose nonnegative consumption plans, mathemat ical arguments show that there always exists a solution to (8 . 4. 1 ) , denoted b y ( ai , (Ji) . Let ci be the consumption plan financed by ( a i , O i ) . The reader is asked in Exercise 8 . 1 to show that , Vt < T, S; (t) = Ei
[u�
t + l ( ci ( t + 1)) ( S; (t + ' u� t (ci (t))
As u � t (ci (t)) is known at time u�t ( ci ( t)) S; (t)
=
[
t,
1 ) + x; ( t + 1 ) ) 1 Jt
l
.
( 8.4. 2 )
relation ( 8 .4.2) is equivalent to
]
Ei utt + l (c i (t + 1 ) ) ( S; (t + 1 ) + x; (t + 1 ) ) 1 Jt . (8 . 4.3 )
228
Foundations for Financial Economics
Relation (8 .4.3) can be understood as follows. The lefthand side of (8.4.3) is the marginal utility at time t of one fewer share of security j, and the righthand side is the marginal utility at time t of one additional share of security j. In the optimal solution, there should be no incentive to deviate. Thus it is necessary that the lefthand side of (8.4.3 ) be equal to the righthand side. The reader should compare (8.4.2 ) with (5.6.2) and (7.14.5 ) . The discount bond prices satisfy a relation similar to (8.4.2): B (t)
=
{ {
[ [
 u tet l ( c i ( t +l ) ) B (t + E1 u1.II (c'(t)) . u � .e±l (c' ( t +l )) I T EI· u � l ( c • ( t )) Jt
]
1)1 J; ] t
if t :::; T  2 ·, 1'
ft=T  1
•
By repeated substitution of the above relation into itself, we have B (t )
=
[ u'. (ci(s)) Ei u�: (c � ( t ) ) B (s ) l ft [ u(,( c'(s)) Ei u�1 (c' (t)) I Jt
]
]
if t :::; s :::; T  1;
if s
=
T.
(8.4.4)
Price processes of longlived securities are in general not mar tingales under any of the 1r i 's, as is evident from (8.4.2) and (8.4.4) . One sufficient condition for price processes plus accumulated divi dends to be martingales under some individual's probability belief is that there exists at least one risk neutral individual who does not exhibit any time preference. In such an event , ( 8.4.2) must apply for this risk neutral individual under his probability belief, say 11" . Since he is risk neutral and does not have any time preference,
Thus (8.4.2) and (8 .4.4) become, 'Vt
0, and the assumption that 71"� > 0, Tlw . In fact , 1r: > 0 , Tlw. Next note that
[u�:
]
(ci (T)) �� (ci (O) ) / B (O) = B(O)j B(O) = 1 ,
= E;
where the third equality follows from (8 .4.4) . Now we want t o show that Sj + D j i s a martingale under 1r • . First we observe that the conditional probability under 1r* of a n event a. E 1. given an event at E Jt with s � denoted by 1r�. (at ) , is
t,
( 8 . 5 . 2) Similarly, we can define 7r�. (at ) , the probability under ditional on at . When a. � at , we have e 1r U �T (ci (w , T)) E 1r•a (a t ) = wE a . 1f� = E w a. � Ewea , 7r�u:T (c'(w , T)) E wEa1 11"� . 'II"� U�x(ci(w, ) ) "" I 7ria. U;s ( Ci ( a . ' s ) ) L.Jw Ea. 'II" ' u' (c,.(a T•)) . 4.
..
1r i of a.
con
(8 .5.3)
,
where ci (at , t) denotes individual i's optimal consumption at time t in event at , and 71"�1 is the probability of at under 7r i . Relation (8 .4.4) implies that
" 0 wE at
7r� U�T (ci (w , T) ) _  B (at , t ), ( 7r ia , u1i t ( ci at , t ) )
Tit � T  1 ,
(8.5 . 4)
231
Multiperiod Securities Markets II: Valuation by Arbitrage
where B (a t , t) denotes the bond price at time t i n event a t . Substi tuting (8 .5.4) into (8.5.3) gives if s � T 
1;
(8 .5 .5)
if s = T .
We are now ready to prove the first main result of this chapter. Substituting (8.5 . 5) into (8. 4.2) gives, for all t � T  1,
S; (t) = Ei
[
u � t+ 1 (ci (t + 1 )) (S; ( t + 1 ) + x;( t + 1 )) 1 Jt u�t (c i (t)) •
= B ( t ) E* [Sj ( t + 1 ) + xj (t + 1) 1 Jt ] ,
l
where E * [ · ! Jt] is the expectation operator with respect to tional on Jt . Rearranging (8 .5.6) gives, for all t � T  1 ,
Sj ( t) = E* [Sj (t + 1) + xj (t + 1 ) ! Jt ] .
(8.5.6)
1r*
condi
( 8 . 5 . 7)
The reader should compare (8 .5.7) with (8 .4.5) . These two relations have the same form, except that one is in units of the consumption good and the expectation is taken with respect to 11" and one is in units of the discount bond and the expectation is taken with respect to 1r * . Following the same line of arguments used in deriving (8 .4 . 7) from (8 .4.5) , we easily get
Sj (t) + Dj (t ) = E* [ S * ( s ) + Dj ( s ) ! Jt] V s = E* [Dj (T) ! Jt] .
2:
t,
( 8 .5 . 8)
That is, Sj + D j is a martingale under 1r • . As for the discount bond, B* ( t ) + D� (t) = 1, Vt , and is certainly a martingale under 1r * . In summary, we have proved that if a price system admits no arbitrage opportunities, then, after a normalization , there must exist a probability 1r • that assigns a strictly positive probability to every state of nature such that price processes plus the accumulated div idends are martingales under 1r* . A probability having the above property is termed an equivalent martingale measure .
232
Founda tions for Fin ancial Econ omics
Before we leave this section, we make one remark. We con structed an equivalent martingale measure by using an individual's probability belief ?ri and his marginal utilities. Under the martingale measure, however, the normalized price processes plus normalized accumulated dividends form martingales for all individuals. 8.6. We now p roceed to prove the converse part of the_ mar tingale property: if B ( t) > 0 for all t < T and if there exists an equivalent martingale measure, 1r * , for SJ + D j , then (B, S) admits no arbitrage opportunities. We will prove the assertion by contraposition. Let there be an arbitrage opportunity c financed by an admissible trading strategy (a , 0 ) . That is, c � 0, c # O , and a(O) B (O) + 0 ( o) T (S(O) + X (O)) 5 0. Recall from Exercise 7. 1 that if c is financed by (a , O) , we have
a (T) + O ( T ) T X ( T) = a(O)B(O) + O (O)T (S (O) + X(O) ) T 1
T
 L c(s) + L a (s) (B(s) s=O T
a=l
B( s 
1 ) + xo ( s ) )
(8 . 6 . 1)
+ L O (s) T (S (s)  S(s  1) + X ( s ) ) . s= l
Relation (8.6.1) follows from the selffinancing budget constraint. The lefthand side of (8.6. 1) is the value of the strategy (a , 0 ) at time T. The righth and side of (8.6.1) is the initial value of the strategy plus accumul ated gains or losses from trading in the l onglived se curities from time 0 to time T minus the accumulated consumption withdrawals from time 0 to time T  1 . S imilar arguments show that in discounted units,
a(T) + O(T ) T X* ( T ) = a (O) + O(O)T (S* (0) + x•( o) ) T
+ L O (s)T (s• (s)  s• (s  1 ) + x• (s)) s=l
where
X* (t) = (xi (t) , . . . , x !,r( t)f and c(t) (t) if t 5 T c• (t) = c ( t ) jB if t = T. 
{
T l
L c* (s), s=O
1;
(8 .6.2 )
Multiperiod Securities Markets ll: Valua tion by Arbitrage
Denoting (D i (t ) , . . . , D ,N (t )) T by D * (t) , we have
233
(3.6.3)
Substituting (8. 6 .3 ) into (8. 6.2 ) , noting that
c • (T ) = a( T ) + O ( T )T X* ( T ) , and rearranging gives T
l: c* (t) = a (O) + O (o )T (S" (O) + X* (O)) t=O
+
T
{ 8 .6.4)
L: O(t)T (S* (t) + D* (t)  S " (t  1)  D" (t  1 ) ) . t=l
Recall that O; (t) is predictable and its value is thus known time t  1 . It follows that
E* [O(t) T (S* (t) + D * (t)  S*(t  1 )  D * (t  1)) 11it] = O(t)T E* [S* (t) + D* (t)  s • (t  1)  D* (t  1 )1.1tt] = O(t) T (E• [s• (t) + D *(t) l .1tt]  S * (t  1)  D * (t  1))
at
(8 .6.5)
= 0,
where the third equality follows from the hypothesis that s• + D * is a martingale under :�r • . Now we the take expectation of (8. 6 .4) under :�r • and use the law of iterative expectations to obtain
E" +
E"
[� ] [t, [t. c ' (t)
�
a(O) H(O)T (S'(O) + X' ( O))
D ( t ) T (S' (t) + D' (t)  S ' (t  1)  D ' (t 
= a (O) + O (O) T(S* (O) + X* (O) )
+ E'
E" JD(t) T ( S ' (t) + D' (t)  S' (t
= a(O) + O (o ) T (S * (O) + X* (o) ) ,

1))]
1)  D' (t  1
))J T.t]] ( 8.6 . 6)
234
Foundations for Fin ancial Economics t=I
I =0
t =2
W5
F igure
8 . 7.1:
(�)
An Example of a Securities Market
where the third equality follows from (8.6.5 ) . Note the following. If c is an arbitrage opportunity, it must b e nonnegative and be strictly positive i n at least one event at E 'ft . B y the assumption that B ( t) > 0 , c • ( at , t ) > 0. The probability 11' • is an equivalent martingale measure and thus assigns a strictly positive probability to at. It follows that the lefthand side of (8 .6.6 ) must be strictly positive. On the other hand, because c is financed by (a , 9) and is an arbitrage opportunity,
As B ( O)
>
a (O ) B ( O) + O (O) T ( S (O) + X (O) ) � 0. 0, this implies that
a( O) + O(O) T ( S * (O ) + X* (O) ) � 0, which is a contradiction to (8 .6.6 ) . Therefore, there are no arbitrage opportunities. We have thus shown that the existence of an equiv alent martingale measure is a necessary and sufficient condition for ( B, S ) not to admit any arbitrage opportunities.
Multiperiod Securities Markets II: Valuation by Arbitrage
235
8 . 7 . As a sufficient condition, the martingale property makes it convenient to make sure that a given price system admits no arbitrage opportunities. For example, the prices and dividends for the three longlived securities considered in Section 7 .12 , shown here again in Figure 8.7.1 , admit no arbitrage opportunities, b ecause there exists an equivalent martingale measure. ( Recall that these three securities do not pay dividends until time and the numbers shown at that time are the dividends paid. Also, every branch in the event tree is of a strictly positive probability.) To see this, we will construct an equivalent martingale measure explicitly. We do this recursively by dynamic programming. A t time at the upper node, let P l , P2 and Ps b e the conditional probabilities for WI J w 2 , and ws , respectively. As the price process plus accumulated dividends for the Dth security is unity throughout, t.he discounted price system is equivalent to the original price system. If there exists a martingale measure, then Pl, p 2 , and Ps must satisfy the following linear equations:
3
1
+ P2 + Ps = 1 , pt + 4p2 + Bps = 5 , 2pt + 3p2 + 4ps =
3
P1
3. ( ) = ( 1/1/33) 1/3
( 8.7.1)
The unique solution t o ( 8.7.1) is Pl
P2
Ps
and is written along the three top branches in Figure 8 .7 . 1 . Repeat ing the above procedure, we solve for all the conditional probabilities and write them along the branches. The reader can verify that those conditional probabilities are the unique set that makes prices plus accumulated dividends a vector of martingales. Note that the con ditional probabilities are all strictly positive. Given the conditional probabilities written along the branches, it is then straightforward to compute the martingale measure 11" "' , which is simply the unconditional probabilities implied by the conditional
236
t.0
Foundations for Financial Economics t•l
t =2
WI
w2 w3 W4
Figure
8. 7 . 2 :
w5
0) 0) (�) (� ) G)
An Arbitrage Opportunity
probabilities:
(8 .7.2) which is equivalent to 7r. Thus there are no arbitrage opportunities. Conversely, the failure of the martingale property at any node implies that an arbitrage opportunity can be constructed there. To see this, we change the prices at time 1 at the upper node to be those written on the event tree in Figure 8.7.2. The price at that node for security 2 is changed to 3. It is easily seen that there does not exist an equivalent martingale measure. The unique conditional probability that makes the prices plus accumulated dividends for security 2 a martingale assigns probability one to state w1 and probability zero to states w2 and ws . Any unconditional probability implied by this conditional probability cannot assign a strictly positive probability to every branch. Moreover, this conditional probability does not make the martingale property hold for security 3. A casual observation immediately reveals why an equivalent martingale measure does not
Multiperiod Securities Markets
II:
Valuation by Arbitrage
237
exist and where the arbitrage opportunity is. The price for the second security at time t = 1 at the upper node is equal to 3, while its dividends in the three possible states at time 2 are all greater than 3. Indeed, the dividends in state w2 and w3 are strictly greater than 3. At this node, the second longlived security dominates the riskless security and the third longlived security in that its total returns in all states following the upper node are greater than those of the other two securities and are strictly greater in at least one state. An arbitrage opportunity is, for example, short selling secur ity i an d buying in security 2 . The reader will be asked in Ex erCise 8 . 2 to supply the details in constructing an arbitrage opportunity. 8.8.
Besides pinpointing the existence of arbitrage opportuni ties, the martingale property, as a necessary condition, has impor tant implications. It allows us to compute the prices of a complex lon glived security over time in a simple way, when the price of the complex security is welldefined over time. Recall that a longlived security is characterized by its payoffs in each timeevent. Thus, a longlived security is equivalent to a consumption plan and we will use these two terms interchangeably. We will show in this section that when there are no arbitrage opportunities, it is precisely those marketed consumption plans / longlived securities that have well defined prices over time. A derivative security is just a consumption plan / longlived security, so it has welldefined prices over time when it is marketed and when no arbitrage opportunities exist. In such an event, we say that a. derivative security is p riced by arbitrage . We now show that a consumption plan has well defined prices over time when it is marketed and there are no arbitrage opportuni ties. First, we show that a marketed consumption plan has a unique cumdividend price at time 0. Let c be marketed and financed by (a, O ) . From ( 8.2 .1 ) , we see that an initial cost for c, by dynamic trading , is
a (O) B (O) + O(O) T ( S ( O) + X(O) ) . This is a. cumdividend price for c at time 0. We claim that when there are no arbitrage opportunities, this price is unique. Suppose this is not the case. Then there must exist another admissible trading
238
Foundations for Financial Economics
strategy ( &, B) that finances c with a different initial cost. Without loss of generality, assume that
a (O) B (O) + 0 ( 0 ) T (S (O) + X(O) )
>
&(O) B (O) + 0(0 ) T ( S ( O) + X (O)) .
I t can b e verified that (&  a , 0  0 ) i s an admissible trading strat egy that finances a consumption plan which is zero throughout and has a strictly negative initial cost. It is then easy to turn this into an arbitrage opportunity and a contradiction. The reader will b e asked t o provide a construction of a n arbitrage opportunity i n Exer cise 8 .4. Therefore, the cumdividend price at time 0 for any mar keted consumption plan is unique. This conclusion also extends to the exdividend price at time 0 by subtraction of the time 0 divi dend / consumption from the cumdividend price. Next, we want to show that any marketed consumption plan has a unique price at any time t. We will work with the exdividend prices. An exdividend price for a marketed consumption plan at any time t is the number of units of timet consumption good needed to begin a dynamic strategy that replicates the consumption plan after time t. From ( 8.2.1) , we know that if c is financed by ( a , O ) then a price for { c ( s ) ; s = t + 1 , , T} at time t is .
.
.
a(t + 1 ) B (t) + O (t + 1 ) T S (t) .
The fact that this cost is unique when there are no arbitrage op portunities follows from the same arguments as above. We can thus define unambiguously the exdividend price of the consumption plan c at time t as
Sc (t)
:=
a(t + 1 )B(t) + O(t + 1 ) T S (t) .
(8.8.1)
Similarly, we define the discounted prices over time for c as follows
s; (t)
{ 0�hW
if t � T  1 if t = T = a(t + 1) + O (t + 1) T s • (t) , = a(t) + O( t ) (s • (t) + x• (t) )  c * (t) , =
(8.8.2)
where the last equality follows from (8.2.1) . Let 11'" 0 be an equiva lent martingale measure, which we know exists because (B , S) does
239
Multiperiod Securities Markets II: Valuation by Arbitrage
not admit arbitrage opportunities by assumption. We claim that s: (t) + E ! =o c* (s) is a martingale under the equivalent martingale measure 1r* . This claim follows from arguments similar to those used in deriving (8 .6.6) . To see this, we know from Exercise 7 . 1 and from (8.8.2) that
a (t + 1) + O(t + 1) T S * (t ) = a(O) + O(O) T (S* (O) + X* (O) ) t
t
+ I: O (s) T (S * (s)  S * (s  1) + X* (s))  �::: c • (s) . •=0 s = 1
Substituting (8 .8.3) into (8 .6.4) gives T
L c* (s)
s= t + l T
=
a(t + 1) + O (t + 1)T S * (t)
+ L O(s) T (S* (s) + D'" (s)  S * (s  1)  D * (s  1 )).
(8 . 8 .3)
(8 .8 .4)
s= t + l
Taking expectations of both sides of (8 .8.4) with respect to 11" 0 con ditional on Jt and using the law of iterative expectations gives T
E * [ L c * (s) I Jtl s=t+l
+ E*
[t
s=t+l
=
a(t + 1) + O(t + 1 ) T S * (t)
E * [ O(s) T (S* (s) + D* (s)  S* (s  1)  n• (s  1 ) ) 1 1st ] I Jt
= a( t + 1) + O( t + 1)T s * (t ) = s: (t) .
l
(8 . 8 . 5) That is, the discounted exdividend price for the consumption plan c at time t is the conditional expectation under 1r • of the sum of discounted future consumption. Note that adding E != 0 c * (s) to both sides of (8 . 8 .5) immediately yields (8 .8.6)
240
Foundations for Fin ancial Economics
which is equivalent to saying that s; (t) + I:! =o c* ( s ) is a martingale under 1r* ; because by the l aw of iterative expectations, Vs ?:?: t, s
E* [s; (s) + 2: c * ( v ) l 1t]
[ [t, ] ] [ l o ' ( v ) I T.
= E' E' = E'
= s; ( t ) +
o ' ( u ) I T.
I r.
(8. 8 . 7)
t
2:.: c• (s) . s=O
We have thus shown that if ( B, S) admits no arbitrage oppor tunities, not only the prices of the longlived securities but also the prices of marketed consumption plans have the martingale property. G iven an equivalent martingale measure, the prices over time of any marketed consumption plan can be computed by evaluating a condi tional expectation under this martingale measure. Although individ uals' probability beliefs about the states of nature may differ, they nevertheless agree on the price processes of marketed consumption plans. This follows since the price processes of marketed consump tion plans are determined by the no arbitrage condition and since the definition of an arbitrage opportunity does not involve any of the individuals' probabilities 1r i 's. 8.9.
Consider the price system for the three longlived securi ties represented in Figure 8.7 . 1 . The unique equivalent martingale measure is defined in (8. 7 .2) , and the implied conditional probabil ities are recorded along the branches in Figure 8.7 . 1 . The three trading / consumption dates are t = 0, 1 , 2. The first security is a dis count bond. As its price is unity at time 0 and time 1 , the interest rate is zero throughout. Recall from Sections 7. 1 2 and 7. 1 3 t h a t the securities market with the three longlived securities is dynamically complete. Therefore , any consumption plan can be manufactured dynamically, or is marketed, and has welldefined prices over time.
241
Multiperiod Securities Markets II: Valuation by Arbitrage t
=
t
0
=
I
t
=
2
0
0. 2 5 0
4.25 5
3.5
Figure 8.9.1: An Example of a Consumption Plan
Now consider the consumption plan c depicted in Figure 8 .9 . 1 . I n this consumption plan, individuals consume only a t time 2 . The prices for this consumption plan before time 2 can be determined by finding a trading strategy that replicates it through dynamic trad ing as we did in Section 7.12. Alternatively, because an equivalent martingale measure has been computed, we can use (8.8.6) to com pute the prices in a straightforward fashion. Note first that because B ( t) = 1, Vt � T  1, the discounted price system is equivalent to the undiscounted one. Thus
Sc (O) =
(�) (�) O+
0.2 5 +
( �)
4.25 +
(�)
1 .5 +
= 2,
sc (w n ,
1) 

{ (�)
0 + (k) .5 t1 + t
3 .5
+ .5 2
0.25 =
(k) 4. 2 5 =
(�) 3 .5
1 . 5 i f n = 1, 2 , 3 , if n 4 , 5 . (8 . 9 . 1) =
242
Foundations for Financial Economics
In Exercise 8 . 5 , the reader will be asked to verify the prices computed in (8.9.1) by constructing a replicating strategy. 8 . 10.
Before we discuss an application of the martingale prop erty in pricing stock options, we digress slightly on some aspects of the theory developed here as they relate to Chapter 7. In Chapter 7 , we discussed dynamically complete markets. A necessary and sufficient condition for markets to be dynamically com plete is that at every node on the event tree representing the infor mation structure, the number of longlived securities having linearly independent random returns the next period is equal to the number of branches leaving the node. We will show in this section that, given a price system that admits no arbitrage opportunities, markets are dynamically complete if and only if there exists a unique equivalent martingale measure. We will see that this conclusion follows easily from elementary algebra. We fix price system ( B, S) that admits no arbitrage opportuni ties, so that there exists an equivalent martingale measure :�r * . Let a t  1 E 1f1 for some t � T and let {a81 , . . . , a8 ,. } be such that a8l: E 1t, as., � a t 1 for k = 1, 2 , . . . , m, and U:'= 1 asl: = a t  1 · Note that m is the number of branches leaving the node a t 1 · Consider the following system of linear equations:
1 s; (a, l 't)+Di (a,1 ,t) s; ( a,1 ,t)+D 2 ( a,1 ,t)
=
1 s; ( a,,. ,t)+Di ( a,,. ,t) S2 (a.,. ,t)+D2 (a,.,. , t)
s;., (a,1 ,t)+D; (a,1 ,t) 1 s;(at l >f 1 )+Di (a1 1 ,t 1 ) s; (at  I 1t 1 )+D2( a cI ot 1 ) s;., (a cI ,t 1)+D; (a cI ,t  1 ) .
(8 . 1 0 . 1 ) The unknowns in this system are p = {Pa . l: , k = 1 , 2 , . . . , m } . A strictly positive solution to the system is a conditional probability for {a. l: ; k = 1, . . . , m } given the event a t  1 , because the first linear
Multiperiod Securities Markets II: Valuation by Arbitrage
243
equation in the system requires that the solutions, p� s, sum to . ,. one. The other linear equations will simply turn S* + D * into a martingale from the node at _ 1 to its subsequent nodes under the conditional probability {Pa . ,. ; k = 1, . . . , m } . Note that we have m unk nowns and N + 1 linear equations in the system, where m is the number of branches leaving node Ut  1 . By the hypothesis that there exists an equivalent martingale measure :�r • , there exists a strictly positive solution to (8 . 1 0 . 1 ). This strictly positive solution is the conditional probability induced by w* condi tional on at 1 · If the matrix on the lefthand side of ( 8 . 10 .1) has m linearly independent rows, the solution induced by :�r * is unique. We can carry out the above analysis for each node in the event tree representing the information structure F and conclude that if there are as many linearly independent longlived securities at each node as the number of branches leaving the n o de , that is, if markets are dynamically complete, there exists a unique equivalent martin gale measure. Conversely, suppose that the conditional probability induced by • 1r on a t  1 is the unique strictly positive solution to (8. 1 0. 1 ) . We claim that the matrix on the lefthand side of ( 8 . 10. 1) must h ave m linearly independent rows. Here we note that m is the maximum number of linearly independent rows the matrix can have, because it has m columns. Suppose this is not the case, that is, the number of linearly independent rows is strictly less than m. By the hyp othesis that ( B, S ) admits no arbitrage opportunities, we know that there exists a strictly positive solution, p, to ( 8 . 10. 1 ) . From linear algebra, we then know that there exists a continuum of solutions to ( 8 . 1 0 .1) but not all solutions to (8.10.1) will be strictly positive. Now take any solution p1 = {p� • I , . . . , p'a•m. } As p is strictly positive, there exists A E [0, 1] such that p = AP + ( 1  .X)p' is strictly positive. It is easily checked that p is also a solution to (8 . 10 . 1 ) . Thus 1r * is not the unique equivalent martingale measure, a contradiction. It then follows that at every node, the number of longlived securities that have linearly independent returns in the subsequent nodes must be equal to the number of branches leaving that node. This implies that the markets are dynamically complete. We have thus shown that the uniqueness of an equivalent martingale measure is equivalent to •
244
Foundations for Financial Econ omics
dynamic market completeness. Before we end this section, we make one final remark . The equiv alent martingale measure of (8.5.1) is constructed by using an indi vidual's ratios of marginal utilities weighted by his probability belief 1ri . When markets are dynamically complete, those ratios weighted by 1ri 's are equal across individuals, and thus the equivalent martin gale measure is unique. 8.11. Recall from Chapter 7 that for every equilibrium in a competitive economy with a complete set of timeevent contingent claims where ma.J:kets open only at time 0, there exists a rational ex pectations equilibrium with reconvening securities markets that has the same allocation, when a set of longlived securities is appropri ately chosen . With the aid of the discussion in Section 8 .9, we can prove a very general converse of the above statement. We will show that for every price system ( B, S) admitting no arbitrage opportu nities, there exists a competitive economy where markets open only at time 0 such that every individual optimally trades, in this static setting, to his allocation in the rational expectations longlived se curities markets economy. The price system ( B, S) does not h ave to be an equilibrium price system , and the markets are not necessar ily complete. Moreover , any solution to an individual's maximization problem in the static economy is also a solution in the dynamic econ omy. Formally, let ( B, S) be a price system that does not admit arbi trage opportunities , and let .M be the space of marketed consumption plans. From ( 8 . 8 .6) we know that
S� (O) + c•(o) � E'
[tc•(s) 1 Vc
E N,
(8. 1 1 . 1 )
where t h e expectation is taken under an equivalent martingale mea sure. Take any individual in the economy, say individual i. Let c1 be his optimal consumption plan in the dynamic economy. Consider
Multiperiod Securities Markets
II:
Valuation by Arbitrage
245
the following static problem: c:E .M
max E; s. t. E'
[t, c'(t)]
=
[t t=O
Ui t ( c ( t) )
l
(8.11 .2)
a' (o) + ii' (O JT ( S' { O) + X' { O)) .
This static problem can be understood as follows. At time 0, individ ual i can choose a consumption plan from those that are market ed . The price of a marketed consumption plan is exactly equal to its initial cost in a dynamic longlived securities markets economy. We claim that ci is also a solution to (8 . I I .2) . To see this, we note that because ci is a solution in the dynamic economy, it must be marketed. Relation (S.B.6) implies that (8 . 1 1 .3) Thus, ci is budget feasible for (8 . 1 1 .2) . Suppose that c1 is not a. solution to (8. 1 1 .2). Then, there must exist a c E .M satisfying the budget constraint of (8.II .2) such that
( 8 . 1 1.4 ) As c E .M and has an initial cost equal to ci (o) B (O) + 01 (o)T (S(O) + X(O) ) , it must also be feasible for individual i in the dyn amic econ omy. This contradicts the fact that c' is a solution in the dynamic economy. Conversely, let c' be a solution to (8.1 1 .2) . We want to show that it is also a solution in the dynamic economy. First, c1 is feasible in the dynamic economy because it is marketed and has an initial cost equal to ai(o) B (O) + U' (o)T (s(o) + X (O) ) . Suppose that ci is not a solution in the dynamic economy. Then there must exist a consumption plan c E .M financed by (a, il) with ( a (o) , O (O) ) = (a'(o) , Ui (o)) such that ( 8 . 1 1 .4 ) is true. Relation ( 8 . 8 .6 ) implies that c satisfies ( 8 . 1 1 . 3 ) and
246
Foundations for Financial Economics
is thus feasible for (8.11.2). This contradicts the hypothesis that ci is a solution to (8.1 1 .2) . Finally we note that, using the same arguments as in Section 7 .8, w e can show that in the dynamic economy, an individual has no incentive to deviate from the consumption plan that is optimal in the static economy. The correspondence between a dynamic economy and a static economy makes some analyses in a dynamic economy easier. We will give an example here. In the remarks at the end of this chapter, the reader can find references for other applications . The power of this correspondence is most pronounced in a continuous time continuous state model. We consider the consumptionportfolio problem for an individ ual i posed in (8.4.1) . Suppose that (B, S) admits a unique equivalent martingale measure 1r * . From Section 8. 10, we know that markets are dynamically complete. Thus the space of marketed consumption plans .M is composed of all the possible consumption plans. Then ( 8 . 1 1 .2) becomes 8.12.
rn:x El
•.t. E'
[t, l ,•
( t)
�
[t, ) ] u., ( < ( t
;(0 ) + 0' (0) T (
S'
(D) + X' ( 0 ) ) ,
(8.12.1)
where E* H is the expectation under 1r * . Standard mathematical ar guments show that there always exists a solution to (8.12.1), because there is a finite number of states of nature and u�t (O) = oo. Denote this solution by c* . It then follows from Section 8 . 1 1 that there exists a solution to the dynamic problem. Next we want to characterize the optimal consumption and port folio policies. We can rewrite the budget constraint in (8.1 2.1) as
Multiperiod Securities Markets II: Valuation by Arbitrage follows:
E*
=
[��
I:
c(t) B (t) + D0 (t)
1r i '11"� w 1r�
l
c ( w, t ) t t =O B (w, t) + Do(w , t)
[; • � � ] [� [ ** ]]  [� l w en
= Ei
= .&  E,
i
T
T
T
247
B ( t)
Ei
ii
(t) D 0 ( t)
(8 . 12 . 2)
c ( t) B (t) + Do (t) i 1t
c (t) 'l (t) B (t + ) D0 (t) '
where if • /i' denotes the random ratio of 1r':., j 1r� , and
Forming the Lagrangian yields
where 'li is a strictly positive Lagrangian multiplier. The first or der conditions, which are necessary and sufficient for ci to be an optimum, are
'l (t) ui' t (,ci (t )) = ii B (t) + Do( t)
v t,
(8.12.3)
and the budget constraint. Relation (8 . 12.3) can be written more explicitly as
(8 .12 . 4)
Foundations for Financial Economics
2413
The reader is asked in Exercise 8 .6
to
verify that
B ( at , t) + Do ( at , t) is the time0 price of a timeevent contingent claim for the time event (t, at ) · Then (8.1 2.3) corresponds to (7.3.7) . Relation (8.12.3) completely characterizes the optimal consump tion policy. The one remaining task is to compute the optimal port folio strategy. As we know the optimal consumption plan already, this is then a standard practice of finding a replicating strategy as demonstrated in Section 7.12. Utilizing the correspondence between a static economy and a dynamic economy allows us to decompose an individual's optimal consumption and portfolio problem into two parts. First, find an optimal consumption plan in a static economy through the standard Lagrangian method. Second, implement this optimal consumption plan in the dynamic economy through a replicating strategy. In the rest of this chapter, we will use the martingale property developed in the earlier parts to discuss a discrete time option pricing theory due to Cox, Ross, and Rubinstein (1979) . Consider a multiperiod securities market with two longlived securities, a risky common stock and a riskless bond. The economy has a long time horizon. However, we only look at a piece of it, say the trading dates t = 0, 1, . . . , T. The risky asset does not p ay dividends from time 0 to time T and has prices over time following the binomial random walk depicted in Figure 8 . 1 3 . 1 . At time 0, the stock price is S (O) > 0. At time 1 , the stock price will be either uS ( O) or dS(O), with u > d, and so on and so forth . That is, at every trading date, the return on the common stock next period will either be u or d. We will assume that u > 1 and d < 1 and interpret the stock price movement from S(t) to uS(t) to be moving up and that to dS ( t ) to be moving down. Note that if S(O) > 0, the stock price will never reach zero. The riskless asset, the bond, does not pay dividends and earns a constant return R. Equivalently, let the bond price at time t be Rt . 8 . 13.
Multiperiod Securities Markets II: Valuation by Arbitrage 1=0
I= I
1 = 2
249
I =3
S ( O) '
'
'
/
Figure
8.13.1:
' /
'
/
'
/
'
A Binomial Random Walk
We assume that the information structure an individual has is the information generated by the stock price. By this we mean that a state of nature is a complete realization of the risky stock price from time 0 to time T. For exampl e, when T = 3, a possible complete realization of the stock price is
S (O) , S ( 1 ) = uS(O) , S ( 2) = udS (O) , S( 3) = u2dS (O) . In this realization, the stock price moves up at time 1 , moves down at time 2, and moves up again at time 3. Thus the state space 0 is composed of all the p ossible complete realizations of the stock price from time 0 to time T. The event tree representin g the information structure is just that depicted in Figure 8 . 1 3 . 1 . At time 0, an indi vidual knows that the state of n ature is one complete realization of the stock price from time 0 to time T . Thus 1o = {0}. At time 1 , the partition .1i has two events. The first (second) event i s composed of all the realizations of stock prices from 0 to T such that the stock moves up (down) at time 1. The other Jt 's can be understood sim ilarly. Note that, for every node on the event tree of Figure 8.1 3.1 b efore time T, there are only two branches leaving the node. More formally, for every at E .1t with t < T , there are two elements of h+l
250
Foun dations for Financial Economics
that are subsets of a t . Note also that the information an individual has at time t is simply the historical price realizations of the common stock. The riskless asset does not generate any information , because its return is purely deterministic over time. Finally, we assume that individuals ' possibly heterogeneous prob ability beliefs assign strictly positive probabilities to all states. 8 . 14.
Now we want to make sure that the prices presented in Section 8 . 1 3 do not admit arbitrage opp ortunities. We recall from Sections 8 . 5 and 8.6 that a necessary and sufficient condition for _no arbitrage opportunities is that there exists an equivalent martingale measure. As usual, we define
S* (t) B* (t)
=:
=:
S (t) R t Rt R t = 1 .
We want to find a probability that assigns to every state o f nature a strictly positive probability such that
(8.14.1) As S * (t ) is strictly positive, (8. 14.1) is equivalent to
E
•
[
]
S * (t + 1) =1 S * (t) I Jt
Vt < T.
( 8 .14.2 )
That is, under the martingale measure, the risky stock is expected to earn a rate of return, in number of the b ond , equal to zero. Given the price dynamics, there are two branches leaving each node b efore time T; and , on the two subsequent nodes, S * (t + 1)/ S* (t) has values uR 1 and dR 1 , which we note are two constants independent of time t and of the node at time t. For (8. 1 4.2) to hold, the conditional probabilities at any node on the event tree at time t < T for the subsequent two nodes, induced by the equivalent martingale measure, denoted by 1r and 1 1r must solve the following linear equation : 
(8.14.3 )
Multiperiod Securities Markets II: Valuation by Arbitrage '=0
I=3
1 =2
I=I
25 1
.,.
S (Ol
d 3 S (Ol
F igure 8 . 1 4 . 1 : The Equivalent Martingale Measure
The unique solution to (8.14.3) is Rd . ud
'lr = 
(8. 14.4)
As 1r and 1  'lf are conditional probabilities induced by an equivalent martingale measure, 1f must lie in (0, 1) . The necessary and sufficient condition for 1f to be in (0, 1) is d < R < u.
(8. 14.5)
Let us assume henceforth that (8.14.5 ) is true. Then there exists an equivalent martingale measure. Moreover, since 1f is the unique solution to ( 8.14.3 ) , the equivalent martingale measure is unique. The conditional probability induced by the equivalent martingale measure at each node for time t < T assigns probability 1r to the stock's subsequent upward move and probability 1  1f to the stock's subsequent downward move; see Figure 8 . 1 4 .1 . It is easily seen that the probability, under the martingale measure, for a state of nature is completely determined by the number of upward moves of the stock price. For example, the probability, under the martingale measure,
Foundations for Financial Economics
252
for a complete realization of the stock prices that has n upward moves is :�r"(l  'I') T n . Two complete realizations that have the same number of upward moves but differ in the timing of those upward moves are of the same probability under the martingale measure. By simple combinatorial mathematics, there are
(T ) n
T!
= n ! (T 
n) !
number of complete realizations that have exactly n upward moves. In addition, the final stock price for each of these complete realiza tions is equal to S (O) u"cP' n . Similarly, given S (t ) at time t , the con ditional probability, under the equivalent martingale measure, that there will be n ::; T t upward moves in the future and, therefore, the stock price at time T is S (t)u"dT  t n is 
(8. 14.6)
8 . 1 5 . Now we are ready to price an European call option written on the common stock . First, we note that because there exists a unique equivalent martingale measure, the markets are dynamically complete. Thus any consumption plan is marketed and has well defined prices over time or is p riced by arbitrage. Consider a European call option written on the common stock with exercise price k and expiration date T. The payoff at time T of this call option is max [S(T)  k, O] .
From (8.8 .6) and (8.14.6) , we know that the price of this call option at time t is
[
E• max[S (T) =
R (T t )
T t
L =O
n

k, O] R (T t ) ! Jt
( T ;:
t
)
]
"' " (1  '�' ) T t n max[ S ( t ) u " dT t n  k , OJ . 
( 8 . 15 . 1)
2:53
Multiperiod Securities Markets II: Valuation by Arbitrage
This clearly depends only on S (t) , the exercise price k, and the time t. Thus we can write the option price at time t as p(S (t) , t; k) . Let j be the minimum number of upward moves such that
That is, j is the minimum positive integer such that .
> 3 
In
S(t)�'i' I n du
t
( 8 . 15 .2)
Then (8.15 .1 ) becomes
p( S (t) , t; k) =
= S (t)
_
� (T: t) � (T ,; ) (';r((l ��)dr•• � ( T: t)
R(T t)
n =: J
kR(T t)
1r" ( 1
_
1r)T tn (s (t)unf1T tn
_
k)
'
1r n ( 1
n=:J
_
1r ) Ttn
,
>
(8.15 .3 ) T  t, the summation
where we have used the convention that if j is equal to zero. Now note the following: In a series of T  t independent trials of an experiment whose success rate is 1r and whose failure rate is 1 1r , the probability that there will be at least j ?: 0 successes is

( 8 . 15 .4 )

This is termed the complementary binomial distribution fun ction with parameters j, T t, and 1r. The bino mial distrib ution func tion is equal to 1 minus the complementary binomial distribution function, and gives the probability that the number of successes is
254
Foundations for Financial Economics
strictly less than j. The trials are termed Bernoulli trials. Substi tuting ( 8.15.4 ) into ( 8.15.3 ) gives
p(S( t ) , t ; k)
=
S(t) �(j; T  t, 1r 1 )  k R  (T  t ) �(j; T  t, 1r ) , ( 8.15.5 )
where j is defined in ( 8.15.2 ) and where
..
.,.1
u = 'If' R
and
1  1f1
=
!! ( 1 R
7r
).
This is the binomial option pricing formula due to Cox, Ross, and Rubinstein ( 1979 ) . Finally, we note that as the number of independent trials in creases to infinity, the central limit theorem from probability the ory shows that the binomial distribution function converges to the standard normal distribution function after a suitable normalization. Cox, Ross, and Rubinstein make use of this observation and show that if we allow trading to occur more and more frequently while keeping the time interval [ 0, T] fixed, then the binomial option pric ing formula converges to the BlackScholes formula if we choose ap propriate parameters of the price processes in the process of taking limits. The limiting formula is
where
ln( S j k er (T t ) ) 1 r,;;; + a v T t, . � 2 av Tt and the two constants r and (T are the continuously CElmpounded riskless interest rate per unit time and the standard deviation of the rate of return on the common stock per unit time, respectively . To use this formula, we first fix the unit d measure of time. Suppose that t is in units of a month, for example. R i s oue plus the monthly riskless interest rate, define r = In R, and take a to be the standard deviation of the monthly return on the stock. We will not present here the limiting procedure mentioned ab0ve. Cox and Rubinstein ( 1985, pp.196208 ) is an excellent source for this procedure, which we encourage the reader to consult. zt
=


Multiperiod Securities Markets II: Valuation by Arbitrage
255
Exercises
8 . 1 . Prove relation (8.4.2) . 8.2. Given the prices depicted in Figure 8.7.2, demonstrate an arbi trage opportunity. 8.3. Suppose that the 0th longlived security also pays dividends xo (t) > O, Vt . Prove that the martingale property is still true with a suitable normalization. 8.4. Show that if a consumption plan is financed by two admissi ble trading strategies with different initial costs, there exists an arbitrage opportunity. 8.5. Verify the prices computed in (8.9.1) by constructing a dynamic replicating strategy. 8 �6 . Verify that, in the context of Section 8 . 1 2 ,
is the time0 price of a timeevent contingent claim for the time event (t, at) · 8. 7. In the context of the binomial model, price a security that gives the right to its holder to purchase at time T a share of the common stock at the minimum price that the stock has reached from time 0 to time T .
Remarks. The martingale characterization o f a price system that admits no arbitrage opportunities was pioneered by Harrison and Kreps (1979) . Their paper, in turn, was motivated by an observa tion made by Cox and Ross (1976) in an option pricing context . In deriving the martingale property, we assume that individuals' pref erences have expected utility representations. This is far more re strictive than necessary. In a finite state model, the only condition
needed is that individuals' preferences are strictly increasing. In an
infinite states model, some other technical conditions are needed ; see Kreps (198 1 ) . The martingale property assumes its full power in a continuous time continuousstates model. Interested readers should
256
Foundations for Financial Economics
consult Duffie and Huang ( 1985), Duffie ( 1985 ) , Harrison and Pliska ( 198 1 ) , and Huang ( 1985a, 1985b ) . A good reference for the correspondence between a dynamic and a static economy is Kreps ( 1979) . Cox and Huang ( 1986, 1 987 ) , Pages ( 1987 ) , and Pliska ( 1986 ) apply this idea to solve an optimal consumption and portfolio problem in a continuous time model. Our brief discussion on the convergence of the binomial option pricing formula to the BlackScholes formula is taken directly from Cox and Rubinstein ( 1985 ) . Reference s
Black , F . , and M . Scholes. 1973. The pricing of options and corpo rate liabilities. Journal of Political Economy 81:637654. Cox, J ., and C. Huang. 1986. A variational problem arising in finan cial economics with an application to a portfolio turnpike theo rem. Working Paper #175186. Sloan School of Management, Massachusetts Institute of Technology. Cox, J., and C. Huang. 1987. Optimal consumption and portfolio policies when asset prices follow a diffusion process. Working Paper #192687. Sloan School of management, Massachusetts Institute of Technology. Cox, J ., and S. Ross. 1976. The valuation of options for alternative stochastic processes. Journal of Fin ancial Economics 3, 1 45166. Cox, J . , S. Ross, and M. Rubinstein. 1979. Option pricing: A sim plified approach. Journal of Financial Economics '7:229263. Cox, J ., and M. Rubinstein. 1985. Options Markets. PrenticeHall, Inc., New Jersey. Duffie, D., and C. Huang. 1985. Implementing ArrowDebreu equi libria by continuous trading of few longlived securities. Econo metrica 5 3 , 1 3371356. Duffie, D. 1985. Stochastic equilibria: Existence, spanning number, and the "no expected gain from trade" hypothesis. Econometrica 54:11611184. Harrison , M., and D. Kreps. 1979. Martingales and arbitrage in multiperiod securities markets. Journal of Economic Theory 20,
Arbitrage Multiperiod Securities Markets II: Valuat ion by
257
381408. Harrison, M., and S. Pliska. 198 1 . Martingales and stochastic inte grals in the theory of continuous trading. Stochastic Processes and Their Applications 11 , 261271. Huang, C. 1985a. Information structure and equilibrium asset prices. Journal of Economic Theory 35, 337 1. Huang, C. 1985b . Information structure and viable price systems. Journal of Mathematical Economics 1 3 , 215240. Kreps, D. 1979. Three essays on capital markets. Mimeo. Graduate School of Business, Stanford University. Kreps, D. 198 1 . Arbitrage and equilibrium in economies with in finitely many commodities. Journal of Math ematical Econ omics 8 , 1535. Pages, H. 1987. Optimal consumption and portfolio policies when markets are incomplete. Working Paper #lilil3il6. Sloan School of Management, Massachusetts Institute of Technology. Pliska, S. 1986. A stochastic calculus model of continuous trad ing: Optimal portfolios. Mathematics of Operations Research 1 1 , 371382.
C H A P TER 9 F INA N C IAL M A RKETS WITH D I F F ERENTIAL I N F O R M AT I O N
In our previous discussions, i t was assumed either implicitly or explicitly that individuals h ad homogeneous information. Differential information among individuals is, however, an impor tant aspect of financial markets. Regulations on insider trading pro vide evidence for this claim. Many mutual fund managers claim to trade on the basis of superior information. In addition, the exis tence of active markets for information, such as advisory services and newsletters, signifies the possibility that individuals are differ entially informed a p riori. The first part of this chapter will develop a model of financial markets that incorporates the heterogeneity of information among individuals. The focus of this discussion will b e o n how private information of individuals gets transmitted to the public through prices and on the appropriate equilibrium concepts. The idea that individuals may be differentially informed is plau sible, and it is even more plausible that the managers of a firm have private information about the firm that is not shared with outside investors . In the the second part of this chapter, we consider an 9.1.
259
260
Foundations for Financial Economics
entrepreneur seeking outside investors to finance his project . The entrepreneur is privately informed about the quality of the project, which he can only credibly reveal by taking some observable action. This action is termed a signal. We will introduce the notion of a sig n alling equilibrium, in which an entrepreneur takes certain actions that would not have been taken were there not i nformation asymme try. 9 . 2 . We will first discuss in general how differential information among individuals can be modeled. In the course of this discussion, we will make extensive use of certain set theoretic notions developed in Section 7 .2. O ur attention will be limited to twoperiod economies where trades occur at time 0 and consumption of a single good occurs at time 1. We assume that before trading occurs at time 0, each in dividual receives a piece of private signal conveying to him some information about the state of nature that will be realized at time 1 . Recall from Chapters 7 and 8 that information revelation over time can be modeled by finer and finer partitions of the possible states of nature. In this context, it is natural to say that two individuals ·are differentially informed if their information structures are not identi cal. In a two period economy, we can therefore model a signal that an individual receives before trading to be a p artition of the state space. Formally, we fix a state space 0, each element, w, of which is a possible state of nature. Assume for now that 0 has a finite number of elements . Individuals are endowed with a common information structure at time 0, denoted by F = { lo , JI } , where lo and l1 are partitions of 0. Assume that 10 = {0} and that l1 is the partition in which each state is an event. Recall that the latter condition simply says that the true state of nature will be revealed at time 1 to all individuals. Individuals' prior beliefs are such that every state of nature is of strictly positive probability. Before trading at time 0 , however, individual i receives a private signal represented by a partition lJ . For example, suppose that 0 = {w1 , . . . , ws } and that lJ = { {w1 , w2 , w3 } , {w4 , ws } }. If, for example, wg is the state of nature that will be realized at time 1, then individual i learns from
Financial Markets with Differential Information
261
his signal that the state to be realized at time 1 is one of {wh w 2 , ws } . Equivalently, a private signal for individual i can also be represented by a random variable Y; defined on 0. For instance, the random variable if w E {w1 , w2 , ws } , Yi ( w ) = if w E {w4 , w5 } ,
{�
conveys the same information as the partition 1j of the example above, where we note that Yi ( w) is the realization of Ys when the state of nature is w. If Ys == 1, individual i learns that the true state of nature is one of {w1 , w2 , w3 } , and if Y = 2, individual i learns that the true state of nature is one of { w4 , w5 } . That is, a realization of Ys tells individual i whether the true state is one of { w1 , w2 , ws} or one of {w4 , w5 } . In this sense, we can say that Ys generates the p artition
1j.
As usual, we assume throughout that individuals strictly prefer more to less. Now we claim that in the current context, there does not exist a competitive equilibrium with complete markets when there exist some nonsatiated individuals i and k such that J"j =/: JQk . ( Here we remark that prices in a competitive equilibrium have only one function, namely to determine an individual's budget constraint. ) To see this, it suffices to consider an example. We take the state space and individual i's signal to be the same as those considered in the last paragraph. Let there be another individual k with signal represented as 10k = { {wb w2 } , {ws , w4 , w5 } }. Suppose to begin. with that either w1 or w2 is the state that will be realized at time 1 . Then, after receiving his signal, individual i learns that the true state is one of {w1 , w2 , w3 } . On the other hand, individual k learns that the true state of nature is one of {w1 , w2 } . Since markets are complete, a competitive equilibrium must assign a price to each state contingent claim. Consider the state contingent claim that pays off in state ws . After receiving his signal, individual k learns that ws is not the true state. However, even after receiving his signal , individual i still believes that ws may be the true state. If the state contingent price for state ws is strictly p ositive, by the hypothesis that utility functions are strictly increasing, indi vidual k will short sell an i nfinite amount of the statews contingent claim and use the proceeds to buy state contingent claims paying off
262
Foundations for Financial Econ omics
in states w 1 and w2 . As a consequence , the market for the statewa contingent claim cannot clear. On the other hand, if the state con tingent price for state w3 is nonpositive, individual i will purchase an infinite amount and the market for it cannot clear. Thus , there does not exist a competitive equilibrium when the true state is either w1 or w2 . Arguments similar to those in the last paragraph show that there cannot exist a competitive equilibrium with complete markets if the true state is any of {w1 , . . . , ws } . Thus, there does not exist a com petit ive equilibrium with complete markets. 9.3.
A casual examination of the nonexistence result of the last section reveals the reason why an equilibrium does not exist: the in dividuals do not learn from observing prices. For example, when w1 is the true state, a zero price for the state contingent claim paying off in state w3 would be the equilibrium price if individual i con cluded that the true state must not be state w3 when he observed that the price for state Wg was 0. In this case , an equilibrium can be constructed as follows when individuals i and k are the only two individuals in the economy : Use Bayes rule to compute the two indi viduals' posterior probability assessments about the states of nature conditional upon the knowledge that the true state is one of {wt . w2 } . There exists, under stand ard conditions, a competitive equilibrium given these posterior probability assessments. The competitive equi librium price system must assign zero prices to states Wg , w4 , and w s . This competitive equilibrium price system i s an equilibrium price sys tem for our economy with differential information , when individuals learn from observing equilibrium prices and the true state is either w1 or w2 • Note that in this equilibrium, both individu als know that the true state is one of {w1 , w2 } and do n o t know which is the true state. That is, the difference in the information conveyed by the signals they received is symmetrized by the equilibrium prices. One can apply the same arguments to cases where the true state of nature is any of {w3 , w4 , ws } and determine that the conclusion reached in the last paragraph is valid in all cases. Moreover, this conclusion is not limited to the special case of two individuals and five states considered and is valid in general. The reader will b e
Fin ancial Markets with Differential Information
263
asked to show in Exercise 9 . 1 that, in fact, a necessary characteristic for a price system to be an equilibrium price system when markets are complete and individuals learn from observing the price system is that the price system symmetrizes information among individuals. 9 . 4 . Note that in the construction of an equilibrium in Sec tion 9 . 3 , unlike in a competitive equilibrium, a price system has two roles: first, determining an individual's budget constraint as in a competitive equilibrium; second, conveying information . An equilib rium where a price system plays these two roles will be termed a rational expectations equilibrium . This term is the same as was used in Chapters 7 and 8 but applies in a different context. In Chapters 7 and 8, individuals are endowed with the same information struc ture. They are said to h ave rational expectations if they agree on the mapping from (w , t) to the price system S (w , t) . Here , individuals are endowed with different information structures. We will not give a general definition of a rational expectations equilibrium, because it would involve some mathem atical concepts that are unnecessary for our purposes here. We will examine , however, a class of models with differential information among individuals and give a formal defini tion of a rational expectations equilibrium in this context . Readers interested in general discussions of rational expectations equilibria can find references in the remarks at the end of this chapter. Analyses in previous sections and Exercise 9.1 show that when markets are complete and individuals have differential information before trading occurs, it is necessary that individuals have rational expectations for an equilibrium to exist. In addition, in any ratio nal expectations equilibrium with complete markets, the price system must symmetrize the difference in information amon g individuals. In such event, the price system is said to be fully revealin g . Moreover, a fully revealing rational expectations equilibrium with complete mar kets always exists under standard conditions. Thus, observationally, there is no distinction between a rational expectations equilibrium and a competitive equilibrium in a corresponding artificial eco n o m y that is ide ntical in every aspect except that individuals are endowed with the common information held in the rational expectations equi librium.
264
Foundations for Financial Economics
Note that although our previous discussion is based on an econ omy with a finite number of states of nature, all our conclusions hold , with some additional regularity conditions, in economies with infinitely many states of nature . 9 . 5 . Now we turn our attention to economies with incomplete markets. Unlike the complete markets case , a ration al expectations equilibrium need not exist and a competitive equilibrium can exist. We will present an example to show the former result in this section and present another example to demonstrate the latter result in the next two sections. We consider an economy with two groups of indiv iduals which are equal in number , the informed and the uninfo rmed. The state space is denoted by 0. There is a single consumption good avail able only at time 1 , whose spot price at time 1 is a random variable denoted by q, which we assume to be normally distributed. An in dividual, informed or uninformed, has a negative exponential utility function defined on time1 random wealth,  exp{ aW}. At time 0, an individual can trade on the futures contract of the single consump tion commodity, whose unit price, the futures price, is denoted by F. Note that a futures contract is a financial asset that promises to deliver a unit of the consumption good at time 1 for a price F, the futures price, determined at time 0. No resources need be committed at time 0 for the purchase of a futures contract. The accounts are settled at time 1 . The markets are obviously incomplete, since there exist infinitely many states and only one security, the futures contract. Note that the structure of this economy is a little different from our usual setup. The single consumption good is not taken to be the numeraire, and the utility function is defined on wealth rather than on consumption . . The reader can think of our model here as only a "slice" of an econ omy with multiple consumption goods where the numeraire is not the consumption good discussed in the "slice." Before trading at time 0, an informed individual receives a pri vate signal denoted by Y . This private signal takes on two possible
Financial Markets with Differential Informa tion
y
values
{
265
1 if w E O t i 2 if w E 02 ; where { 0 1 , 02 } is a partition of n . Assume that Ot and 02 are of equal probability. Conditional on Y = n, n = 1 , 2, q is normally distributed with mean mn and variance a 2 • Assume that m 1 =I m 2 . Conditional on Y = n, an uninformed individual is endowed with kn > 0 units of time1 consumption, with n = 1, 2 . Assume that k t =F k2. An informed individual, however, is endowed with a constant k > 0 units of time1 consumption. An uninformed indi vidual does not receive any private signal before trading takes place at time 0 and his prior beliefs about q are, therefore, an even mix ture of two normal random variables, which is generally not a normal random variable . The price at time 0 of the futures contract F may depend upon the realization of the signal received by the informed individuals. If F(Y = 1) is not equal to F(Y � 2 ) , an uninformed individual can infer the private signal received by an informed individual from the futures price. In such event, the equilibrium price for the futures contract is fully revealing . Conditional on = n, an informed individual i solves the fol lowing problem: max El exp {  a W }jY = n] 6; (9.5 . 1 ) s.t. W = fh (q  F) + qk, where (}i i s the number of futures contracts purchased by individual i at time 0. Conditional on Y = n, W is normally distributed with mean and variance: E[W jY = n] = ( Oi + k ) mn  Oi F, Var(W jY = n ) = (Oi + k) 2 a 2 • Note that the conditional variance is independent of the realization of Y , which is a general property of normal random variables. The distribution of exp {  a W } conditional on Y = n is lognormal. Direct computation yields E[  exp {  a W} j Y = n]
Y
=
266
Foundations for Financial Economics
As an exponential function is strictly increasing in its exponent, (9.5 . 1 ) becomes
(9.5 .2) This is a strictly concave program, and its unique solution is charac terized by the first order condition:
9.
_
' 
mn au 2
F
_
k
.
(9 .5.3)
Suppose there exists a rational expectations equilibrium . with a pr£ce fu n c tion a l F (Y ) . We first claim that F(Y) must be fully revealing. Suppose this is not the case. That is, there exists a ra tional expectations equilibrium price functional F(Y) with F (Y = 1) = F (Y = 2) . From (9.5.3) , we know that the optimal demand for the futures contract for an informed individual when Y = 1 is clif f erent from that when Y = 2, since m 1 :;;i: m2 • By the hypothesis that the fu tures price does not vary across two possible realizations of an informed individual's private signal, an uninformed individual max imizes expected utility according to h is prior b eliefs about q. Thus his optimal demand for the fu tures contract does not vary across the realizations of Y. It then follows that the markets for the futures contract cannot clear so that a rational expectations equilibrium, when one exists, must be fully revealing. Next let F(Y) be a price functional for the fu tures contract such that F(Y = 1) :;;i: F (Y = 2) . Facing the price functional F (Y ) , an un informed individual learns the realization of Y received by an informed individual . Conditional on Y = n, an uninformed individ ual j's optimal demand for the futures contract is a solution to the following problem max E[ exp {  a W } J Y 6J.
s.t.
= n]
W 9; (iJ  F) + ilkn , =
(9.5.4)
where 9; denotes the number of fu tures contracts purchased by indi v idual j. Using identical arguments as in the analysis of (9 . 5.1 ) we get , condit ional on Y = n ,
9,·
=
m..:..":.._ cF __, au2
_
k
n.
(9.5.5)
267
Financial Markets with Differential Information The equilibrium futures price, conditional on Y mined by the market clearing condition
It is �
F( Y
= n
)
Consider the following data: k1 = 3, and kz = 5. Then
(};
+ 9; = 0.
=
au 2 . mn  (k + kn) 
m1
F(Y = n )
=
=
3,
0.5,
m2
=
2
=
n =
4,
a = 1,
u2
n,
=
is deter
(9.5.6) 1, k
=
2,
1, 2.
This contradicts the fact that F(Y) i s fully revealing. Thus there does not exist a rational expectations equilibrium! Note that the above example is not robust in that with a slight change of data there can exist a rational expectations equilibrium. For example, consider the same data as above except that mz = 5. Then 0 . 5 � f n = 1; F(Y = n ) = 1.5 1f n = 2;
{
and is a fully revealing rational expectations equilibrium price func tional. The crux of the matter, however, is that a rational expec tations equilibrium may not exist for a given specifi cation of the parameters of the economy. 9.6.
In the previous section, we discussed an example of nonex istence of a rational expectations equilibrium when markets are in complete. Recall from Section 9. 2 that when markets are complete and individuals are asymmetrically informed, there does not exist a competitive equilibrium. We now demonstrate, through an exam ple, that there can exist a competitive equilibrium when markets are incomplete. In this competitive equilibrium, however, individuals ignore the informational content of a price system. Consider the following model of financial markets. There are two periods, time 0 and time 1 . At time O , individuals trade in one risky and one riskless asset. At time 1, individuals consume an d the economy ends . One share of the risky asset pays X units of the
268
Foundations for Financial Economics
single consumption good at time 1 , where X is a random variable. The risky asset has a total supply of one share. The riskless asset is in zero net supply and pays one unit of the single consumption good for sure at time 1. There are I individuals in the economy, indexed by i = 1 , 2, . . . , I . Individual i is characterized by his utility function ui ( z ) =  exp {  ai z } on time 1 consumption and his endowment of a number of shares of the risky asset, denoted by Oi . Individuals are not endowed with the riskless asset. All individuals have a common probability belief P, under which X is normally distributed with mean mx and variance a ; . Before trading takes place, individual i receives a private signal
Yi ,
(9.6.1)
where X and (ei ){= 1 ar:e mutually independent and normally dis tributed and E[ei] = 0, Vi. The variance of ei is denoted by a;, . Implicit in the above setup is a state space !1 , on which all the random variables are defined. This state space has infinitely many elements, because the random variables are normally distributed. Each realization of Yi , denoted simply by Y i , tells individual i that some states are possible and some are not . Individual i then updates his probability assessments about X according to Bayes rule. If the realization of his private signal is Yi , individual i's problem IS
(9.6.2)
where we have normalized the price of the riskless asset at time 0 to b e 1, S x is the price of the risky asset at time 0, and ()i is the number of shares of the risky asset individual i chooses to hold. With the time0 price of the riskless asset normalized to be 1, the rate of return on the riskless asset is zero. Note that, conditional on f'i = Yi > X is normally distributed with mean and variance as follows: (9.6.3) (9.6.4)
where
269
Fin ancial Markets with Differential Information
Note that the conditional mean of X given Vi = Yi is equal to the unconditional mean of X, m z , plus the "beta" of X on Vi, denoted by Pz11; , times the deviation of yi from mz . Similarly, the condition al variance of X given Vi = Yi is equal to the unconditional variance, a ; minus P z JJ ; times a ; . As we noted before, th � conditional variance , of (9.6.4) is independent of the realizations of J'i. Arguments similar to those used in deriving (9 .5.3) and (9.5.5) show that .
()' =
E[ X I'Yi _= _Yi ]  Sx aiVar(X\Yi = yi )
Putting
Rx
=
.
• = 1,2,
.
.
. , 1.
(9.6.5)
X/Sx ,
where Rz is the return on the risky asset, and substituting this ex pression into (9.6.5) gives O '·S z
_

E[Rx i'Yi = Yi]  1 ai Var(Rz \ 'Yi = Y i)
i = 1 , 2 , . . , 1. .
(9.6.6)
Relation (9.6 .5) says that individual i will invest in the risky asset if and only if he expects, conditional on his signal, that the time1 payoff of the risky asset will be strictly greater than its time0 price. His demand for the risky asset is a decreasing function of its price. On the other hand, (9.6.6) implies that the dollar demand for the ri�ky asset is strictly positive if and only if the risk premium is strictly p ositive. (Recall that the riskless interest rate is z ero. ) In addition, the more risk averse individual i is ( or the larger ai ) , the smaller the absolute values of ()i and ()iSz are. These results are consistent with our analysis in Chapter 1 . Note also that (9.6 .5) and (9.6.6) are independent of individual i's endowment, oi , because negative exponential utility functions exhibit constant absolute risk aversion and there is no wealth effect . 9. 7. Having computed individuals ' demands for risky assets, we are now ready to construct a competitive equilibrium. As we
270
Foundations for Fin ancial Economics
have normalized the p rice of the riskless asset to be equal to one, a competitive equilibrium is composed of a price, Sz , for the risky asset at time 0 such that the market for the risky asset clears. By Walras law , once the market for the risky asset clears, the market for the riskless asset clears also. O ur task , therefore, is to find Sz such that (9.7 . 1)
We can accomplish this by summing (9.6.5) across individuals and equating it to 1 : 1
=t =
i= l I
E[X/ Yi � �;J_ Bz a; Var (X / Yi 

 y;)
I
L E[X/�_:= y; ]  Bz L a; Var (X a; Var (X / Yi y; ) / Yi y; ) i= l
Solving for Sz gives
=
_1 _
i= l
( L: I
i= l
a;
 1
Var(X / Yi
=
y;)
=
)
1
.
(9.7.2 )
Relation (9.7 .2) gives a competitive equilibrium price for the risky asset. This equilibrium price for the risky asset is equal to the certainty equivalent of its time1 payoff. The certainty equivalent is a weighted average of individuals' expectations of X given the signals they received minus a risk a diustm e n t fa c t o r . The weight for individual i's conditional expectation of X is
which is larger the smaller a; and Var (X / Y"i = y;) are, ceteris paribus . That is, the less risk averse individual i and the more precise the signal Yi (as captured by a smaller Var(X/fi = y;) ) are, the more highly weighted individual i's estimate about X is in the equilibrium
Fin ancial Markets with Differential Information
271
price. S imilarly, the risk adj ustment factor will b e smaller when individuals are less risk averse and their signals are more precise. The competitive equilibrium constructed above makes a gre at deal of intuitive sense. The comparative statics discussed all seem to go the right direction. Unfortunately, this competitive equilibrium is not stable in the following sense. Note that the price of the risky asset in the competitive equilibrium is a linear function of individ u als' signals. To see this, we simply substitute (9.6.4) into (9.7.2) and observe that Var (X I fi = Yi) is independent of the realization of Jii . In order to trade, individuals must know the equilibrium price of the risky asset, since their demands as characterized in (9. 6.5) depend on that price. If they understand the relationship between the price and the signals (Yt , . . . , Y1 ) , as would happen if, for exam ple , the economy were repeated and individuals learned from history, they would realize that the risky asset price contained valuable in formation about the payoff of the risky asset. In such event , after observing the equilibrium price in the competitive equilibrium, indi viduals would acquire extra information and form their demands for the risky asset using their new posterior beliefs. It then follows that the competitive equilibrium price may fail to clear the market and the competitive equilibrium may break down . S uppose instead that there is a price function
which gives prices for the risky asset for all possible realizations of the ioint signal, (Yt , . . . , Y1) , in that , for a given realiz ation of the j oint signal, ( Y t , . . . , YI ) , Sz (Yl , . . . , YI ) is the price for the risky asset. Moreover, the functional relationship b etween Sz and the j oint signal is kn o wn to individuals in that for every realiz ation ( y t , . . . , YI ) of (Yt , . . . , Y1) , an individual i calculates his optimal demand for the risky asset using the posterior belief conditional on his own signal fi = Yi and the price Sz(Yt , . . . , Y1) = Sz (Yl , . . . , yl ) and the market clears. Then the price functional gives an equilibrium price for the risky asset for all possible realizations of the joint signal. This equi librium is stable in the sense that it will not break down even after many repetitions of the economy, because individuals know the func tional relation between the equilibrium price and the j oint signal and
272
Foundations for Financial Economics
have already taken it into consideration in calculating their optimal demands. This equilibrium is a ratio nal expectations equilibrium. Individuals are said to have rational expectations if, in equilibrium, they understand the functional relation between the equilibrium price ( of the risky asset) and the j oint signal. In a rational expectations equilibrium, all individuals have rational expectations. 9 .8 . We will show that there exists a r ational expectations equilibrium in the economy constructed in Section 9.6, when we add the assumptions that a � ; = a� , Vi. The proof is by construction: construct a price functional and show that it is indeed a rational exp ectations equilibrium price functional. Before proving this asser tion , however , we discuss some results in probability theory that will be useful. Let Z be a random variable, let z denote its realization, and let fi (yi , z l :z:) be the j oint density of Y; and Z conditional on X. The random variable Z is a sufficient statistic for /i ( Yi , zl :z:) if there exist functions 9 1 ( · ) and 9 2 ( · ) such that for all Yi and z,
(9 .8 . 1 )
I f Z i s a sufficient statistic for /i( Yi , zl :z:) , the conditional density of X given Y; and Z , denoted by h ( :z: I Yi , z) , is independent of Y; . To see this, we note that by Bayes rule, . h ( :z: I y, , z )
9 ( x )fi (yi , z l :z:)  + oo 9 ( ) ( Yi , zl ) ' Loa x fi :z: d:z:
_
(9.8.2)
where 9 ( · ) is the density function for X. As Z is a sufficient statistic for fi(yi , zl :z: ) , we substitute (9.8.1) into (9.8.2) to get
��
��
, h ( :z: l y; , z ) = + !, ( ) l ( Yi , z ) 2 ( :z:) J oo 9 :z: 9 1 (Yi , z 9 2 z, :z:) d:z: 9 (:z:)92 ( z, :z:)
(9 .8.3)
The righthand side of (9.8.3) is independent of Yi , which was to b e proved.
273
Financial Markets with Differential Information
Now we claim that
is a sufficient statistic for /i ( Yi , !i l x ) , the joint density of fi and Y conditional on X, where y is a realization of Y . Intuitively, this follows because, with the additional assumptions about the distribu tions of £i , individuals are getting signals that are independent and identically distributed conditional on X. Thus the mean of the sig nals is more informative than any individual Yi. More formally, we want to show that (9.8. 1) holds when we take Z to be Y . Condi tional on X x , fi is normally distributed with mean x and variance a; , and Y is normally distributed with mean x and variance a; I[. Conditional on }( = x 1 the covariance of Yt and Y is a; I I. Thus con ditional on x, ('iii , Y) are bivariate normal with a mean vector ( x, x) and a variancecovariance matrix
= X=
Therefore,
Define 9 1 ( Yi • Y ) 
and 9 2 ( i/ , x)
=
_
=
I ( 2 11" ) _ 1 u� ,;r=l exp
{� /
{
1 I 2  2 a � ( I _ 1 ) ( Yi  2 YYi )
exp  a� ( _ 1 ) (2 xy  x 2 + I(y  x) 2 )
Then /i ( Yi 1 i! l x)
=
_
}
}
,
.
91 (Yi 1 y)g 2 (y, x) , and we have proved our assertion.
274
Foundations for Financial Economics
As Y 1s a sufficient statistic for /; ( yi , il ! x ) , the density of X conditional on Pi = Yi and Y = y, h ( x l yi , y) , is independent of Yi · As a consequence,
E[ XI "Yi Var(X! Yi
= =
Yi , y
=
Yi • Y
=
y] y)
=
=
=
E[X ! Y
=
y]
mz + fJzg (Y  mz) ,
Var(X! Y
= Uz2 
(J
=
2
y)
(9.8.4)
zjj Uz ,
where is the "beta" of X on Y . The interpretation of (9.8.4) is similar to that for (9.6.3) and (9 .6.4) . Using similar arguments, we can show that Y is a sufficient statistic for the joint density of ("Y1 , . . . , Y1 , Y) conditional on X and, thus,
E [X ! Yl
=
Yl l . . . ' YJ
=
YI . y
=
y]
Var ( X lY1
=
Yl • . . . , Y1
=
YI , Y
=
y)
=
=
=
E [X ! Y
=
y]
mz + fJzg (y  mz) , =
(J 2 2  U z  zjjU'z · 
Var (X ! Y
y)
(9.8.5)
9 . 9 . Now w e are ready t o construct a rational expectations equi librium. F irst we note from (9 .7.2) that if every individual received a signal equal to Y , the competitive equilibrium price functional would be a linear function of Y . This functional will be our candidate for a rational expectations equilibrium price functional. The reasoning is as follows: As individuals have rational expectations, once they see the equilibrium price announced by the auctioneer, they can invert the price to solve for the realization of Y . Knowing Y , which is a sufficient statistic for fi (Yi , il l x) , a rational individual's optimal de mand for the risky asset will be independent of his own signal. In fact, individuals' optimal demands are equal to those in an artifi cial eco n o m y where all individuals received the signal Y . Thus our
275
Financial Market s with Differential Information
price functional clears the market for the risky asset and is a rational exp ectations equilibrium price functional. Formally, consider an artificial eco nomy identical to the one currently being analyzed except that all individuals receive a signal equal to Y . As there is no differential i nformation among individuals in this artificial economy, we simply look for a competitive equilib rium. Given a realization fi of Y , the optimal demands for the risky asset are ; O =
E[ X IY
:= yJ  s"' ,
i = 1 , 2, . . . , I, Var ( X i ? = y) and the competitive equilibrium price for the risky asset i s
S:z:
=
E[ X !Y
a;
=
fi]  Var (X ! Y
=
fi )
( L :. ) 1
i=l
•
(9.9 . 1 )
1
(9.9.2)
where we have used (9.7.2) and (9.8.4) . Observe that S:z: is a linear function of y with fixed coefficients in (9.9.2) . We will use S:z: (Y) to denote the value of this linear functional when Y = y. Note that there is a onetoone correspondence between y and S:z: (i/), and S:z: (Y) is therefore invertible. The price functional of (9.9.2) is intuitively appealing. The price of the risky asset is higher the higher the realized average signal y, ceteris parib us. For a given change of y, the response of S:z: is proportional to f3:z:g , which lies in (0, 1) by the definition given i n (9.8.4) . The proportional increase of S:z; is higher t h e smaller a; is, or equivalently the more precise the signals are. Also, ceteris paribus , the more risk averse individuals are, the lower the price for the risky asset, since they require a higher risk premium to hold the risky asset. The comparative statics of S:z: with respect to a; and a; are left for the reader. Note that the competitive price functional of (9.9.2) is also an equilibrium price in another artificial economy where individuals share their signals before trading , that is, where individuals observe
276
Foundations for Financial Economics
the j oint signal cvl , . . . , YI) before trading . This is so, because y is also a sufficient statistic for the joint density of (l\ , . , YJ , Y ) given X. The reader will be asked to prove this claim in Exercise 9.2. Now we claim that the price functional of (9.9. 2) is a rational expectations price functional. G iven this price functional, individual i's rational expectations demand for the risky asset is .
fJ;
=
.
E[X I Yi = y; , Sz ( Y ) = Sz (ii)]  Sz ( ii ) a;Var (X If; = Y; , Sz (Y ) = Sz ( ii )) E [XIY.: = y; , Y = :fi]  Sz (:fi) a; Var (X I f; = y; , Y = :fi) E[X i Y = :Y]  Sz (:fi) a;Var (X IY = :ii )
(9.9.3)
where the first equality follows from the fact that Sz (:fi) is an invert ible function of :fj , and the second equality follows from (9.8 .4) an d the fact that Y is a sufficient statistic for fi (y; , :fi l x) . Note that the ri ghthand side of the third equality is identical to the righthand side of (9.9 .1) when we take Sz to be Sz (:fi ) . Thus a rational individ ual 's optimal demand when faced with the price functional Sz (Y) is identical to that in the artificial economy. We know that the price functional clears the market in the artificial economy. Thus the price functional clears the market in the rational expectations economy and is a rational expectations equilibrium price functional. 9 . 1 0 . The rational expectations equilibrium constructed in Sec tion 9. 9 has the following properties. First, the information conveyed by the equilibrium price is superior to any private signal in the sense that the price information is a sufficient statistic. Thus, given the equilibrium price , private information becomes redundant. Second, the rational expectations equilibrium is identical to a competitive equilibrium in an artificial economy where individuals share their private information. These two properties amount to saying that the rational expectations equilibrium price system symm etrizes the differences in information among individuals and thus is a fully re vealing price system. A rational expectations equilibrium with a
Fin ancial Markets with Differential Information
277
fully revealing price system is said to be a fully rev ealing rati o n al expectations equilibriu m .
A fully revealing rational expectations equilibrium seems eco nomically attractive. The equilibrium price system aggregates di verse private information among individuals efficiently in that it re veals a sufficient statistic for the diverse information . Moreover, the equilibrium allocation cannot be improved on by a social planner who has access to the j oint private signals. A fully revealing rational expectations equilibrium, however, has the following problems. As the equilibrium price system conveys in formation that is superior to an individual's private signal, the op timal demand for the risky asset is independent of the individual's private signal! The optimal demand depends only upon the equilib rium price . If an individual's optimal demand is independent of his or her private signal, how can the e qu i li b i· i u m price system aggregate individuals' diverse private signals? In addition, an individual will not have an incentive to collect private information if it is costly to do so  a fully revealing equilibrium price system renders the informa tion collection activity an unprofitable proposition. It then follows that if no individuals collect private information , certainly there is no diverse private information to aggregate. These are the p aradoxes associated with a fully revealing price system. The paradoxes can be resolved if the price system aggregates in formation only partially in the sense that the price information is n o t a sufficient statistic for an individual's private signal. I n such event, the optimal demand of an individual for the risky asset will depend not only upon the price information but also upon his own private signal. This solves the first paradox. Since the optimal demands for the risky asset depend upon indivi? uals' private signals, individuals have incentives to collect private information even at a cost. This solves the second paradox . Now the question is under what scenario will a price system be partially revealing? Note that in the fully revealing rational ex pectations equilibrium considered earlier, there is only one source of uncertainty, namely the time1 payoff of the risky asset X. Indi vid u als receive private signals about X, and the equilibrium price system reveals a sufficient statistic about individuals' private signals. The
278
Foundations for Financial Economics
price system is fully revealing, because it is a nontrivial linear func tion of the sufficient statistic and is thus invertible. When the price of the risky asset increases, individuals can infer that this occurs be cause the demand for the risky asset increases, which is in turn due to more optimistic private signals, on average . Suppose instead that there i s a n additional source of uncertainty in the economy, for example , the aggregate supply of the risky asset. Then when the price of the risky asset increases, individuals · may not be able to tell for sure whether it is because the private signals are more optimistic or because the aggregate supply is smaller. In such an event, the price system does not provide information that is superior to an individual's private signal and is not fully revealing. We will show in the next section that such an equilibrium exists. Note that the assumption of an uncertain aggregate supply is not the most natural. For example, in a stock market economy, the total supply of a common stock is the number of shares outstanding and is common knowledge. This somewhat contrived assumption , however, provides tractability to our analysis. 9.11.
Consider again the economy constructed in Section 9.6. Assume that individuals have rational expectations and the same co,. efficient of absolute risk aversion. Assume in addition that individual i's endowment of shares of the risky asset is a draw of a normally dis tributed random variable Vi . The r andom variables (V1 , . . , VI ) are indep endent and identically distributed with mean zero and variance u� . The zero mean assumption is made for convenience and can b e r e i axed easily. A realization of Vi is denoted b y Vi. I t follows that the total supply of the risky asset is a realization of the random variable .
z
=
I
L: vi , i =l
which i s normally distributed with mean zero and variance Iu� , where we recall that I is the total number of individuals. A real ization of Z is denoted by z. We also assume that X, \ii , and the li 's are mutually independent and li has mean zero and variance ui , which is constant across individuals. Given that a realization
Financial Markets with Differential Information
of CVt , . , VI ) is ( v � , risky asset is .
.
.
.
.
, VI ) ,
279
the market clearing condition for the
I
I
i=l
i=l
I: Bi ::::: .L
Vj .
We want to show that there exists a partially revealing rational expectations equilibrium. As in the fully revealing case, our proof will be by construction. 9 . 1 2 . We fi rst conjecture that the equilibrium price system is a linear function of the sufficient statistic Y and the aggregate supply z:
(9.12.1)
where "{, b 1 > 0, and b 2 > 0 are some unknown constants. This price functional is an increasing function of Y and a decreasing function of the aggregate supply Z. Given this conjectured price functional, we can compute an individual's optimal demand conditional on a real ization of the price functional, his private signal, and his endowment of shares of the risky asset. We then ask what the values of "{, b 1 1 and b 2 should be such that the optimal demands of individuals clear the market. G iven that the realizations of and Z are vi , y;, and z , respectively, and given t h e price functional of (9 . 1 2 . 1 ) , individual i's optimal demand for the risky asset is
Vi, Y,,
( 9 . 1 2 .2) The derivation of (9. 1 2.2) is identical to that for (9.6.5) , except that now individual i forms expectations about X conditional on his own private signal as well as on his endowment of shares of the risky asset and the realized price of the risky asset. As (9 . 1 2 . 1) is a linear func tion of normally distributed random variables, Bz (Y , Z) is normally distributed with mean and variance:
Foundations for Fin ancial Economics
280
Therefore, (X, Vi, ¥;, So: (Z, Y)) are multivariate normally distributed with a mean vector
and a variancecovariance matrix

( Vu21 V12V22 ) v
"'
0 a2 b l; ;
(
0 a u2 0  b2a�
b,a; a "'2 0  b2a� b 1 ( a; + ail I ) a; + a� bHa ; + a; I I ) + Ib � a� b 1 (a; + ail I)
)
'
where Yu = a; , Vn is a 1 X 3 vector, v2 1 is a 3 X 1 vector, and v2 2 is a 3 x 3 matrix. The conditional expectation and conditional variance of rela tion (9. 12 . 2) can be explicitly computed using multivariate normal distribution theory. They are
E[Xl'Vi =
v; ,
Vi = y; , S"' = Bz (fi, z)]
(
i
)
y ; � m"' b 1 (Y  mo:)  b2 z = m z + r  1 { ( hb 2 a ; a � a : ( 1/ I  1 ) ) v ; + ( b�a ; ( a �  a � a : ) ) ( y;  mz ) + b 1 a ; a � a : ( 1J I  1) ( b 1 (y  mz)  b 2z) } ::: S'o + S'1Vi + S'2 Yi + S'3Y + S'4 Z ,
= mz + V12Viz 1
and
(9.12 . 3)
Fin ancial Markets with Differential Information
281
where
Note that ( �o, . . . , �4 ) are nonlinear functions of 61 and 62 and are independent of individual indices and that the conditional vari ance is independent of the realizations of Vi , 'Yi, S� n and individual indices. To simplify our notation, we denote the conditional variance of ( 9. 1 2 . 3 ) by H 1 , which we note is a nonlinear function of b1 and b2• H is the inverse of the conditional variance and is usually termed the p recision of the i nformation contained in vi, Y, , and Sz 9.13. The optimal demands for the risky asset of individuals ( 9 . 1 2 . 2 ) and ( 9.12 . 3 ) a r e based on a conjectured price functional ( 9 . 1 2 . 1 ) . For the conj ectured price functional to be a rational expectations equilibrium price functional, it must be th at the optimal demands clear the market for the risky asset. That is, we must have derived in
I
'"""' �o + �l vi + �2Yi + �3Y + �4z  "'/  6 1fj  62z � aH 1 i=1 (�o  "'f) / + (�2 + �3  6 1 ) /y + (/�4 + �1  l b2 ) z aH 1 = z, _
(9 . 1 3 . 1 )
for every possible realization of Vi , fi, Y , and Z. As Z is independent of Y, for (9 . 1 3 . 1 ) to hold for every possible realization of Y and Z, we must have
�0  "'{ = 0 (9 . 1 3 . 2) �2 + �3  61 = 0 1 /�4 + �1  /62 = aH • , �4 ) and H 1 are all nonlinear functions of 6 1 and
Recall that ( �0 , Thus, ( 9.1 3.2 ) is a system of three nonlinear equations in as many unknowns ("'!, 61 , 62 ) . •
62 •
.
.
282
Foundations for Financial Economics One can verify that a ( unique ) solution to (9.13.2) is
(9. 13.3)
It is then easily verified that the values of 7, b1 , and b2 in (9. 13 .3) are not only necessary but also sufficient for a ration al expectations equilibrium. That is, there exists a rational expec tations equilibrium with the price funct ional
where 7 , h , and 62 are defined in (9.13.3). Note that i n this equilibrium, h > 0 and 62 > 0. Thus, the equilibrium price for the risky asset is an increasing function of Y , the sufficient statistic, and a decreasing function of Z , the aggre gate supply of the risky asset . Because the aggregate supply of the risky asset is also a random variable, there is no onetoone relation ship b etween the price of the risky asset and the sufficient statistic Y. When the risky asset price increases, an individual is uncertain whether it is b ecause on average everybody is getting a better sig nal or because the aggregate supply of the risky asset is smaller. Therefore, an individual's optimal demand for the risky asset de pends on the information conveyed by the price of the risky asset, on h i s private signal, and on his own endowment o f the risky asse t. This resolves the paradox that if a price system is fully revealing, an individual's optimal demand is independent of his own private signal and the price system cannot, therefore , aggregate diverse pri vate i nformation among individuals. Moreover, as the price system aggregates diverse private information only partially, there exists an incentive to collect i nformation even when it is costly to do so. This resolves the paradox associated with costly information discussed in Section 9. 10.
Financial Markets with Differential Information
283
9 . 1 4 . From previous discussions, we learned that fully reveal ing rational expectations equilibria are not very interesting as they are not likely to arise  recall the paradoxes associated with fully re vealing equilibria. We have constructed a partially revealing rational expectations equilibrium in Section 9 . 1 3 for a very special economy  one with negative exponential utility functions and normally dis tributed returns and private signals. Unfortunately, little is known about partially revealing rational expectations equilibria outside the special case discussed . The combined effects of negative exponential utility functions and multivariate normally distributed returns and private signals make a linear price functional a feasible solution to the equilibrium problem. This gives tractability to an inherently very difficult problem. On the other han d , the tractability is not with out cost . The normally distributed return on the risky asset implies unlimited liability and negative consumption in equilibrium. We also know very little about rational expectations equilibria in multiperiod economies. Even in the special case discussed above, the multiperiod e xtension is a formidable task . This extension, when successful, will give rise to a much richer model. Questions such as to what extent historical prices contain information ab out future prices and whether historical volumes of trade play any informational role can only be answered in models of a multiperiod economy. 9.15. We concentrated o u r discussion i n earlier sections on securities markets equilibrium when individual traders possess di verse private information about the return on the risky asset. In the remainder of this chapter, we will consider situations where en trepreneurs possess inside information about projects for which they seek financin g . Outside investors would benefit from knowing the true characteristics of the proj ects. However, the entrepreneurs can not be expected to convey truthfully their inside information to the outside investors, as there may be substantial rewards for exagger ating positive qualities of their proj ects. If the qualities of proj ects can be verified ex post , financing contracts between the entrepren eurs and the outside investors can be written with terms dependent upon the e x post verifiable qualities. On the other hand, if the qualities of projects cannot be verified
284
Founda tions for Fin ancial Economics
ex post, the markets for financing may break down and proj ects with good qualities cannot be carried out resulting in inefficiency. To see this , we consider the financing of a family of projects whose quality is highly variable. While the entrepreneurs know the quality of their own proj ects, the outside investors do not. Therefore, the financing cost must reflect some average qu ality. For entrepreneurs with projects having above average quality, the financing cost may be too high to j ustify undertaking the proj ects. As a consequence, the high quality projects are withdrawn from the financing markets, and the average quality of projects seeking financing is lowered. This process may continue until the only projects seeking financing are of the lowest quality. This is the socalled the lem o n 's p ro blem or the problem of a dverse selection. For projects of good quality to be financed, the private infor mation possessed by entrepreneurs must. bP. transferred to the out side investors when ex post verifiability of the quality of projects is lacking. This can be done by observable actions taken by the en trepreneurs. One such action, observable because of disclosure rules, is the willingness of the entrepreneurs to retain the ownership of the proj ects. The larger the proportion of a proj ect retained by an entrepreneur, the less diversified his portfolio is. The cost of a non diversified portfolio is smaller for entrepreneurs having good quality projects , because they will be compensated by, for example, higher expected returns on their projects. Outside investors can then infer the quality of the project from the proportion of a project retained by an entrepreneur. From the entrepreneurs' perspective, they use their actions to signal the outside investors about the qualities of their projects . O utside investors can announce a schedule of financing cost de pending on the proportion of the project retained by an entrepreneu r. Through choosing financing from the schedule, an entrepreneur will reveal his or her private information about the proj ect quality. This is called scre e ning by the outside investors. Cle arly, signalling and screening are two sides of a coin. In the subsequent sections, we will develop a simple model of financial markets in which entrepreneurs seeking financing possess private information about the quality of their projects . A sign alling
Fin ancial Markets with Differential Information
285
equilibrium will be established . This equilibrium differs in important ways from models without asymmetric information.
9 . 1 6 . Consider a family of investment projects indexed by "quality" , p E [p, jl j . All investments require a capital outlay of K at time 0 An investment project with quality J.J has a time1 cash flow X = p + €, where p is the expected future cash flow and l is a normally distributed random variable with mean zero and vari ance u;. Each investment p roject is accessible to an entrep reneur, who plans to hold a proportion of the equity, raising the remainder of the equity from outside investors, once he decides to undertake the project. A project will be 100% equity financed. An entrepreneur has private information and knows the true value of p. O utside investors do not know the true value of J.J. They, however, are informed that p has a strictly positive density function on �. p ] . The ex post realization of X, observable to both an en trepreneur and the outside investors, does not allow outside investors to tell the value of p with certainty. However, outside investors will respond to a signal sent by an entrepreneur about the true value of p, if they believe that it is in an entrepreneur's best interest to send such a signal. The signal we will examine here is the proportion a of the equity retained by an entrepreneur. O utside investors believe that there exists a functional relation between J.J and a and make inferences about p from the observed a. A signalling equilibrium is a situation where outside investors' beliefs about the functional rela tion between a and p are indeed correct. A signalling equilibrium is said to be a separating equilibrium if the functional relation between a and p is strictly monotone. In such an event, the true J.J will be learned in equilibri um. We will only be interested in separating equi libria and will therefore use the term signalling equilibria to refer to separating equilibria. S uppose that outside investors infer the value of J.J according to the schedule g ( a } , in that when a is observed, they conclude that p = g ( a ) . We assume that the value of equity is determined by the CAPM relation: .
.
V (a) =
g (a) 1
+ TJ
,\ ,
(9. 16 . 1}
286
where
Foundations for Financial Economics , = A 

_
C ov (X, rm)
(E[T m l, 2 um C ov (l, Tm )  ] (E[Tm 2 um
_
_
Tf )
Tf ) '
(9. 16.2)
where Tm denotes the normally distributed rate of return on the market p ortfolio, and u! is the variance of Tm . We assume that· JL  ..\   K 1 + ,,
0. Relation (9. 1 7 . 5 ) must be satisfi e d by a , given that outside investors infer I' using the functional g (a) when a is obs�rved . The second order necessary conditions are
 2 g' (a) + ( 1  a ) g" (a)  a; � 0 ,  a ! � 0, a ;. (2 g1 ( a )  ( 1  a ) g" ( a ) + a; )  ( C ov(£, rm ) ) 2 ?:: 0 ,
(9 . 1 7 .6)
288
Fo un dations for Fin ancial Econ omics
where g1 and g11 denote the first and second derivatives of g, respec tively. Now we turn to analyze the problem of the outside investors.
9 . 1 § . As mentioned in S ection 9.15, the outside investors' prob lem is to screen entrepreneurs. That is, they want to announce a sch edule g ( ) to determine the quality of projects seeking financi ng as a fu nction of a, the proportion retained . The schedule should have the property that an entrepreneur, knowing the value J.L , will optimally choose an a(JJ) such that ·
(9.18.1)
when faced with the schedule g(· ) . I n other words, given t h e sched ule g (·) , the optimal behavior of an entrepreneur is to tell the truth. Equi valently, entrepreneurs reveal their true typ es by selfs election . Relations (9 . 17.5) and (9 . 1 8 . 1 ) together characterize a signalling equi libriu m. Now we turn to closed form solutions o f signalling equili bria. Su bstituting (9.18 . 1 ) into (9. 17.5) gives (1  a) g 1 (a)
=
a'"'(.
( 9 . 1 8 . 2)
There exists a family of solu tions to (9. 1 8 .2) parameterized by an arbitrary const ant C:
g (a )
=
'"'( Jln (1  a) + a) + C.
(9.18 .3)
Any member of ( 9 . 1 8 .3) is a candidate for an equilibrium schedule and makes the program of (9. 1 7 .2) a strictly concave program, for which the second order conditions are automatically satisfied. Thus the first order conditions are also sufficient for a unique optimum i n this case. Note th at every member of (9 . 1 8.3) i s strictly increasing and strictly convex in ( 0, 1) . As a approaches 1 , g (a) asymptotically approaches infinity: lim g ( a ) = + oo. a. . 1
If a member of (9. 18. 3 ) is a signalling equilibrium, its relevant domain may be a subinterval of [0, 1 ) . To see this, we first note
289
Fin ancial Markets with Differential Informat.ion
g(al
1
A " .,... ,.  
o
Figure
/
I I I I
I
9.18 .1:
a
.....
Solutions to (9 .18 .2)
that an entrepreneur with Jl < A + (1 + r 1) K will not undertake the project and seek outside financing, as, in equilibrium, the proj ect has a strictly negative net present value . The project that just breaks even in equilibrium has a quality Jl = A + (1 + r1) K. Thus the relevant domain of an equilibrium schedule is (9. 1 8 .4) the
inverse image of the quali ties of projects that have a positive net present value. Three members of (9. 1 8 .3) are depicted in Fig ure 9 . 1 8 . 1 . According to (9.1 8 .4) , the relevant domains of A and A1 are [0, 1) , while the relevant d omain of Au is (g, 1 ) , where g is the solution to "f ( ln ( 1  a:) + a:) + c" = A + ( 1 + TJ) K ,
where d' is the constant that corresponds to A11• Not every member of (9 . 1 8.3) , in its relevant domain, is an op timal solution to the outside investors' problem, however. Consider the schedule A1 in Figure 9 . 1 8 . 1 . The const ant C' associated with this schedule is strictly greater than ,\ + (1 + r1 )K. According to this
290
Foundations for Financial Economics
schedule, if an entre p reneur retains no equity, outside investors will conclude that 1L = g ( D ) = C' > A + (1 + r) K.
Thus , an entrepreneur with 1L < C' will find it profitable to undertake his project by investing K and choosing a = 0 to sell the project for
V (O)
>. g (O) 1 + TJ >. + (1 + r J )K  >. = K. > 1 + TJ
=
By doing so, he makes a strictly positive profit of V (D) K. Thus, all the entrepreneurs with a 1L < C' will undertake the project and choose a signal level of a = 0. This violates (9. 18 . 1 ) , and outside investors incur losses. (Recall that, by (9.16.3), there exists a strictly p ositive proportion of entrepreneurs having 1L < >. + ( 1  T J ) K . ) In Exercise 9 . 4, the reader will b e asked to show that the problem associated with schedule A' cannot be avoided by requiring a min imum level of strictly positive retained equity, since entrepreneurs with a 1L smaller than but arbitrarily close to the minimum level will find it profitable to undertake the proj ect and retain the minimum level. This violates (9. 1 8 . 1 ) . Thus we must have 
c ::::: >. +
(1 + T J ) K.
( 9.18 . 5)
Next, consider the schedule A in Figure 9 . 1 8 . 1 . The constant C that corresponds to A is equal to >. + (1 + rJ ) K. S imilar problems arise. An entrepreneur h aving JL < >. + (1 + r J) K can undertake the project and retain no equity. He j ust breaks e ven in this transaction and is indifferent as to whether or not to undertake the project. This again v iolates (9. 1 8 . 1 ) at a = 0. In Exercise 9.4, the reader is also asked to show that, if we require a > 0, then entrepreneurs with JL < ..\ + (1 + r f )K · will find it suboptimal to undertake the project. Hence schedule A on the domain (0, 1) satisfies (9. 1 8 . 1 ) and is a sign a lling equilib riu m.
Next, we consider schedule A " in Figure 9 . 1 8 . 1 , which corre sponds to a constant C" < ..\ + (1 + rJ) K. The kind of problem associated with schedules A and A' does not arise. One can show
Financial Markets with Differential Information
291
that entrepreneurs with p. < A + (1 + T J ) K will find it unprofitable to undertake their proj ects and ret ain a strictly positive proportion of the equity. Thus A11 with a domain [g_, 1) is a signalling equilibrium schedule . The same arguments apply to any member of the family (9. 18.3) with a constant strictly less than A+ ( 1 + r 1 ) K on its relevant domain. Hence there exists a continuum of signalling equilibria. 9 . 1 9. We identified a continuum of signalling equilibria in Sec
tion 9 . 1 8 . In each equilibrium, an entrepreneur who chooses to un dertake his project will get the desired financing by revealing the true quality of the project. However, all equilibria but one are inefficient in two senses. First, we note that the prop ortion of equity retained c an b e viewed as an indication o f the signalling cost. To s e e this, w e note that, for a project having positive net present value, if there were no information asymmetry, the first order conditions of an entrepreneur's problem would be
 au ;
 aCov ( £, im) = 0, E[rm]  T J  au!�  aCov (l,i"m) = 0 . A
(9.19.1)
These imply that a = 0. That is, without information asymme try, the optimal solution involves no retained equity. In signalling equilibria where an entrepreneur chooses to retain a strictly posi tive proportion, his expected utility will be strictly lower than what would be attained without information asymmetry. Moreover, for a fixed p. , it is easily seen that the expected utility attained for a low a is higher than that which can be attained for a higher a . As a consequence, the proportion retained is a proxy for the cost paid for obtaining outside financing. For a fixed proportion, a, the sig nalling cost to an entrepreneur with a high I' is less than that to another with a low I' · This is the reason an entrepreneur with a high I' will optimally choose a high signal  he or she can b etter afford it! Among the continuum of signalling equilibria, the one that has the least signalling cost is schedule A in Figure 9 . 1 8 . 1 . Compare schedules A and A11 , for example. For a given level of p, schedule A" requires a strictly higher retained proportion. Thus A is the most efficient in the sense that the signalling costs in equilibrium are the
292
Foundations for Financial Econ omics
lowest. In fac t, schedule A is the only equilibrium which is competi tive in th at if schedule A were offered by some investors while some others were offering a different schedule, all entrepreneurs would do business with the investors offering schedule A. Schedule A is the only equilibrium schedule under which all projects with a strictly positive net present value are undertaken. To see this, we note that an entrepreneur with a JJ which yields a zero net present value is indifferent between undertaking his project and r emaining at his status quo when facing schedule A. He will find it suboptimal to undertake his project if he has to retain a strictly p ositive proportion. On the other hand, it can be shown that an en trepreneur with a JJ that yields a strictly positive net present value fi nds it optimal to undertake his proj ect when faced with schedule A. Thus all proj ects with a strictly positive net present value will be undertaken. On the other hand , when facing the schedule A", an en trepreneur with a JJ equal to A + (1 + r 1 )K will find it suboptimal to under take his proj ect , because he has to retain a proportion g. This is so, because he is j ust ind ifferent between accepting the proj ect and remaining at his status quo when no proportion must be retained. When a strictly positive proportion must be retained, he will cer t ainly find it suboptimal to undertake his proj ect . By a continuity argument, one can show that an entrepreneur with a JJ strictly above but close to A + (1 + 'I )K will also fi nd it suboptimal to undertake his project. Thus, in equilibrium, some projects having a strictly p ositive net present value will not be undertaken . This is clearly inefficient relative to schedule A. We h ave thus identified a unique efficient signalling equilibrium. Although efficient among all the signalling equilibria, it results in an expected utility loss for the entrepreneur compared to the case without information asymmetry. 9 . 2 0 . In the previous discussions of this chapter, we only touched upon very limited aspects of informational issues related to fi nancial markets. Readers interested in the general area of information eco nomics as applied to financial markets will find Bhattacharya ( 1987) useful reading. Among the subj ects not discussed previously, the m o ral hazard
Financial Markets with Differential Information
29 3
problem is perhaps the most important. Moral hazard usually arises in a gaming situation where some players involved can take an un observable action that affects the payoffs to be shared . A classical moral hazard problem arises in the context of insurance. The prob ability of, say, a fire can be infl u enced by the care exerted by the insured. There is no incentive for an individual who is fully insured to exert any care. Thus the insurance premium should vary for differ ent JeveJs of coverage. Arrow ( 1 970) contains the original discussion of moral hazard issues in the context of health insurance. In the entrepreneurs/outside investors example discussed in Sec tions 9.16 to 9.19, there is no moral hazard problem, as the quality of a project is outside an entrepreneur's control. It is, however, more reasonable to think that the quality of a project depends on some effort on the part of an entrepreneur. For example , we can think of p. as a function of an action taken by an entrepreneur. When the action taken by an entrepreneur is not observable and not ex p ost verifiable , the problem of moral hazard arises. The outside investors' problem is then not only to screen entrepreneurs who have exerted different levels of efforts but also to design a compensation scheme to provide the right incentives. Spence and Zeckhauser (197 1 ) , Holm strom ( 1979) , and Mirrlees ( 1976) are early discussions of general moral hazard/incentive problems, to which we refer interested read ers. Recent discussio1is on this subj ect in the context of financial markets are Allen ( 1 985) and Bhattacharya and Pfl eiderer ( 1985) .
Exercises
9 . 1 . Consider a complete markets twoperiod economy, where trad ing occurs at time 0, and the consumption of a single consump tion good occurs at time 1 . Show that a necessary characteris
tic for an equilibrium price system when individuals learn from observing the system is that it symmetrize the differences in information among individuals.
294
Foundations for Financial Economics
9 . 2 . Consider an economy identical to the artificial economy of Sec tion 9.9 except that individuals observe the joint signal (Y1 1 b efore trading. Show that (9.9.2) also gives the competitive equilibrium price for the risky asset. 9.3. Consider an economy identical to the one constructed in Section 9 . 11 except that the coefficients of absolute risk aversion are not equal across individuals. Conj ecture that the price functional for the risky asset is •
•
.
, Y1)
I
s, (Yl , . . . , YJ , Z) = 1 + L b;f; + boZ. (bo,
i=l
S how that i f . . . , bJ) is a solution to a system of nonlinear equation in as many unknowns, then there exists a rational ex pectations equilibrium with the conj ectured price functional. 9.4. Show that the problem associated with schedule A ' discussed in Section 9.18 cannot be avoided by requiring a minimum level of retained equity Q > 0. Also show that the similar problem associated with schedule A in Figure 9.18. 1 disappears when we require that a > 0 . The example of nonexistence i n Section 9 . 5 i s due to Kreps (1977 ) . The competitive equilibrium of Sections 9.6 and 9. 7 i s taken from Lintner ( 1969 ) . Discussions i n Sections 9.8 to 9.10 are freely borrowed from G rossman ( 1976) and G rossman and Stiglitz ( 1980) . The closed form solution of the noisy rational expectations equilibrium of Sections 9 . 1 1 to 9. 1 3 is taken from Diamond and Ver recchia ( 198 1 ) . The signalling model of Sections 9 . 1 5 to 9. 19 is due to Leland and Pyle (1977 ) . In the same context as Exercise 9 . 3 , Hellwig ( 1980) shows the existence of a noisy rational expectations equilibrium. He obtains a closed form solution when the number of individuals in the economy increases to infinity. Our discussion in this chapter on some applications of informa tion e conomics to financial markets is very limited in scope. Readers interested in rational expectations equilibrium should consult Jordon and Radner (1982) and the references therein. Admati ( 1987) has Remarks.
Fin ancial Markets with Differential Information
295
a more recent summary account of the literature including applied work . Pfleiderer ( 1984) studies comparative statics on volumes of trade in a rational expectations model. Kyle ( 1 985) uses monopolis tic competition to model speculation and information transmission through prices. Signalling models began with Akerlof ( 1970) and Spence ( 1973) . Riley ( 1975) and Rothschild and Stiglitz ( 1975) also made important contributions. For a more recent discussion on signalling equilibria, see Cho and Kreps ( 1987) . Early applications of signalling models to financial economics include Bhattacharya (1976, 1980) � nd Ross ( 1977 ) . Bhattacharya and Ritter (1983) is the first to consider models where a signal to financial markets is also observed by competitors in the product markets. Gertne r , Gibbons, and Scharfstein ( 1987) also model this situation. Readers interested in this class of models as well as applic a t i o n s of other ideas in information economics to fi n an cial markets should consult Bhattacharya ( 1987) and the references therein. References
Akerlof, G. 1970. The market for "lemons." Q uarterly Journal of Economics 89, 488500. Admati, A. 198 5 . A noisy rational expectations equilibrium for multiasset securities markets. Econometrica 53, 629657. Admat i , A . 1987. Information in financial markets: The rational expectations approach. In Frontiers of Financial Theory. Edited by G . Constantinides and S. Bhattacharya. Rowman and Little field. Totowa, New Jersey. Allen, F . 1985. Contracts to sell information. Mimeo. Wharton School , University of Pennsylvania. Arrow, K. 1970. Essays in the Theory of RiskBearing. North Holland. Amsterdam. Bhattacharya, S . 1976. Imperfect information, dividend policy, and the "bird in the hand" fallacy. Bell Journal of Economics 10:259270.
296
Foundations for Financial Economics
Bhat t acharya, S. 1 980. Nondissipative signalling structures and div idend policy. Q uarterly Journ al of Economics 9 5 : 124. Bhattacharya, S. 1987. Financial markets and incomplete informa tion : A review of some recent developments. In Frontiers of Financial Theory. Edited by G . Constantinides and S . Bhat tacharya. Rowman and Littlefield. Totowa, New Jersey. Bhattacharya, S . , and P. Pfleiderer. 1985. Delegated portfolio man agement . Journal of Economic Th eory 36: 125 . Bhattacharya, S . , and J . Ritter. 1983 . Innovation and communi cation: Signalling with partial disclosure, Review of Economic St udies 5 0 : 3 31346 . Cho, 1.K . , and D . Kreps. 1987. Signalling games and stable equi libria. Quarterly Journal of Economics 5 2 : 179221 . Diamond, D., and R. Verrecchia. 1981. Information aggregation in a noisy rational expectations economy. Journal of Financial Economics 9, 2 2 12 35. G ertne r , R., R. G ibbons, and D . Scharfstein. 1 9 8 7 . Simultaneous signalling to the capital and product markets. Mimeo. Sloan School of Management, Massachusetts Institute of Technology. Cambridge, Massachusetts . Grossman, S. 1976. On the efficiency of competitive stock markets when agents have diverse information. Journal of Finance 3 1 , 573585. Grossman, S., and J . Stiglitz. 1980. O n t h e impossibility of in formationally efficient markets. American Economic Review 70, 393408 . Hellwig, M. 1 9 8 0 . O n the aggregation of information in competitive markets. Journal of Economic Theory 2 2 , 477498. Holmstrom, B. 1979. Moral hazard and observability. Bell Journ al of Economics 10:749 1 . Jordon, J . , and R . Radner. 198 2 . Rational expectations i n microe conomic models: An overview. Journ al of Economic Theory 2 6 , 20 122 3 . Kreps, D . 1977. A note on fulfilled expectations equilibria. Journ al of Economic Theory 44, 3243. Kyl e , A. 1985. Informed speculation with imperfect competition. Working Paper, Woodrow Wilson School, Princeton University.
Financial Markets with Differential Information
297
Princeton, New Jersey. Leland, H . , and D. Pyle. 1 977. Informational asymmetries, fi nan cial structure, and financial intermed iation. Journal of Finance 3 2 :37 1387. Lintner, J. 1 969 . The aggregation of investors' diverse judgement and preferences in purely competitive security markets. Journal of Financial and Quantitative Analysis 4 : 103124. Mirrlees, J . 1976. The optimal structure of incentives and authority within an organization. Bell Journal of Economics 7: 1051 3 1 . Pfleiderer, P. 1984. Private information , price variability and trading volume. Mimeo . G radu ate School of Business, Stanford Univer sity. Stanford, California. Riley, J. 1 975. Competitive signalling. Journal of Economic Theory 1 0 : 1 74186. Ross, S. 1977. The determination of financial structure: The incentive signalling approach. Bell Journal of Economics 8 :2340. Rothschild , M . , and J. Stiglitz. 1975. Equilibrium in competitive insurance markets: An essay on the economics of imperfect in formation. Quarterly Journal of Economics 9 0 : 8 1 2824. Spence, M . 1973. Job market signalling . Quarterly Journal of Eco nomics 8 7, 355379. Spence, M . , and R. Zeckhauser. 1 9 7 1 . Insurance, information , and individual action. American Economic Re view 6 1 :38G387.
CHAPTER 10 E C O N O METRIC I S S UE S I N TESTING THE C APITAL ASSET P RI C I N G M O D E L
1 0 . 1 . I n this chapter, econometric issues related t o testing the Capital Asset Pricing Model { CAPM) will be discussed. Many test statistics will be given interpretations in the framework of Chapter 3. To provide a conceptual basis for interpreting these econometric is sues, we shall first briefly discuss the testable implications of the CAPM. Throughout our discussion, we will assume that there exists a riskless lending opportunity.
10 . 2 . In Chapter 3, we examined the mathematics of the port folio frontier. There we derived the first order condition necessary and sufficient for a portfolio to be on the frontier and proved that any p ortfolio that is a linear combination of frontier portfolios is it self a frontier portfolio. In Chapter 4, we provided conditions on the distributions of asset returns necessary and sufficient for individ uals' optimal portfolios to be frontier portfolios. In such an event,
299
300
Foundations for Financial Economics
lio is the marke t dearin g condit ion implie s that the marke t portfo thus nd is a lios portfo al optim a conve x combi nation of individ uals' the once diately a fronti er portfo lio. The CAPM then follow s imme us, mar ket portfolio is identified to be on the portfo lio frontie r. 1� ns the CAPM is merely the combi nation of the first order cond1t g cleann t marke the and r for a portfolio to be on the portfolio frontie condition.
!�
1 0 . 3 . Using timeseries sample means, variances, and covari ances we can construct a portfolio frontier, referred to as an ex post portfolio frontier. A p ortfolio on an ex post portfolio frontier is an ex post frontier portfolio. An ex post frontier portfolio could be con structed even if means, variances, and covariances were generated randomly. If the betas on individual assets are measured relative to an ex post frontier p ortfolio, it follows from the mathematics of the p ortfolio frontier that the average realized rates of returns on these assets would have an exact linear relation to these betas. This is a mathematical fact and has nothing to do with an equilibrium pricing relation. Thus, it is tautological to say that there exists a single "factor" that will "explain" rates of return. In addition, such a frontier portfolio may have large negative weights for many assets . Thus , without an underlying valuation theory, we cannot predict, a priori , that there is a positive association between average realized rates of return and betas measured relative to a prespecified market proxy portfolio that is welldiversified and has positive weights for all assets. This observation also relates to an important conceptual problem in testing the Arbitrage Pricing Theory that does not pre identify the "factors" based on economic models of portfolio choice. For example, if factor analysis is used to determine the set of fac tors "explaining" ninetyfive percent of the variation of the rates of returns on individual assets, a linear combination of these factors may closely approximate the rate of return on an ex post frontier portfolio. A test of the proposition that these factors are the sole determinants of risk premiums might fail to be rejected even if the underlying returns were generated randomly.
Econometric Issues in Testing the CAPM
301
1 0 . 4 . Financial theories provide internally consistent models of asset prices that have testable implications. A positive theory of the valuation of risky assets should not be judged by the realism of its assumptions. Indeed, incorrect assumptions are sometimes necessary to abstract from the complex and detailed circumstances and to build a model that focuses on more important aspects. For example, although we are well aware of the fact that many individuals have different probability beliefs, we abstract from this consideration by assuming homogeneous probability b eliefs . This assumption and other equally implausible assumptions permit the derivation of the CAPM. A long and detailed list of realistic assumptions that are impossible to model is merely an institutional description and has in itself no predictive value. The correct test of a p ositive theory of asset pricing is the accuracy of its predictions concerning security returns and / or security prices. If the assumptions used to derive a theory abstract from the most critical considerations, then the theory's predictions would very likely prove inaccurate. The more general are the assumptions necessary to derive a given theory, the less precise its predictions will be. For example, allowing for the possibility that borrowing is either prohibited or is done at a rate strictly higher than the lending rate is more general than assuming unconstrained b orrowing, or equivalently, unlimited borrowing and lending at the same rate. Recall from Section 4 . 1 4 that the CAPM based on unconstrained b orrowing predicts a proportional relation b etween risk premiums ( with respect to the riskless rate) and betas . The more general model that allows for constraints on borrowing merely predicts a linear relation between risk premiums and betas with a positive intercept.
1 0 . 5 . Positive theories have strong predictions and weak pre dictions. A strong prediction is a prediction whose validity is implied by and implies the underlying theory. Thus , strong predictions are equivalent to necessary and sufficient conditions for the underlying theory. A strong prediction of the CAPM is two fund separation. Two fund separation may be refuted by the finding that an individ ual's optimal portfolio is not spanned by the optimal portfolios of two other individuals. While to the best of our knowledge no researcher
302
Foundations for Financial Economics
t akes such a prediction seriously, another strong prediction that the market portfolio is on the portfolio frontier has been subjected to extensive testing. Since meaningful positive theories are often based on unrealistic assumptions, their strong predictions are unlikely to be perfectly accurate. Indeed, failure to reject statistically a strong prediction of a positive theory is usually due to a lack of power of the statistical test. Strong predictions of positive theories, like two fund separation, are not often examined b ecause their rejection is usually obvious a priori . In contrast, a weak prediction is a prediction whose validity is "broadly" implied by but does not imply the underlying theory. An example of a weak prediction of the CAPM is that ex post betas measured relative to a broadly based market index are positively related to the average ex post realized returns. This weak prediction does not imply a.n exact linear relation b etween ex ante expected rates of return and betas. For example, ex ante expected returns may be related to betas and a second variable that is independent of the betas. Under stationarity conditions, ex post betas would then be related to ex post average returns. Empirically verified weak predictions have, however, yielded useful applications in financial economics . 1 0 . 6 . In its most general form , the CAPM implies the following crosssectional relation between ex ante risk premiums and b etas, when there exists a riskless asset:
(10.6.1) where r; is the excess rate of return on security j , r m is the excess rate of return on the market portfolio of all assets, rz�(m ) is the excess rate of return on the frontier portfolio having a zero covariance with respect to the market portfolio, f3;m is Cov (r; , rm ) /Var ( rm ) , the "beta" of security j (with respect to the market portfolio) , and E[·] is the expectation operator. Here we caution the reader to note that, for notational simplicity, we have used r; to denote the "excess" rate of return on asset j with respect to the riskless lending rate in contrast to the notation used in earlier chapters, and likewise for rm
Econometric Issues in Testing the CAPM
303
and izc( m ) · In empirical work , one often uses the treasury bill rate as the short term riskless lending rate. 1 0 . 7 . When there are unlimited borrowing and lending oppor tunities at a constant riskless interest rate, the traditional version of the CAPM predicts that the expected rate of return on z c ( m ) is equal to the riskless rate, or, equivalently, that risk premiums on assets are proportional to their market betas. In this case, { 10.6.1) holds with E[i.,c(mJ ] = 0 and E[im] > 0. (10.7.1) The constrained borrowing version of the CAPM makes the less pre cise prediction that the risk premium on zc ( m ) is positive and that the difference between the risk premium on the market p ortfolio and the risk premium on zc ( m) is strictly positive. In this case, (10.6.1) holds with
( 10.7 . 2) Thus, the predictions of the constrained borrowing version of the CAPM are somewhat less precise than those of the traditional version of the model. It should, therefore , come as no surprise that the empirical tests are more consistent with the constrained borrowing version of the model. There are, in general, three major conceptual problems associated with the testing of the CAPM. First, the CAPM implies relationshi ps concerning ez ante risk premiums and betas, which are not directly observable . Second, empirical tests use timeseries data to calculate mean excess rates of return and b etas·, however , it is unlikely, that risk premiums and betas on individual assets are stationary over time. When timeseries data are used to calculate betas and mean rates of return on assets, it is implicitly assumed that the CAPM holds period by period, since the CAPM is a two period model. Third, many assets are not marketable and tests of the CAPM are invariably based on proxies for the market portfolio that exclude major classes of assets such as human capital (the capitalized value of wage and salary income), private businesses, and private real 10 . 8 .
304
Foundations for Financial Economics
e state.
Different approaches for dealing with these problems have
b een taken in the literature. 10.9.
Concerning the first conceptual problem, the unobserv
ability of ex ante expected returns and betas, the assumption of ra tional expectations is implicitly made . Under ratio nal exp ectations,
the realized rates of return o n assets i n a given time period are draw ings from the ex ante probability distributions of returns on those assets.
Here we remind the reader that the defi nition of ratio nal
expec tations can b e found in Section 10 .10.
7.7.
A s for the second problem, the nonstationarity of risk
premiums and betas on individual assets, two complementary ap
proaches h ave evolvf!d. The first. approach is t.o form portfolios that, are constructed to have stationary betas and to assume that the
risk premiums on these constructed portfolios, on the market port folio, and on
( )
zc m
are stationary over time.
Note that constant
risk premiuros are consistent with equal p ercentage point changes in the riskless interest rate and in the expected rates of return on risky assets. The second approach is to interpret the tests in terms of the distributions of asset returns conditional on a coarser inform ation set and assume that these distribution are timestationary. Even when risk premiums and betas conditional on information sets available to investors over time are nonstationary, they can be stationary con ditional on a coarser information set.
We will discuss this second
approach in detail be low.
As mentioned in Section
10.8,
empirical testing of the CAPM
implicitly assumes that the CAPM holds period by period. In par ticular, we assume that, at each time
t  1,
an individual i maximizes
exp ected utility of time t random wealth conditional on the informa
1.
The first order condition for his time t
tion that he has at time
t
random wealth,
be optimal is
wt, to

( 10. 10. 1 ) where
7;t
uit (·)
denotes individual i's utility function for time t wealth,
denotes the rate of return on asset j from time t

1
to time t,
Econometric Iss ues in Testing the CAPM
305
fzc( m ) t denotes the rate of return on t h e zero covari ance portfoli o with respect to the market portfolio from time t  l to time t� and :Ft l denotes the i nformation p ossessed by individuals at time t  1 . Taking exp ectations o n both sides of
aggregate wealth at time
where
t 
1
gi ves
(10.10. 1 )
condi tional o n the
( 1 0 . 1 0. 2 )
M,_t deno t es the tim e t  1 aggregate wealth. Assume that distri W:, �" and f"''"'' are multivariate normal cond i tio nal o n M,_,.
butions of
Using the defi nition of covar iance, the Stein's Lemma, and the line of uncorrelated conditional on
l?�  1 (E [fjt 1 Mt  d  E[fzc(m)t i .�ft  d ) where
19; =
M, and
Mt] we can rewrite (10.10.2) as
argument used i n deriving (4 . 1 5.3), and assuming that
=
C ov (
rzrrmJt are
�V/ , Tjt  Tzc( m ) t i ll1t  I ) ,
( 1 0 . 1 0.3)
 E [ui� (vVi )liY.ft 1] E[u� t (vVi ) I A1t 1]
is t h e ith i ndivi dual ' s global absolute risk aversion for t h e t i m e t utility fun ction on wealth conditional on time t 
S ummi n g
( 10.10.3)
i=l
gi ves
where
I
L: wf
=
1
total wealth.
across i and using the market cleari ng condition =
M"t
E [rit ! Mt 1 ]  E [f zc(m )t l .iVt  d 1 Cov( fit, A1t l Mt  1 ) ,. t •=1
(t ; )
Vj
( 1 0 . 1 0 .4)
306
Foundations for Financial Economics
is the rate of return on the market portfolio from time t  1 to t. Relation (10.10.4) certainly holds when we take asset j to b e the market p ortfolio and thus
(10 . 10 . 5)
Vj.
Substituting ( 10 .10.5) into ( 10 . 10.4) gives E [i;t i Mt 1 ]  E[i•c(m) ti Mt  1 ] = .B;mt (Mt t ) ( E[imt i Mt 1 ]  E[r•c(m)t i Mt 1 ] ) ,
Vj,
(10 . 10.6)
where is the time t beta of assets j condit ional on Mt 1 · Note that in t ( 10 . 10 .6) the expect ed return s and betas are measu:_ed with respec the When · M on t 1 to the distribution of asset return s condit ional ate wealth are aggreg on onal conditi s return asset of distrib utions indepe ndent of aggreg ate wealth , (10.10. 6) becom es E[i;t+l]  E[in(m)t+ l] = .B;mt ( E [ imt+ l]  E [r•c(m)t+ 1 ] ) , where .B; mt
=
Vj, (10 . 10.7)
Cov (i;t+ l , imt) " v ar ( rmt )
is the unconditional beta for asset j. Note that (10. 10.7) is specified under the unconditional distribution and is thus termed the uncon dition al GAP M. If the unconditional distributions of asset returns are stationary over time, .B;mt of ( 1 0 .10.7) is a constant independent of time. It then follows that even though risk premiums and betas conditional on the information set available to investors may be non stationary, a timeseries of returns on assets may be used to test the unconditional CAPM. Also note that (10.10.2) does not imply (10 .10 . 1) . This is b e cause a zero conditional expectation conditional on an information
Econometric Issues in Testing the
307
CAPM
partition does not imply that the expectation conditional on a finer information partition will be identically zero. Along the lines of argu ment used in deriving ( 10. 10.6), (10.10.1) together with multivariate normality of the rates of return on assets conditional on lt  1 implies that E[i;t l 1t  t ]  E[rzc(m)t l 1t  t] = .Bjmt (E frmt l 1tt]  E[izc(m)t i Jt1] ),
Vj,
(10. 10.8)
where the random variable
is the conditional beta of asset j at time t. Relation (10.10.8) will be termed the conditional CAPM. The validi ty of the uncond i t ional CAPM does not imply the validity of the conditional CAPM. Thus, the market portfolio being on the portfolio frontier based on the unconditional distributions of asset returns should be viewed as a weak prediction of the conditional form of the CAPM but a strong prediction of the unconditional form of the CAPM. Most existing tests of the CAPM and our discussion which follows focus on the unconditional form of the model. 1 0 . 1 1 . Concerning the third conceptual problem, the unobserv ability of the true market portfolio, three related approaches have been taken. The first approach ignores the problem by implicitly as suming that the disturbance terms from regressing the asset returns on the return on the market proxy portfolio are uncorrelated with the true market portfolio and that the proxy portfolio has a unit beta. If the market proxy is a portfolio constructed from the individual assets or portfolios contained in the test sample, this assumption is equivalent to assuming that the market proxy is the minimum vari ance unit beta portfolio of the set of all feasible portfolios constructed from the assets in the test sample. This implicit assumption will be discussed in more detail in Section 10.12. The second approach merely interprets the test as a test of whether the market proxy is on the portfolio frontier. The third
308
Foundations for Financial Economics
�
�
AP�
with ac t:o r appr oach is to view the test as a test of a . s ngle . tmgm shabl e a presp ecifie d facto r. This appro ach is empm cally md1s from the secon d appr oach . 1 0 . 1 2 . Now we will demonstrate explicitly the implicit assump tion made when the unobservability of the true market portfolio is ignored . Suppose that the market proxy, m, has unit beta and that the disturbance terms in the regressions
r; = a; + fJ; im. + e; ,
i
=
1 , 2, . . . , N
are uncor relate d with the true mark et, wher e Pim.
_
=
Cov(r i• r m.) Var (rm.)
(10. 1 2 .1)
(10.1 2.2 )
is the "beta" of asset j with respect to the market proxy . By d efini tion, the true beta of asset j is
fJ; m
=
Cov(r; , rm) Var(rm) '
Subst itutin g ( 10.12 .1 ) into ( 10.12 .3 ) gives
Cov(rm. , rm) fJ; m = fJ;m. Var(rm) '
( 10.12.3 )
( 10. 12.4 )
where we h ave used the hypothesis that e; is uncorrelated with rm . Since the market proxy has unit b eta, we know that
Cov (im., im) Var ( im)
Substituting this into ( 10.12.4 ) gives
= 1. (10.12.5)
that is, the b eta with respect to the market proxy is equal to that with respect to the true market portfolio. Thus, even though the true market p ortfolio is not observable, the true betas can be estimated if
Econometric Issues in Testing the CAPM
309
the market proxy has unit beta and the disturbance terms of (10. 1 2 .1) are uncorrelated with the true market portfolio. Next we show that if the proxy, m, is constructed from the in dividual assets or portfolios contained in the test sample and is the unit beta minimum variance portfolio among all the portfolios con structed from the test sample, the betas measured with respect to m are equal to the true betas. Consider a set of N risky assets which is a proper subset of the set of all assets. Let V be the uncondi tional variancecovariance matrix of their rates of return, w be the N X 1 vector of the weights of a portfolio, Bm be a N X 1 vector of betas measured relative to the true market portfolio, and Bm. be a N X 1 vector of b etas measured relative to the rate of return on the minimum variance unit beta portfolio. Let wn. be the portfolio weights on risky assets for the minimum variance unit beta po rtfol i o Since covariances are additive, the beta of m is a weighted average of betas of individual assets, that is, fJm. = wm.Bm. Then wm. is a solution to the following problem, .
min !wTVw
{w} 2
s.t. w T Bm
= 1.
(10.12.6)
The weights on risky assets in (10.12.6) are not constrained to sum to unity b ecause the portfolio weights can be later rescaled with a riskless asset without changing the beta or the variance . Assuming that V is nonsingular, wm, is the unique solution to (10.1 2.6) and satisfies the following first order conditions:
(10.12.7) and
(10. 12.8) where ..\ is the Lagrangian multiplier for the constraint of ( 10. 12.6) . Relations (10.12 .7} and (10. 12.8} imply that ..\
1
= BJ;_V1 Bm
(10. 12.9) ·
3 10
Foundations for Financial Economics
Next note that, by definition, the N x 1 vector of betas with respect to the minimum variance unit beta portfolio, denoted by Bm. , is
(10.12. 10)
where the second equality follows from (10.12.7 ) , and the third equal ity follows from (10.12. 9). Relatio n (10.12 .10) demonstrates that the betas of all assets in this proper subset of assets with respect to the minimum variance unit beta portfolio (constru cted from the same subset of assets) are identica l to their betas with respect to the true market portfoli o. Thus, even if the market portfolio is not observa ble, the betas can still be estimat ed when an appropr iate market proxy is used. In Exercis e 10.1 we ask the reader to solve (10.12.6 ) by replacin g the constrai nt with w T B m = p for some constan t p =1 0. The solution is a portfolio with a beta equal to p that has a minimu m variance . Denotin g this solution still by m, the reader is also asked to show that Bm. = Bm / p . Most tests of the CAPM have used a timeseries of monthly rates of return on common stocks listed on the New York Stock Exchange (NYSE). The CAPM suggests three related empiri cal models. The first is a crosssectional regression model involving average monthly excess rates of return, f; ( average realized rate of re turn in excess of the short term lending rate) , and betas with respect to the NYSE index, denoted by f3;m. . This model is 10.13.
r; = a + bfJ;m. + u; , r; 
E[r;], a = E[rzc(m) J , b = E[rm.]  E [rzc(m) l ·
u; =
j = 1, 2, . . . N,
(10.13.1)
The second i s a series of monthly crosssectional regressions of the realized excess rates of return on the betas. That is, for all
311
Econometric Issues in Testing the CAPM
t
= 0 , 1 , . . . , T, j = 1 , 2, . . . N, ( ii;t = r;t  E[i; lizc m)t . imt ), a t = Tzc(m)t • bt = Tmt  rzc ( m)t•
(10.13 .2)
where we have used r;t , im.t . and Tzc (m)t to denote the random excess rates of return at time t on security j, the NYSE index, m, and the minimum variance zero covariance portfolio with respect to the NYSE index, respectively. The third is a seri�s of timeseries regressions for each asset or p ortfolio in the sample: Vj = 1 , 2, . . . , N , r;e
=
ai
+ f3;m.im.t + ejt
a; = E[rzc ( m ) ) (1  fJ;,n ) , i; t = r;t  E[i; t limt) ·
t
=
1 , 2,
. . .
T,
(10. 13.3)
In this model, the expected excess rate of return on zc ( m) and the betas are assumed to be constant over time. Note that the {J;,n's are treated as a fixed independent vari able in the crosssectional regression models given in ( 10. 1 3 . 1 ) and (10.13.2), while they are parameters to be estimated in the time series regression model given in ( 10.13.3). Note also that in the above model specifications, we have used a tilde on top of r;t , r,nt , and Uj t to signify the fact that they are random variables. When the same symbols appear without a tilde, they represent realizations of random variables or the observations in a sample. We will refer to the u; , ii;t > and i; t of (10.13. 1 ) , (10.1 3.2) , and (10.13 .3) , respectively, as disturbance terms. In later sections, when they appear without a tilde , the same symbols represent "residuals" from . a fitted linear model. 1 0 . 1 4 . An example of the type of test of (10.13.1) is Blume and Friend ( 1973) . The focus of their study was on a weak prediction of the traditional CAPM. Specifically, they tested the predictions that a = 0 and b > 0. Thus, they were testing whether E [rzc( m) J = 0 and
312
Foundations for Financial Economics
E[rm. ]  E[rzc (m ) l > 0 . They found that both a and b were strictly p ositive and this finding was statistically significant. An example of a test of (10.13 .2) is Fama and MacBeth ( 1973) . The primary focus of their study was the same weak form predictions. They tested the predictions that T
L at/T t= l
=
T
2: btfT
0 and
t=l
>
0
when return distributions are stationary over time. Thus, like Blume and Friend they were essentially testing whether E[ rzc( m.) ] = 0 and E[ rm ]  E[ rzc( th ) ] > 0. They found that the means of both O t and bt were significantly strictly positive, where at and bt denote estimates of llt and be , respectively. An example of the third type of test is Black, Jensen and Scholes ( 1972), who tested the weak prediction that N "' � .
t=l
Q • '
N (1  P•rn) J
o
.
Equivalently, they were testing whether E [rzc( th) ] = 0, which they rejected . Another example of the third type of test is Gibbons ( 1 982) , who tested whether
The Gibbons' test was a test of whether or not the market proxy p ortfolio was on the p ortfolio frontier of the assets included in his sample. The tests discussed above are closely related . Before discussing them in detail, we will first develop a general framework for analyzing the methods used by these authors . 1 0 . 1 5 . In addition to the conceptual problems associated with . testmg the CAP , there are three major econometric problems . These problems Will be briefly discussed in this and the next two
�
Econometric Issues in Testing the CAPM
313
sections and their detailed solutions will b e developed in later sec tions. The first major econometric problem is that the distmbance terms of { 1 0 . 1 3 . 1 ) and { 1 0 . 1 3 .2) are heteroscedastic and correlated across assets, b ecause variances of rates of return differ across assets and asset returns are correlated. Therefore, ordinary least squares (OLS) estimators of a and b and at and b t are inefficient (h ave higher variances) relative to generalized least square (GLS) estimators. This indicates that tests based on OLS estimators would be less powerful than tests based on GLS estimators. The lower power means that there is a higher probability of failing to reject the null hypotheses that E [rzc( m.) ] = 0 and E[ rm ]  E [r.�c(m ) ] > 0 when the alternative hypotheses that E [ fzc( m) J > 0 and E[rm]  E[r z c(m ) l :::; 0 are true. Therefore , empirical findings that reject the null hypotheses cannot be attributed to a lack of power in the testing p:rocedures. Moreover, estimates of the variances of the regression coefficients are biased if standard formula for the OLS estimators are used. Thus, values of the tstatistic given by "canned" OLS regression packages are biased. Note, however, that if the f3; m.'s are observed without error, the OLS estimators of a and b and a t and b, are unbiased. Possible solutions to the first problem are to use G LS estima tors or to use OLS estimators and t o calculate the correct variances of the coefficients. The former approach was taken by Litzenberger and Ramaswamy { 1 978) based on a restricted variance covariance matrix. The latter approach was adopted by Fama and MacBeth ( 19 72 ) , Black, Jensen and Scholes ( 1972) , and Kraus and Litzen berger { 1976) . These solutions, however, require estimation of the variancecovariance matrix. We will discuss these estimation proce dures in considerable detail in later sections. 1 0 . 1 6 . The second major econometric problem is that we do not observe the true betas but rather their estimates which contain measurement errors. Under this condition, O LS and G LS estimators of a and b and at and bt are biased and inconsjstent. There are three p ossible solutions to this problem. First, use grouped data to reduce the variances of the measurement errors in betas. Second, use an in strumental variables approach. Third, use an adjusted GLS approach
Foundations for Financial Econ omics
3 14
that t akes account
of the variances of the measurement errors in
ta.s. The first appro ach was used by Blume and Friend
be·
( 1973), Fama and Macbeth (1 972) , Black , J ensen and Scholes ( 1972) , B l ack an d Scholes (1974) , and Kraus and Litzenberger (1 976) among others. The second app roach was used by Rosenberg and Marathe (1979) .
�y
The third approach was adopted by Litzenberger and Ramaswa
(19'1'9 ) .
This problem and alternative solutions will be discussed
considerable detail later in this chapter.
m
10 . 1 7 . The third major econometric problem is that the CAPM
implies a nonlinear constraint on the return gener ating process as expressed in relation
(10.13.3).
In testing whether or not a mar
ket. proxy p ortfolio is on the p or tfolio frontier, estimation of b e tas,
variances
and covariances should take account of this constraint. the p arameters that
H owever , the constraint itself depends upon are to b e estimated.
This problem can be solved by usin g a mroci
mum likelihood estimation that talces these interactions into accou nt . This appro ach was developed by Gibbons by S t amb augh Shanken
(1985) ,
10.18.
( 1982),
(1982) and later extended (1982) , Kandel (1 9B4) ,
J obson and Korkie
and MacKinlay
(1987).
In order to fo cus o n the problems o f heteroscedasticity
and correlation of the disturbance terms , assume until further no tice that the betas are fixed known independent variables observable without error. Also assume that our test sample consists of
N assets
having linearly independent realized monthly rates of return. When the variancecovariance matrix is estimated using real ized monthly rates of r eturn, we require that the number of observations in the sample exceed the number of asse ts. variancecovariance matrix
U nder these conditions, the
is nonsin gular . Furthermore,
as
vari
ances are strictly positive even when portfolio weights do not sum to unity, the variancecovariance matrix is p ositive definite. We will consider a p o oled cross sectional timeseries regression of realized monthly excess rates of r e turn on betiiS and other independent vari
(see
ables. The other independent variables that have been e x amined in the literature include residual risk
Fama a.nd Macbeth
(1973) ) ,
Econometric
Issues
315
in Testing the CAPM
( see Black and Scholes (1972) , Litzenberger and Ra (1979, 1982), and Miller and Scholes (1982}) , log of firm size ( see B anz (1981), Reinganum (1983) and Schwert (1983) ) , and systematic skewness ( see Kraus a n d Litzenberger (1978) a n d Friend and Westerfield (1980)) . dividend y ield maswamy
1 0 . 19 .
To consider tests o f extended forms o£ t h e
CAPM
that
involve more than one independent v ariable, the econometric model is expressed
as:
r = Xb + u.,
E [ii]
=
0,
(10.19.1)
i i s a TN X 1 vector o f monthly random excess rates o £ return , rj h T is the numb er of months, N is the number o£ assets, b is a k X 1 vec tor of coeffi cients, X is a TN x k mat.rix of independent variables , ii is a. TN X 1 vector o£ disturbance terms, and 0 is a T X N vec tor of zeros. The first column o£ X is a column of 1 1 s, the where
(k1 ) th
second column contains the the
j
' fliml s, and the kth column contains
independent variable. Note that the betas are allowed
to differ each month and tha t we have used of asset
flim.t to denote the beta
with respect to the market proxy i n month t.
This is
consistent w i th many studies that use different estimates of betas f or each monthly c rosssectional regression. const ant
However, betas being
iB consistent with a station ary multi variate distribution o£
asset returns, and this case w ill be considered in d etail later.
The variancecovariance matrix of the disturbance t erms is V
:=
E[iiuT ] E [u n u u] E [u21 u u ]
E [ uu u21] E[u 2 1u21]
E [ uu fiNT ] E [u2l uNT ]
E [ uNI Uu] E[un u u]
E [ u N. u 21 ] E [u nu21]
E [uN. u NT ] E [ti n u NT]
E [uNT un ]
E [uNTU2a ]
E[ u NTu: NT]
H V is known , the GLS estimator of b is
'b(aLsJ
=
(XTv •xr •xT v •r.
(10. 1 9 . 2}
3 16
Foundations for Financial Economics
The variancecovariance matrix of the estimator b ( GLS ) is Tylx)  1 Var(b ( aLs) ) = (X 4
(10.19.3)
·
The OLS estimator of b is
b eaL s ) = (X Tx ) l xT r . A
(10 .19 .4)
The variancecovariance matrix of b ( OLS ) is
Var(b (oL s ) ) = ( X TX)  lxTvx (xTx) 1 A
·
(10. 19.5)
The G LS estimator of ( 10 . 19.2) reduces to the OLS estimator of (10.19.5) when the OLS assumption on the disturbance terms that (10.19.6)
for some a! > 0 is satisfied, where ITN is the TN X TN identity matrix. It can be verified that if (10. 19.6) is not true, then the diagonal elements of Var (b (oL s ) ) is greater than the corresponding diagonal elements of Var {b ( GLS ) ) , and thus OLS estimators are not efficient compared to GLS estimators. For expositional purposes, the analysis in most of this chapter will be done under the assumption that there are two independent variables so that b is a 3 X 1 vector and X is a T N X 3 matrix whose second and third column are vectors of the two independent variables. The second independent variable will henceforth be referred to as the dividend yield, for convenience, and the dividend yield for security j at time t will be denoted by d;t . In this case, we can write
Our discussion will not be changed if the second independent variable is residual risk , the log of firm size, or systematic skewness. Note that in practice the true variancecovariance matrix is un lmown and its elements are estimated using realized excess rates of
317
Econometric Issues in Testing the CAPM
return. In small samples, GLS estimators based on a n estimated variancecovariance matrix are not necessarily more efficient than OLS estimators. However, as the number of months in the sample increases (T + oo ) , the estimates of the variances and covariances based on timeseries data approach the true variances and covari ances. Thus, GLS estimators based on timeseries estimates of vari ances and covariances are said to be asymptotically efficient. Note that asymptotic efficiency is achieved as the number of time periods approaches infinity, not as the number of assets approaches infinity. As the number of months in the sample increases to infinity, the sample distribution of asset returns approaches the true underlying distribution. 1 0 . 2 0 . We will assume that excess ra.tes of return are serially uncorrelated, which is consistent with most empirical worl< . There fore, V is a block diagonal matrix, and its inverse, v 1 , may b e expressed as
v 1 1
v1 =
0
0
0
y 1
t
0
0
0
1 vT
where v; 1 denotes the inverse of the period t variancecovariance matrix of asset returns. The element in the jth row and kth col umn of Vt is the covariance of the excess rates of return on the jth and kth asset at time t. Because V is block diagonal, separate GLS regressions may be run crosssectionally using data in each period, and the pooled timeseries crosssectional GLS estimates that are b ased on a combined sample of monthly observations for all months and on model ( 1 0 . 1 9 . 1 ) are weighted averages of the individual p e riod estimates. The weights are proportional to the variances of the individual period estimates. Formally, consi der the period t submodel of the two indepen dent variable version of (1 0.19 . 1 ) : (10.20.1)
318
Foun dations for Financial Economics
where
The GLS estimator
bA t(G LS)
of bt , denoted by bt(GLS) , is =
T (x tT vt 1xt )  1xt vt1 rt .
(10'. 20.2)
The variancecovariance matrix o f the estimator is
(10.20 .3) Note that, similar to the discussion in Section 10.19, when Vt = u�,IN for some u� , > 0, where IN is the N X N identity matrix, the G LS estimator of (10.20.2) reduces to an OLS estimator. Note also
that when Vt is stationary over time, its elements can be estimated by the timeseries variances and covariances of the excess rates of return on �sets. Denoting by bk t( GLS) the kth element of b t (GLS) and by bk (GLS) the kth element of b(GLS) , it is easily verified that T
bk (GLS) = L Wtbk t(GLS) > t==1 where
(10.20.4)
1
(10.20.5) Wt = Var(b�ot(GLS)) E;=l ( 1 /Var(�Jot(GLS) )) The varian ce of bk(GLS) • denoted by Var( bk ( GLS) ) , can b e verified to be
T
Var(bk (GLs) ) = L wi Var (bkt(GLS) ) . t= 1 Note that the Wt 's of (10.20.5) are the positive weights on the b k t ( GLS) 's that minimize the variance of the linear comb ination of the A
319
Econometric Issues in Testing the CAPM
b kt (GL s) 's on the righthand side of (10.20.4) . Note also that when Var( bA:t (GL S) ) is a constant over time, Wt = 1/T and the estimates
of the pooled model of (10. 19. 1) are simple averages of the separate monthly estimates. That is, T A b k t(GLS) bA k( GLS)  """ L., . t=l T
(10.20.6)

Moreover, we have in this case,
Var ( bA A:( GLS) )
_ 
Var( bA:t( GLS)) T
·
(10.20.7)
Note from ( 10.20.3) that Var(b t) is a constant over time if Xt and Vt are time stationary. Indeed, this will be the case in point in Exercises 10.6.210.6.4. Blume and Friend (1973) applied O LS to the crosssectional re lation between timeseries average excess rates of return and betas of ( 10. 1 3 . 1) but relied on a "canned" OLS program to generate the variances. Thus, they obtained estimates of a and b which would be unbiased in the absence of measurement error but obtained bi ased tstatistics. In contrast, Fama and Macbeth ( 1973) applied OLS to monthly excess rates of return and betas by using ( 1 0 . 1 3.2) for each month. They then calculated the timeseries means of the regression coefficients and their timeseries variances. Except for small monthly differences in betas, the average of the estimates of at 's and bt 's would be identical to the estimates of a and b of Blume and Friend ( 1973) . However, the estimators of the variances of the OLS estimators of b in Fama and Macbeth ( 1973) take account of the full variancecovariance matrix. Although emphasizing a time series b ased interpretation, Black , Jensen and Scholes ( 1972) also ob tained OLS estimates of a t 's and calculated their timeseries mean and variance. They found that a was strictly positive and statisti cally significant. In Exercise 10.2 , the reader is asked to verify that the Black , Jensen and Scholes estimator of the intercept term a is identical to that given in (10.20.6) and that their estimator of its variance is identical to that given in ( 10.20 .7) .
320
Foundations for Fin ancial Economics
1 0 . 2 1 . The relationship between portfolio theory and the lin ear regression model is helpful in interpreting tests of the CAPM. We will see that the GLS procedure is identical to a portfolio prob lem. This follows as any linear estimator is a linear combination of the observations of the dependent variable. Since the dependent variable is the excess rate of return , a GLS estimator of a regression coefficient is the excess rate of return on a portfolio (whose weights sum to either zero or unity depending on the parameters estimated ) . It is well known that GLS estimators are BLUE (for Best Linear Un biased Estimators) . The "best" means that a GLS estimator has the minimum variance among all the linear unbiased estimators. Thus, each coefficient in (10.20. 1 ) is estimated as the linear combination of the securities excess rates of return that has minimum variance subject to the unbiasedness condition. The unbiasedness condition is th at; the expected value of the estimator be equal to the true value of the coefficient of interest. The estimator for b 1ct is identical to the rate of return on a port folio obtained by solving the following portfolio problem in month t :
.min
{w,·, ;J = 1 ,2 , . ,N} ..
s.t. E
N
N
L L WjtWif CTjlt J=l I =1 .
[t l
(10. 2 1 . 1 )
w;tr;t = b,c� ,
J=l
where Wjt denotes the portfolio weight o n asset i a t time t and u jlt denotes the covariance at time t of excess rates of return on securities i and l . Using the two independent variables model of Section 1 0 . 1 9 and the definition of a linear estimator, the unbiasedness constraints for the coefficients of the two independent variables are implemented as follows: F irst, we note that N
L j=l
w;t rit = bot
N
L
j= l
w;t
N
+ bu L w;tf1;m.t j= l
N
N
j=l
i= l
+ b2t L Wjt d;t + L Wjt Ujt ·
(10.21.2)
Econometric Issues in Testing the CAPM
The unbiasedness constraint o n the coefficient for betas is N
E[L w;ti;t]
i=l
=
321
bu,
or, equivalently, N
N
N
bot L w;t + bu L w;tfJ;t'h.t + b2t L w;td;t btt,
i=l
i=l
=
i= l
(10.21 .3)
.where we have used the fact that the disturbance terms have zero expectations. For (10.2 1 . 3) to hold for arbitrary betas and dividend yields, it is necessary and sufficient that N
N
L W;t = a, L w;tfl;m.t = 1 ,
i=l
i=l
and
N
L w; td ,., = a
i=l
.
(1a.21 .4}
Therefore, (1a. 2 1 . 1 ) is equivalent to minimizing the same objective function as in ( 1a.21.1) subject to the constraints of (1a.21.4} . Note that Wit = a implies that the "portfolio" is a selffinancing portfolio; w;tfl;m.t = 1 implies that the "portfolio, has unit beta with respect to the proxy; and w;t d;t = 0 implies that this "portfolio, has a zero dividend yield. Of all feasible selffinancing portfolios meeting these conditions, the portfolio with the smallest variance would be the one that meets the efficiency property of a G LS estimator. The corresponding unbiasedness conditions for the coefficient on the dividend yield are
E f= t L;f:, 1
N
Ef=, 1
N
L Wit = a, L w,.,fJ,..,ht = a ,
i=l
i=l
and
N
L w;t d;t = 1 .
i=l
(1a.21 .5}
L W;td;t = a.
( 1a.21 .6)
The corresp onding unbias edness conditi ons for bot are N
N
i=l
i=l
L Wj t = 1 , L Wjtflimt = a,
and
i
322
Foun dations for Financial Economics
•
In words, the G LS estimate of b 2t is the excess rate of return on a selffinancing portfolio that has zero beta and unit dividend yield and that of bot is the excess rate of return on a "normal" portfolio that has zero beta and zero dividend yield. The standard deviations of the rates of return on these portfolios are , of course, the standard deviations of the estimated coefficients. To understand the intuition behind the analogy between the G LS procedure and a portfolio problem recall that the G LS estima tors are the best linear unbiased estimators. Since we have no prior knowledge of the betas and the dividend yields, unbiasedness can only be assured by the constraints of (10.21 .4)( 10.2 1 .6 ) . The best linear unbiased estimator is the excess rate of return on the minimum variance portfolio that meets those constraints. 1 0 . ::1l ::1l . In the GLS procedure discussed above, it w as assumed that V t is known. In practice, the variancecovariance matrix of asset returns is not known and must be estimated. To simplify this estimation, it can be assumed that the variancecovariance matrix has a certain special structure and the GLS estimation is then said to be based on a restricted variancecovariance matrix. In this section, we discuss a special structure of the variances and covariances of asset returns  the single index model. We will assume , throughout this section , that betas, variances , and covariances are stationary over time so we will drop their time subscripts. Assume also that excess rates of return on assets satisfy
r;t _
_
Cov(e; c , e��:t )
= a; + f';.;;,lm.t + e;c , =
{
0,
u �i
for j I k, for j = k.
Vt
=
I,
2,
. .
. , T, (10.22 . 1 )
Note that u �I. is independent o f t . Under the assumption o f the single index model, the covariances among individual asset returns are explained by their common covariances with the market index: if i I k; i f j = k;
(10.22.2)
where u i is the time stationary variance of the rate of return on
323
Econometric Issues in Testing the CAPM
the market proxy. From ( 10.22.2), we can see that the variance covariance matrix is completely specified by knowledge of the f3i m. 's, u i , and the u;1 's . The single index model is an overly simplified model of the re turn generating process. If im.t is a linear combination of the ii t 's , this linear combination will not h ave a disturbance term. That is, a linear combination of the eit ' S will be identically equal to zero. This means that covariances of asset returns cannot all be zero and contradicts (10.20 . 1 ) . However, this assumption greatly reduces the amount of information required in the estimation process. An al ternative procedure is to use a multiple index model incorporating industry specific indices in addition to the market index. The variance of a linear estimator in the context of the single index model may be expressed as N
Var(L WjtTi t ) i =l
=
=
N
N
L w]t (f3]m. ui + u�) + L L Wjt Wlct (f3imf3km. ui) i= l i=l k #i N
N
i= l
i= l
(L Wjtf3i m. )2ai + L w]t u �r
Thus, the problem of minimizing the variance of an estimator re duces to the minimization of Ef=,1 wjt u;; , because the unbiasedness condition necessitates that Ei =l Wjtf3im = 1 or 0 depending on the coefficient being estimated. Note that when the assumption of a single index model is not made , an additional term involving the covariance of the disturbance terms, N
L L Wj t Wktt1ejelo ' i =l k#i
must be added, where Ue ; e �o denotes the time stationary covariance of li t and ekt · When the assumption that u;1. = u; , a constant, is made, a GLS estimator reduces to an OLS estimator. This is the case considered by Black and Scholes (1972), where the independent variables are betas and dividend yields. 1 0 . 2 3 . We h ave discussed the estimation procedure for the model (10.20 . 1 ) for general Vt . When Vt is a diagonal matrix, GLS
324
Foundations for Financial Economics
estimators for the model ( 10.20. 1) may be obtained by making a simple heteroscedasticity correction and using O L S . This is accom plished by deflating the excess rate of return, b eta, and dividend yield of an asset by the standard deviation of its disturbance term, uu1 , . Note that this involves an OLS regression with the constant term suppressed. This procedure is a weighted least squares regres sion (WLS) , where the weights are the standard deviations of the disturbance term of each asset. When uu1, is constant across assets, this estimator is, of course, identical to the OLS estimator with the undeftated variable and a constant term. The WLS estimators are A b t(W LS) = (x•t T x•t ) lx•t T rt.. •
(10.23.1 )
where and x;
=
(
u:�·u u,.Nt
Note that in this case
and
E(u:u: T ) = IN ,
where as usual IN denotes the N x N identity matrix. Note also that by the definition of X 't and r't and the assumption that Vt is diagonal, we have (10.23.2)
Recall the lefthand side of ( 10.23.2) is the GLS estimator of be given in ( 10.20.2) , and the lefthand side is the WLS estimator of be given in ( 10.23. 1 ) . Thus, the WLS estimator are identical to the G LS estimator when the variancecovariance matrix is restricted to be diagonal.
Econometric Issues in Testing the CAPM
325
1 0 . 2 4 . The previous analysis treated the betas as a fixed in dependent variable observable without error. In actuality, only esti mates of the true betas are available. Even if the estimates of betas are unbiased, the resulting OLS or G LS estimators of the coefficients would be biased. Moreover, the O LS or GLS estimators are incon sistent. These problems will be discussed in the following sections.
Consider the case of a single independent variable. As sume that the crosssectional model we desire to estimate for month t is (10.25.1) j = 1 , 2 , . . . , N. 10.25.
However, instead of using the u nobservable true beta, fJ;m.t . we u s e an unbiased estimate, /J;m. t · To simplify the discussion, assume that an equally weighted index o f the N assets used i n the test i s used a s Uae market proxy and that ( 10.25 . 1 ) satisfies the OLS assumption that E[ii,iii) = u!I N throughout our discussion about the measurement errors. The equally weighted average of the excess rates of return on these N assets would obviously equal the excess rate of return on the market index, that is, r;e / N = rm.t · Furthermore, because covariances are additive, the equally weighted average of the betas (both true and estimated) of the N assets used in the test is equal fJ;m.t/N = 1. The case of a value weighted to unity, that is, market index would not change the conclusions substantively. The OLS estimate of b, is
Ef=,1
Ef=t
(10.25.2) The estimated beta, /J;m.t , may be expressed as the true b eta, fJ;m. t , plus a measurement error, v;1, (10.25.3) Since /J;m.t is unbiased,
E[ii;t] = 0 ,
j = 1 , 2 , . . . , N.
(10.25.4)
326
Foundations for Financial Economics
Before we proceed , substitute the right hand side of (10.25.3) for
P;m.t in (10.25.2) to get "' N
(r;t  rn.t)(I'J;n.t  1 ) "' N (r;t rn.t )";t + L..Jj = I L..J j =1 N N b t(OLS) = N 2 • 2 1 (I'J;.;.e ) 2 "' N � "' N ( 1'J;n.t  1 ) v; t "' + + L..Jj =1 N L..J j =1 N N L..J j =1 (10.25.5) Our purpose now is to show that even under a set of plausible as sumptions about the measurement errors, b t(OLS) is inconsistent. We will take probability limit of (10.25.5) when the number of assets in cluded in the sample goes to infinity. Note that as the number of assets in the sample increases , the equally weighted index and the corresponding asset betas will change. We will not, however, denote this dependence for notational simplicity. We will make the following plausible assumptions about ii; t : A
if j i k , =0 < 2 < Uvt 00 I" f J  k 1 Cov ii; t , u��:t = 0 Vj, k. < uu Var V;t 2  4t l J. = 1 , 2, . . . E Vjt Vft. t l
(
�
[
�
) ( )
{
·
_
_
_
( )
(10.25.6) (10.25.7) (10.25. 8)
( })
Note that a t and a t are upper bounds of Var ii ; t and Var ii t across assets, respectively. Relations (10.25.6) and (10.25.7) basically say that measurement errors are cross sectionally uncorrelated and are uncorrelated with the disturbance terms. We will further assume that t . I Im E =1 2 m = 0 . (10.25.9)
f P]
N+oo
N
That is, the sum of the squares of true betas increases more slowly than the square of the sample size as the sample size goes to infinity. Note that a sufficient condition for ( 10.25.9) is that the true betas are uniformly bounded as the the number of assets in the sample goes to infinity. In Exercise 10.3, we ask the reader to demonstrate that relations ( 10.25.4) and (10.25.6)(10.25.9) imply  "'N ; =1 r; t  r;nt ) v; t . p m L..J
h
N+oo
(
N
=
0,
( 10.25. 10)
327
Econometric Issues in Testing the CAPM 1.
N+ oo
p 1m
Ef=t P;m.tVjt = 0 , N
E p 1m J = l 1.
.
N+oo
and
.
p 11m
N+ oo
N
( 10.2 5 . 1 1)
N 
tl ·
t J = 0,
Ef=t vj, = N
( 10.25.12)
. 1 1m N+oo
E f= 1 Var(ii;t) N
( 10.25. 1 3)
respectively. Now define tl . '\'I! t L...iJ =l ( 1'Jrh  1"1m 1't ) =
uv ar ( tl
N+oo
1) 2

N
( 10.25.14)
the asymptotic sample variation of the true betas. We assume that Var(pt) =/: 0. Taking the probability limit of (10.25.5) by using (10.25.10)( 10.25.13) gives •
�
ph m b t(OLS )
N+oo
=
bt v.pf;;. \ , 1+� Var(Pe )
"� Var iiJ f ) Var tit  ) = 1.1m L...i J 1
where
(
_
N+oo
N
(
·
( 10.25.15)
( 10.2 5 . 16)
Thus, b t(OLS) is not a consistent estimator of bt . From ( 10.2 5 . 1 5) , w e see that the magnitude of inconsistency can be improved either when Var vt is small or Var t is large. The OLS estimator of at is
( )
(P )
Moreover, since b t(OLS ) is not consistent , O.t(OLS) is not either. 1 0.26. There are at least three approaches to the problem of measurement errors in the betas. The first approach groups assets into portfolios. Under the assumption that the measurement errors are uncorrelated across assets, the variance of the measurement er rors for the portfolio betas would approach zero as the number of
328
Founda tions for Financial Economics
assets in each portfolio increases. The second approach is an instru mental variables approach, in which an instrumental variable that is highly correlated with the true beta but uncorrelated with the measurement errors is used to obtain consistent estimators of ae and The third approach estimates the variances of the measurement errors and subtracts the average variance of the measurement er ror from the crosssectional sample variation of the estimated betas. These approaches will be developed in Sections 10.27 to 10.32.
be.
1 0 . 2 7 . The most frequently used approach to the problem of measurement errors in variables is grouping assets into portfolios. If the measurement errors associated with the betas satisfy (10.25.5) and are uncorrelated with the criterion used to group the assets into portfolios, the variance of the measurement errors associated with the portfolio betas approaches zero as the number of assets in the portfolios increases. It then follows from (10.25.15) that bt (OLS ) will be consistent. To see this, note that the variance of the measurement error of the beta for portfolio g consisting of L assets is
Var(v;e) L...., ( ) � L2
Var Vgt Then
=
j= l