Professional Documents
Culture Documents
Drobetz Global Stock Markets
Drobetz Global Stock Markets
With a foreword
by Prof. Dr. Heinz Zimmermann
Das Werk einschließlich aller seiner Teile ist urheberrech~ich geschützt. Jede
Verwertung außerhalb der engen Grenzen des Urheberrechlsgesetzes ist
ohne Zustimmung des Verlage~ unzulässig und strafbar. Das gift insbeson-
dere für Ver-vielfältigungen, Ubersetzungen, Mikroverfilmungen und die
Ein-speicherung und Verarbeitung in elektronischen Systemen.
www.gabler.de
www.duv.de
Höchste inhal~iche und technische Qualität unserer produkte ist unser Ziel. Bei der Produktion und
Verbreitung unserer Werke wollen wir die Umwelt schonen. Dieses Buch ist deshalb auf säure-
freiem und chlorfrei gebleichtem Papier gedruckt. Die Einschweißfolie besteht aus Polyethylen
und damit aus organischen Grundstoffen, die weder bei der Herstellung noch bei der Verbren-
nung Schadstoffe freisetzen.
Die Wiedergabe von Gebrauchsnamen, Handelsnamen, Warenbezeichnungen usw. in diesem
Werk berechtigt auch ohne besondere Kennzeichnung nicht zu der Annahme, dass solche Na-
men im Sinne der Warenzeichen- und Markenschutz-Gesetzgebung als frei zu betrachten wären
und daher von jedermann benutzt werden dürften.
Foreword
"While the state-preference approach is perhaps more general than the mean-
variance approach and provides an elegant framework for investigating theo-
retical issues, it is unfortunately difficult to give it empirical content." I
The state of the art in asset pricing has substantially changed over the past years.
While the seminal CAPM represents an equilibrium model derived under rather
restrictive assumptions on preferences or return distributions and places a lot of
emphasis on the efficiency of a somehow arbitrary market portfolio, subsequent
models were much less restrictive with respect to the underlying economic struc-
ture. For example, the arbitrage pricing theory maintains the linear relationship
between risk and return simply by assuming the absence of arbitrage profits.
While empirically more tractable than the CAPM, the main drawback of arbitrage
pricing models is that they do not provide much insight into the economic and
dynamic nature of risk premia. The "conditional" CAPM provides an elegant
econometric framework to characterize how changing economic conditions de-
termine the variability of multiple risk premia. However, this framework still re-
quires some rather ad-hoc assumptions about the economic nature of the pricing
kernel. An ingenious next step in asset pricing modeling was therefore to revert
the question to be addressed. Instead of placing strong restrictions on distribu-
tions and preferences, observed returns are used to derive restrictions which must
be imposed on the stochastic properties of the pricing kernel. A simple Euler-type
equation is typically used to characterize that approach.
1 Michael C. Jensen: "Capital Markets: Theory and Evidence". In: Bell Journal of Economics
and Management Science. Vol. 3, Nr. 2. 1972. p. 357.
VIII Foreword
ings have important implications for investors as well as finns in their financing
decisions. With respect to the methodology, the thesis demonstrates how eco-
nomic models and patterns - such as state dependent preferences or consumption
smoothing - provide the indispensable framework to prevent modem valuation
methodologies from being empty boxes. I hope that many young scholars are mo-
tivated by this monograph to study the fascinating interaction of capital markets
and the economic process.
Acknowledgements
The turning point, however, was the possibility to join the team of Prof. Dr. Heinz
Zimmermann at the Swiss Institute of Banking and Finance (sib£) of the Univer-
sity of St. Gallen. As my academic teacher and thesis adviser I am greatly in-
debted to him. His economic intuition and his ability to structure complex finance
issues so as to make them accessible to a wide audience of finance illiterates has
fascinated me since the first days I met him. But most of all, I deeply appreciate
his liberal approach and his great generosity, giving me the opportunity to enjoy
the spotlight he would have deserved at various occasions. I am sure he knows
what I mean. I also benefited from numerous discussions with him - he is an ex-
tremely tough discussant, and it took me much longer to convince him about the
usefulness of Hansen-Jagannathan bounds than the quality of Austrian red wine.
Likewise, I am sincerely grateful to Professor Dr. Alex Keel, who graciously ac-
cepted the task of refereeing this thesis. To both of them, I would like to express
my sincere thanks for the many constructive comments and useful suggestions
x Acknowledgements
There are many more people lowe thanks. First, my colleagues at the Swiss In-
stitute of Banking and Finance (s/bt) made my life a lot easier. I thank Peter
Oertmann, Thomas Kraus, Felix Maag, and Thomas Portmann for their guidance
during my early days at the department. The wonderful friendships with Steffen
Graf, Karl Keiber, and Patrick Wegmann helped me over the many 'ups' and
'downs' in St. Gallen. They mean much more to me than just office colleagues.
Patrick Wegmann was a challenging partner in numerous academic discussions,
and I am afraid he was more often right than myself. Second, I am grateful for the
friendship with my former professor Werner Baer from the University of Illinois
at Urbana-Champaign. Although I never really went into his area, he was my first
true academic friend and mentor. His intellectual modesty impresses me.
Contents
1 Introduction 1
Bibliography 309
XVII
List of tables
5.1 Asset pricing models with a single unspecified state variable 184
5.2 Comparison of unrestricted and restricted volatilities 186
5.3 Correlation matrix of unrestricted excess return predictions 187
5.4 Asset pricing models with two and three unspecified state variables - 193
G7 countries
5.5 Asset pricing models with two and three unspecified state variables - 195
non-G7 countries
XVIII List of tables
List of figures
2.1 The equity premium puzzle and the risk-free rate puzzle 33
6.1 Feasible region for stochastic discount factors implied by G7 data 219
6.2 Hansen-Jagannathan bound and efficient frontier 225
6.3 Volatility bounds and the addition of new assets 235
Introduction
Global stock markets have witnessed major changes over the last three decades.
In particular, three important developments must be taken into account. First,
international stock markets have grown tremendously.l In 1974, the New York
Stock exchange was the only significant stock market in the world, accounting
for more than 60 percent of a world market capitalization of less than a trillion
dollars. At the start of 1999, the world stock market capitalization grew to over
23 trillion dollars. The share of U.S. equity fell to less than 30 percent in 1988,
and moved back to 50 percent by the end of 1998. The Pacific region made up
more than 30 percent of the world stock markets, but it shrank to 15 percent by
the end of 1998. Europe makes up one-third of the entire world market. At the
same time, the international interest in the emerging stock markets of Latin
America, Asia, and Eastern Europe has developed in several waves. The capi-
talization of all emerging markets accounts for roughly 10 percent of the total
world stock market capitalization. Second, international portfolio investment is
becoming increasingly popular. In a fully efficient international stock market,
buying a world market portfolio would be the natural strategy. However, it is
only recently that the benefits of international diversification are exploited on a
larger scale. For example, international mutual funds were one of the fastest-
growing segments of the mutual funds industry in the 1990s. This trend will be-
come even more pronounced once European countries start privatizing their
pension systems. Third, international stock markets have become increasingly
integrated, which is to say that - roughly - the pricing of stocks is independent
of the market they belong. Improvements in technology, liberalization of capital
controls, and the increasing globalization of financial services have clearly en-
hanced this trend. In particular, this is the case for the emerging stock markets.
Despite the many crises experienced, many practitioners still believe that there is
a 'diversification free-lunch' in these markets. However, these times (if they
ever existed) seem to be over.
Asset pricing - the study of markets for financial assets including stocks, bonds,
foreign currencies, and derivatives - is a field in which there is an intense and
fruitful interaction between empirical and theoretical research. Financial econo-
mists study many different phenomena, including the high rewards that investors
have received for holding stocks, the apparent predictability of stock and bond
returns at long horizons, and the unusual patterns in option prices. In each area,
empirical puzzles can stimulate new thinking about investor behavior and the
functioning of capital markets. During the last two decades theoretical and em-
pirical developments in asset pricing have been based on one or more of three
central concepts. First, the law of one price posits that the prices of any two
claims which promise the same future payoff must be the same. The law of one
price arises as an implication of the second concept, the no-arbitrage principle.
Loosely speaking, arbitrage allows to make riskless profits on an arbitrarily
large scale. Arbitrage opportunities tend to be eliminated by trading in efficient
markets, because prices adjust as investors try to exploit them. The third concept
behind asset pricing models is financial market equilibrium. Optimal portfolio
holdings are derived from an optimization problem. A necessary condition for
financial market equilibrium is that the first-order conditions of the investors'
consumption-investment optimization problem are satisfied (assuming investors
can freely buy and sell any security). This requires that investors are indifferent
(at the margin) to small changes in their portfolio holdings. The utility loss from
consuming less today and buying another unit of the financial assets must equal
the marginal increase in (discounted) expected utility from consuming more to-
morrow. Financial market equilibrium also assumes a market-clearing condition.
Intuitively, the aggregate of investors' desired portfolio holdings must equal the
'market portfolio' of securities in supply. These concepts are very powerful. Un-
fortunately, they are so general that they place almost no restrictions on real
world financial data. Therefore, all research in asset pricing must attempt to un-
Introduction 3
derstand the economic forces that determine the interplay between risk and re-
turn on financial markets, that is, to find meaningful cross-sectional restrictions
for expected returns. In addition, financial ratios of stock prices to book values,
earnings, or dividends are used in time series studies of the stock market as a
whole. These ratios, along with other variables including yield spreads between
long- and short-term or between low- and high-quality bonds, have some ability
to forecast aggregate stock returns. There is still a lot of research necessary to
explore the equilibrium implications of the empirical evidence for optimal port-
folio choice. In particular, the question is whether a rational investor should try
to 'time the market', i.e. adjust the portfolio share in stocks in respo~se to
changes in the financial ratios that predict returns. While there is preliminary
empirical evidence supporting timing strategies, financial economists do know
very little whether these conditional asset returns could be consistent with gen-
eral equilibrium. From an economic point of view, therefore, the most important
task is to explore where the variation in expected returns comes from, and
whether it can persist in the face of market-timing responses by investors.
plans based on expected real returns, but there is no reason that the magnitude of
adjustment could cause predictable changes in returns to fully disappear. Those
investors are providers of 'insurance'. Recession averse investors can reduce
their exposure to stocks that are particularly sensitive to declines in real activity,
but to do so they have to pay a premium to investors who deliberately take on
recession risk. After all, any strategy that works in the stock markets must have
a real economic function.
This book attempts to answer some of these questions. The title of this thesis
"Expected returns, consumption, and the business cycle on global stock mar-
kets" already indicates that I argue for a risk-based explanation rather than some
form of irrationality. In particular, the main proposition is that investors hold
assets to smooth their consumption patterns over time. People do not dislike
stocks per se, but because they perform poorly in recessions. To make sure in-
vestors hold the market portfolio, the expected rate of aggregate return adjusts
accordingly over the business cycle. This is the rationale for consumption-based
asset pricing, an approach that has dominated financial theory at least since the
seminal analysis by Breeden (1979). Expecting that the economy will do well in
the future, investors attempt to increase current consumption by selling off part
of their holdings. In equilibrium, however, the agent has to hold his pile of as-
sets, which can only happen - accepting a risk-based explanation - if asset
prices are low and expected returns are high. In contrast, investors anticipating
lower real activity in the future will attempt to transfer wealth to this anticipated
period of scarcity, buying additional assets. Today's stock prices must then be
high, implying low expected rates of return. In a nutshell, expected stock returns
vary countercyclical. Most important, the predictability of long horizon returns
is not at all inconsistent with the concept of market efficiency. In other words,
stock prices need not follow a random walk in efficient markets as long as the
predictable component of returns is somehow related to the business cycle.
Chapter 2 contains a survey of the major results of modern asset pricing theory.
The stochastic discount factor framework is kept throughout. This is the most
Introduction 5
direct way to obtain the consumption-based asset pricing model, revealing the
interplay between consumption smoothing and risk correction. Unfortunately,
the standard model cannot explain the historical excess return of stocks over the
risk-free rate of return, an observation referred to as the equity premium puzzle.
Extensions of this model are shown, including non-expected utility, heterogene-
ous agents, and incomplete markets. The second major part of the chapter shows
how the stochastic discount factor framework can be linked to the more tradi-
tional beta-pricing ·approach. This implies that one cannot believe that, for ex-
ample, the capital asset pricing model does hold, but the consumption-based
model as a class does not. Both frameworks can be extended to apply in a con-
ditional context, which is the central theme of this book. Conditional asset pric-
ing models not only predict that differences in conditional risk determine differ-
ences in conditional expected returns, but they also imply that expected returns
vary over time, given the information available at the beginning of each period.
Chapter 4 starts with the standard instrumental forecasting model and discusses
its economic interpretations. A large part of this chapter is devoted on how to
identify proper economic instrument variables and interpret their prediction
power in empirical work. The crucial point is that predictability ought to be re-
6 Chapter 1
Chapter 6 further explores the apparent equity premium puzzle on global stock
markets within the framework of HansenlJagannathan (1991). They derive effi-
ciency regions for stochastic discount factors to fall into in mean-standard de-
viation space. There is a direct link between such volatility bounds for stochastic
discount factors and their more traditional counterpart, the efficient frontiers for
portfolio returns. It is also shown that conditioning infonnation can be easily
incorporated into the analysis by adding scaled returns. Most important, modem
versions of 'spanning tests' using volatility bounds for stochastic discount fac-
tors can potentially answer a variety of issues in international finance. First,
spanning tests allow determining whether it is harder to price international eq-
uity. In fact, this is an unparametric test for stock market integration. Second,
the benefits of international diversification can be assessed both from an eco-
nomic and a statistical point of view. This is particularly interesting for a broad
sample of emerging stock markets. Third, the importance of currency hedging
Introduction 7
Chapter 7 presents a regime switching model with two different states of the
economy, defined over the conditional moments of logarithmic consumption
rates. Intuitively, these states are interpreted as booms and recessions of real ac-
tivity. Given the estimates for the properties of the stochastic discount factor,
one can solve for the equilibrium asset prices and returns. Clearly, this model
cannot explain the magnitude of the equity premium on global stock markets.
However, if the consumption-based approach to asset pricing has some merits, it
should at least capture the mean reversion of long-horizon stock returns. Monte-
Carlo simulations are used to generate equilibrium variance ratios, which can
then be compared to their historical values. Hence, this model links stock return
predictability on the basis of slow-moving instrument variables with consump-
tion smoothing and the business cycle. Unfortunately, the portfolio implications
of predictable stock returns are not completely clear at the current state of re-
search.
9
Chapter 2
The central goal of asset pricing is to determine the prices or values of claims to
uncertain payments. This is challenging because the timing and the risk of un-
certain payments have to be taken into account simultaneously. A low price im-
plies a high rate of return, so one can also think of asset pricing as explaining
why some assets exhibit higher returns than others. Intuitively, if we lived in a
world without risk, the price of an asset would simply be the sum of its future
cash flows, discounted using the risk-free rate. Over the last three decades theo-
retical and empirical developments in modern asset pricing have taken place
within a well established paradigm. This paradigm emphasizes the structure
placed on financial asset returns by the assumption that asset markets do not
permit the presence of arbitrage opportunities - loosely, opportunities to make
riskless profits on an arbitrarily large scale. In the absence of arbitrage opportu-
nities, there exists a stochastic discount factor that relates payoffs to market
prices for all assets in the economy. This can be understood as an application of
the Arrow-Debreu model of general equilibrium to financial markets. A state
price exists for each state of nature at each date, and the market price of any fi-
nancial asset is just the sum of its possible future payoffs, weighted by the ap-
propriate state prices. Further assumptions on the structure of the economy pro-
duce stronger results. For example, if markets are complete, the stochastic dis-
count factor is unique. If the stochastic discount factor is linearly related to a set
of common shocks, then a linear factor model describes asset returns. If the
economy is populated by a representative agent with a well defined utility func-
tion, the stochastic discount factor is related to the marginal utility of aggregate
consumption. Even recent developments in behavioral finance, which emphasize
nonstandard preferences or irrational expectations, can be understood within this
paradigm. Unfortunately, because these conditions of modern asset pricing are
10 Chapter 2
so general, they place almost no restrictions on real world financial data. There-
fore, the challenge for future research is to understand the economic forces that
determine the stochastic discount factor, or put another way, the rewards that
investors demand for bearing particular risks. Specifically, financial economists
need to understand and measure the sources of macroeconomic risks that drive
asset prices. Recent papers argue that expected returns vary across assets in
ways that are linked to macroeconomic variables, such as 'recession' or 'finan-
cial distress' factors. l
This chapter draws heavily on Cochrane (1999a). In fact, his book draft provides
an excellent synthesis of the many strands in the literature and is an almost un-
limited source of wisdom in all areas of asset pricing. Some parts of this chapter
also refer to the survey articles by Ferson (1995), FersonlJagannathan (1996),
Cochrane (1997), and Campbell (1998). The chapter adopts the structure of Co-
chrane (1999a), which deviates substantially from traditional textbook treat-
ments. The consumption-based asset pricing model is introduced in section 2.1.
The related stochastic discount factor framework is maintained throughout the
chapter. This model is directly related to the life cycle-permanent income the-
ory. Since this link is so central to modem asset pricing, the rationale for the
consumption smoothing argument is presented in section 2.2. The interplay be-
tween consumption smoothing and risk correction is demonstrated in section
2.3. Section 2.4 shows important properties of the stochastic discount factor.
The results of this section are important for interpreting many aspects of the em-
pirical tests in later chapters of this book. There are two famous puzzles the
complete market model imposes, the equity premium puzzle and the risk-free
rate puzzle. They are both demonstrated in section 2.5. The poor performance of
the standard model led to a growing body of literature assuming heterogeneous
agents and incomplete markets. Section 2.6 presents the major results. Section
2.7 presents a model that incorporates most aspects of the asset pricing literature
over the last three decades and, hence can serve as the starting point for future
research. Some general equilibrium considerations are presented in section 2.8.
Section 2.9 is central to the rest of this book. In particular, it shows how the sto-
chastic discount factor framework can be linked to the more traditional beta-
pricing approach. The empirical part of this book contains tests of both fonnula-
tions. An introduction to conditional asset pricing is given in section 2.10, which
is probably the central issue of this book. Conditional asset pricing models not
only predict that differences in conditional risk determine differences in condi-
tional expected returns, but they also imply that expected returns vary with
changes in conditional risk and time varying risk premiums. Finally, section
2.11 shows that all existing factor pricing models imply different specifications
of the stochastic discount factor. That is, one cannot believe that the capital asset
pricing model (CAPM) and the intertemporal capita! asset pricing (ICAPM) do
hold, but the consumption-based model does not.
The basic equation of modem asset pricing was originally developed in a semi-
nal paper by Lucas (1978). He assumes that individuals hold assets in order to
optimize their intertemporal consumption patterns. The ultimate goal is to find
the value of an uncertain payoff Xt+I. Being invested in a stock, the payoff next
period is the future stock price plus the dividend, X'+I = P'+I + Dt+'. Clearly, X'+I
is a random variable. Therefore, the investor does not know how much he gets
exactly from his investment, but he or she can assign probabilities to a set of
possible outcomes (,states of the world'). Investors are modeled by a Neumann-
Morgenstern utility function, denoted as U (.), defined over current and future
values of consumption C, and Ct+I' respectively:
where u(.) denotes instantaneous utility. The utility function captures the desire
for more consumption, rather than a desire for intennediate objectives such as
means and variances of portfolio returns. The investor does not know his wealth
tomorrow, implying that consumption Ct+1 is also random. Period utility u(.) is
increasing, reflecting a desire for more consumption (non-satiation), and con-
cave, reflecting the declining marginal value of additional consumption. The
curvature of the utility function also generates aversion to risk and intertemporal
substitution. In other words, the investor prefers a steady consumption stream
both over time and across states of nature. ~ is a subjective time discount factor
12 Chapter 2
and accounts for an investor's impatience. Power utility is the most widely used
specification of period utility u (.) :
y stands for the constant coefficient of relative risk aversion. The investor can
freely buy and sell any amount ~ of the payoff X'+I at price P,. Denoting the
investor's original consumption with Y" his intertemporal maximization prob-
lem can be written as follows:
Substituting the constraints in (2.3b) into the objective function (2.3a), and set-
ting the derivative with respect to ~ equal to zero, yields the well known first-
order condition for the optimal consumption and portfolio choice: 2
where m'+1 is the stochastic discount factor, the pricing kernel, or the marginal
rate of intertemporal substitution. The first order conditions in (2.4) and (2.5) are
referred to as the Euler-equation. They represent the central formula of modem
asset pricing. All asset pricing models fall out as specifications of the stochastic
discount factor ml+ l . Both equations describe the standard marginal conditions
for an optimum. P, u' (C,) is the utility loss from buying another unit of the as-
set, and E, (~u' (CI+ I)XI+ I) is the increase in (discounted) expected utility ob-
tained from the payoff corresponding to an additional unit of the asset in t + I.
The investor buys more or less of the asset until the first order condition holds,
i.e. until the marginal loss equals the marginal gain. In equilibrium, the price of
2 For a thorough textbook treatment of all steps in the derivation - also extending the
setup towards an infinite-horizon recursive economy using dynamic programming ar-
guments - see chapter 2 in Altug/Labadie (1994).
The theory of asset pricing 13
an asset is such that the investor is indifferent (at the margin) to buying or sell-
ing an additional unit of the asset. Therefore, m,+) also denotes the intertemporal
marginal rate of substitution (IMRS). The crucial insight is that the relative at-
tractiveness between current and future consumption is reflected in asset prices.
In other words, changes in consumption should mirror changes in asset prices.
This is also the main theme in this book, and all empirical chapters refer to this
central notion of modem asset pricing in some way or the other.
An asset's gross return is computed by dividing the payoff X'+I by its price PHI.
In other words, a return can be interpreted as a payoff with price one. Therefore,
it is possible to write (2.5) in terms of returns rather than payoffs:
(2.6)
This is the famous consumption capital asset pricing model (CCAPM). The dis-
crete time statement of the model goes back to Grossman/Shiller (1982). The
continuous time version was developed by Breeden (1979) and Stulz (1981). It
is important to notice that this model is extremely general. The only necessary
assumption is that the investor can consider small marginal investments or dis-
investments. However, no assumptions have been made about whether markets
are complete, whether a representative investor exists, whether investors have
labor income, and whether the market has reached equilibrium or not. Finally,
note that the model in (2.6) is not a 'two-period' model. It holds for any two pe-
riods of a mUlti-period model.
The discussion so far has focused exclusively on the relation between asset
prices, asset payoffs, and consumption, without asking which exogenous shocks
drive the economy. In other words, equation (2.5) determines what the prices
should be, given the joint distribution of consumption and asset payoffs. This is
the endowment economy in the spirit of Lucas (1978) and MehralPrescott
(1985). Nondurable consumption appears every period, and there is nothing
agents can do to transform current consumption goods into next period's con-
sumption goods. Asset prices must adjust accordingly until investors are just
14 Chapter 2
happy to hold their endowment. 3 This is reverse to the mechanics of the perma-
nent income hypothesis. Assuming some functional form for technology, con-
sumption must adjust to these technologically determined rates of return. 4 If the
marginal product of capital changes, the consumption process must change as
well, but only to the extent that aggregate wealth is affected.
Marginal utility next year is expected to be the same as marginal utility this
year, except for a trend associated with the constant rate of time preference and
the constant real interest rate. Therefore, with linear marginal utility Hall's fa-
mous random walk result obtains as:
(2.8)
£'+1 is a rational forecast error, that is, the innovation in permanent income. Ex-
While the endowment economy framework does not fully account for the in-
vestor's consumption-investment decision, the permanent income story falls
short of explaining asset returns. 6 Clearly, none of these extreme possibilities
will be correct in the real economy. Investors can save or transform consump-
tion from one date to another, but at a decreasing rate. As investment rises, rates
of return decline. Fortunately, (2.5) characterizes any equilibrium outcome. If
E, (m'+IX'+I) is known, it can be used to determine P,. In contrast, if P, is
known, (2.5) can be used to determine consumption and savings decisions. If the
statistical models for consumption and/or asset returns are right, i.e. if they coin-
cide with the equilibrium consumption and/or equilibrium return process, both
approaches will give correct predictions for the joint consumption-price process.
In practice it is easier to work with the Euler-equation in (2.5) only (assuming
linear technology) and stop short of solving for a general equilibrium, which re-
quires many critical assumptions for the investors' entire environment. While
such a model would be appealing from a theoretical point of view, recall that the
first order condition P, = E, (m'+IX'+I) for an equilibrium must hold for any asset
6 The permanent income hypothesis does also not account for precautionary saving (see
Kimball (1990».
16 Chapter 2
and any production technology. This explains the almost exclusive use of Lucas'
(1978) endowment economy framework in the finance literature.
Rational investors are assumed to have sufficient foresight to make asset trades
that protect them, at least partially, against future contingencies affecting their
economic well being. Specifically, the investor can buy assets with payoffs that
are themselves uncertain, but that tend to be high in bad times of nature. In fact,
this is exactly the intuition of the Euler-equation in (2.5). This first-order condi-
tion contains the expectation of a product and, hence can be written as the prod-
uct of expectations plus a covariance term:
Time subscripts are omitted, because they are not needed at the moment. From
(2.6), since the risk-free rate of interest, denoted as R f , has zero covariance with
the pricing kernel, it is just the reciprocal of the stochastic discount factor:
E(X)
P=~+cov(m,X). (2.11)
f
This is a central result, showing that the price of any asset consists of two parts.
The first part on the right-hand side is the standard discounted present value
formula in a risk-neutral world. The second term on the right-hand side repre-
sents a risk adjustment. An asset with payoffs that are positively (negatively)
correlated with the discount factor has a high (low) price. Substituting for minto
(2.11) in terms of marginal consumption utility gives:
(2.12)
The pricing formula in (2.12) provides the unifying link between asset pricing
and macroeconomics. Virtually all issues relating to business cycles, economic
The theory of asset pricing 17
(2.13)
The representation in (2.13) reveals that - at least for small (marginal) portfolio
changes - the covariance between consumption and payoff determines the im-
pact of adding a bit more of each payoff on the volatility of consumption. How-
7 For example. to study business cycles one includes at least labor. leisure. and some sort
of technological shock.
8 This is not always true for wealth when the investment opportunity set is allowed to
vary over time. See the discussion in Breeden (1979). p. 278.
9 See Breeden (1979). p. 278.
18 Chapter 2
Using R f =ljE(m) and substituting in for consumption shows that all assets
have an expected return equal to the risk-free rate plus a risk adjustment, or a
risk premium:
The expression in (2.12) also shows that even if the payoffs of a risky asset are
very volatile, but uncorrelated with the stochastic discount factor, it pays an ex-
pected return equal to the risk-free rate:
The intuition is that an additional unit of such an asset does not have any first-
order effects on the variance of an investor's consumption stream. Ultimately,
this is the well known distinction between systematic risk (which gets a risk cor-
rection) and idiosyncratic risk (which does not get a risk correction). Any payoff
can be decomposed into a part that is perfectly correlated with the discount fac-
tor, hence the projection of X on m, denoted as proj(Xlm), and an idiosyncratic
part £ that is uncorrelated with the stochastic discount factor. In other words,
the following regression is run:
The theory of asset pricing 19
x =proj(Xlm)+£, (2.17)
which implies that the price of residual or idiosyncratic risk £ is zero, and the
price of X is the same as the price of its projection on m. In contrast to the sys-
tematic part of a payoff stream, the idiosyncratic component is that part which is
uncorrelated with m and is not reflected in a risky asset's price. JO
Traditionally, asset pricing focused on means and variances. It turns out that the
mean-variance paradigm and the Euler-equation approach are completely com-
patible. To see the link, start by writing (2.6) as follows:
cr(m)
and E(R) = R f -corr(m,R) E(m) cr(R), (2.19)
(2.20)
positively correlated with the pricing kernel (i.e. marginal utility), putting these
assets on the lower branch of the efficient frontier.
Contingent claims or state securities are securities that pay one unit of the con-
sumption good in exactly one future state, and zero in all the others. The current
price of the contingent claim is denoted with PC (s ), where s denotes the state of
nature. In a so called complete market investors can buy any contingent claim
and insure any type of risk.12 Standard results of state-preference theory show
that the stochastic discount factor m can be interpreted as contingent claim
prices divided by the associated probabilities, and that P = E P (rnX) can be un-
derstood as a bundling of contingent claims. 'P' means that the expectation is
taken with respect to the statistical probabilities. Furthermore, the stochastic dis-
count factor is often understood as a transformation to risk-neutral probabilities,
denoted by a so called martingale measure 'Q', such that P = EO (X)/R f • Tech-
nically spoken, this is a change of numeraire. 13
12 To be more specific, completeness only requires that there are just enough securities to
span or synthesize all contingent claims. In particular, European options with every pos-
sible strike price span all claims contingent on the underlying asset's price (Breeden!
Litzenberger (1978)).
13 See Zimmermann (1998) for a more detailed introduction.
The theory of asset pricing 21
The stochastic discount factor m is defined as the ratio of contingent claim price
to probability:
( ) _ PC(s) (2.23)
m s - 7t(s) .
With this notation the bundling equation in (2.21) can be written as follows:
P(X) = 2: s7t(s)m(s)X(s) = E P
(rnX), (2.24)
where 'P' means that the expectation is taken with respect to the statistical prob-
abilities. Equation (2.24) shows that in a complete market a stochastic discount
e
factor m in P = (rnX) exists. In particular, m is just a set of contingent claims
prices scaled by the probabilities, which explains why the stochastic discount
factor is also referred to as a state price density in the related literature. 14
(2.25)
(2.26)
where 'Q' means that the expectations are taken with respect to risk-neutral
probabilities 7t' instead of the statistical ones. The bundling equation (2.26) says
that investors do not require a risk premium under the new measure, and the
risk-free rate of return can be used to discount future payoffs. In other words,
payoffs can be priced as if investors were risk neutral, replacing the statistical
probabilities by their risk neutral counterpart. Combining (2.23) and (2.25) gives
the transformation:
This result is central: risk-neutral probabilities 1t* give greater weights to states
with higher than average marginal utility m. The 1t* account for risk aversion,
which is equivalent to paying more attention (relative to the actual probability of
occurrence) to unpleasant states. Finally, (2.27) explains why the stochastic dis-
count factor is also referred to as a Radon-Nikodym derivative or a change of
measure from statistical probabilities n to risk-neutral probabilities 1t* .
(2.31)
(2.32)
While the left hand side of (2.32) gives the rate at which investors can substitute
between state 1 and state 2 consumption by purchasing and selling contingent
claims, the right hand side denotes the rate at which investors are willing to
make this switch across states. Of course, an optimum requires that both ratios
are equal. In a nutshell, the stochastic discount factor m is the marginal rate of
substitution between date and state contingent claims.
(2.33)
If consumers all have power utility as in (2.2), consumption itself should be per-
fectly correlated. Specifically, in a complete market the change in each individ-
ual's optimal consumption rate is perfectly positively correlated with the change
in every other individual's optimal consumption rate and even aggregate con-
sumption. Is Assuming power utility and plugging into (2.33) gives:
(2.34)
In a complete market, where consumers can freely trade every contingent claims
all consumers share all risks and equalize their consumption growth. It can also
be shown that risk sharing is Pareto-optimal. The first order condition for a be-
nevolent social planner who wishes to maximize everyone's utility given the
available resources equals the expression in (2.33).16 Constantinides (1982)
shows that the full insurance property of complete markets implies that hetero-
geneous agents are able to equalize their marginal rates of substitution over time
and across states and, at least for consumers with Neumann-Morgenstern prefer-
ences, that the equilibrium conditions of a heterogeneous-consumer, full infor-
mation economy are the same as those in a representative-consumer, full infor-
mation setup.17 Accordingly, an investor's optimal portfolio maximizes the cor-
relation of his or her consumption with aggregate consumption, implying that
idiosyncratic fluctuations in consumption and marginal utility that are uncorre-
lated with aggregate consumption are also uncorrelated with asset returns. IS In a
nutshell, only aggregate consumption shocks matter for a risky asset's valuation
in complete markets; any idiosyncratic risk (e.g., labor income risk) can be in-
sured away by taking appropriate positions in the asset market. Complete asset
markets allow investors to share all risks, and this risk-sharing outcome is Pa-
reto-optimal.
(2.35)
In words, (2.35) guarantees that investors cannot make profits by simply repack-
aging portfolios. It is straightforward to see that the Euler-equation implies line-
arity: if X3 =X I + X 2 , it is innocuous to write E[mX 3]= E[ m(X, + X 2 )]. Hence,
the existence of a stochastic discount factor m implies the law of one price. The
reverse argument is more important, but not so obvious. The law of one price
17 This is the rationale for the assumption of a representative agent who owns the econ-
omy's average endowment. If asset markets are complete and investors face the same
prices, then, for a broad class of period utility functions, equilibrium prices and aggre-
gate per capita consumption behave as if there was a single representative agent, despite
different wealth levels across individuals. In addition, if different investors have the
same power utility function as in (2.2), then even if they have different wealth levels
they can be aggregated into a single representative investor with the same utility func-
tion as the individual investors. See also Rubinstein (1974), PrescottlMehra (1980), and
Grossman/Shiller (1982).
18 See Breeden (1979), theorem 2, p. 283. Grossman/Shiller (1982) show that this result
holds under quite general circumstances.
19 For a more precise definition see Ingersoll (1987), p. 59.
The theory of asset pricing 25
implies the existence of a discount factor. Denote the payoff space with X, i.e.
the set of all payoffs investors can purchase. If markets are complete, contingent
claims to all S states of nature are available and X = 9is . However, markets are
incomplete and X is only a proper subset of complete markets 9is . It is also as-
sumed that investors can form any portfolio of traded assets. Therefore, if
X" X 2 EX, then aX, + bX 2 EX, for any real a and b.
Existence of the discount factor: Given (i) free portfolio formation and (ii) the
law of one price, there exists a unique payoff X' E X such that P(X) = E(X'X),
for all XE X.
The proof follows along Cochrane (1999a)?O Assume that the payoff space X is
generated by portfolios of N basis assets. The basis payoffs are collected in a
vector X =(X, X 2 ••• X N )'. P denotes the corresponding price vector. The payoff
space is then X ={e'X} . A pricing kernel that is in the payoff space must be of
the form X· = b'X, where b has to be determined so as to price the basis assets.
This requires that P = E(X'X) = E(XX'b) holds. b can then be expressed as:
b = E(XX'r' P. (2.36)
X· = P'E(XX'r' X. (2.37)
There is a unique X· in the payoff space X. However, there may be other dis-
count factors not in X. Unless markets are complete there is an infinite number
of random variables E, such that E( EX) = 0 holds, that satisfy the Euler-
20 See Cochrane (1999a). p. 50. For the original sources see Ross (1978). Harrison/Kreps
(1979). Harrison/ Pliska (1981). and HansenlRichard (1987).
26 Chapter 2
In a nutshell, the law of one price implies a linear pricing function, and a linear
pricing function implies the existence of at least one and usually many stochastic
discount factors.
21 See chapter 6.
22 See chapter 2 in Ingersoll (1987).
The theory of asset pricing 27
Again, the proof follows Cochrane (1999a).23 No-arbitrage implies the law of
one price. But then there is a unique X· such that P=E(X·X). Suppose that
X· ~ 0 for some states. Then one can form a payoff X that is 1 in those states,
and zero in all the others. While the payoff is strictly positive, its price is nega-
tive:
according to the bundling equation in (2.21). However, this violates the no-
arbitrage condition. _
In complete markets, no-arbitrage implies that there exists a unique m > 0 such
that P = E( rnX)?4 Intuitively, positivity of the discount factor follows from the
microeconomic assumption of non-satiation. Absence of arbitrage is a conse-
quence of utility maximization. For a sensible class of utility functions the mar-
ginal rate of substitution is always positive, hence:
It has been shown in section 2.1 how a model that relates asset returns to the
stochastic discount factor can be derived from the first-order condition of a sin-
gle investor's intertemporal consumption and portfolio choice problem. This
section presents the canonical consumption-based model that has dominated as-
set pricing research over the last two decades. In their seminal papers Han-
sen/Singleton (1982, 1983) assume that asset returns are jointly homoscedastic
and lognormal. When random variable X is conditionally lognormally distrib-
uted, the following property holds:
This expression was first derived by Grossmann/Shiller (1981, 1982). With the
lognonnality and homoscedasticity assumptions, taking logs of (2.43) gives:
(2.44)
where lowercase letters denote logs.26 In particular, the log stochastic discount
factor is log m,+1 = log ~ - yilc,+I, with C, = log ( c. ), r..'+1 denotes the continu-
ously compounded return on asset i, a; is the unconditional variance of log re-
turn innovations var( r.,'+1 - E,rl,'+I), a; is the unconditional variance of log con-
sumption innovations var( C,+I - E,C,+,), and ale is the unconditional covariance
of innovations cov( r.,'+1 - E,r,,'+1 ,C'+I - E,c,+,), Then, the real risk-free rate of in-
terest obeys:
(2.45)
because the return innovation variance a: and the covariance a fm are both zero.
Subtracting (2.45) from (2.44) gives the expression for the constant log risk
premium on any asset i over the risk-free real interest rate:
(2.46)
sion times the covariance of the asset with consumption growth. According to
the life cycle-permanent income hypothesis, the covariance with consumption
growth is the relevant measure of risk. As discussed above, a 'risky' stock with a
high standard deviation (J, may nevertheless command a low expected return if
it is uncorrelated with consumption growth?8
The underlying intuition can be shown using quarterly consumption and stock
market data from 1947.02 to 1996.03 for the United States, taken from Camp-
bell (1998). The average excess return on stocks over short term debt (adjusted
for Jensen's inequality by adding one-half of the sample variance) was 7.852
percent per annum, with a volatility of 15.218 percent. The annualized standard
deviation of log consumption growth was 1.084 percent, and the correlation
between stock returns and consumption growth a mere 0.193. In their original
study MehralPrescott (1985) argue that coefficients of relative risk aversion y
up to 10 are plausible. Plugging these numbers into (2.46), the largest equity
premium attainable with a coefficient of relative risk aversion of 10 is 0.3 per-
cent per annum, far below the observed 7.852 percent. To duplicate the histori-
cal equity premium, a coefficient of relative risk aversion of roughly 246 is
needed. This seems extremely implausible by any standards. Campbell (1998)
notes that the equity premium puzzle arises from the smoothness of consump-
tion rather than the low correlation between consumption and stock returns. For
example, setting the correlation to I still requires a coefficient of relative risk
aversion of 47. On the other hand, a consumption volatility comparable to the
stock market volatility produces the 7.852 percent equity premium with a coeffi-
cient of risk aversion around 15 (which, of course, is still extremely high).
Given the high stock market volatility of 15.218 percent, even with 50 years of
data the 7.852 percent equity premium is still measured with a 2.15 percent
standard error. In particular, a statistical 95 percent confidence interval of the
risk premium is between 7.85+2x2.15=12.15% and 7.85+2x2.15=3.55%.
Unfortunately, this inaccuracy in measurement does not solve the equity pre-
mium puzzle?9 A quick check confirms that a coefficient of relative risk aver-
sion of 19 is needed to replicate a mere 2.9 percent equity premium. Rietz
(1988) argues that the peso problem might explain the equity premium puzzle.
Risk averse investors demand a high risk premium to compensate for the ex-
treme losses that may occur during an unlikely, but disastrous, market crash. To
the extent that equity returns were high without major crashes, equity owners
got compensated for the crashes that did not take place. Assuming a crash where
output falls by 50 percent of its value with a probability of 0.4 percent, Rietz
generates an ex-ante risk premium consistent with that observed in U.S. stock
market data with a coefficient of relative risk aversion of five. 30 Goetz-
man/Jorion (l999a) shed some doubt on the Rietz argument. If past crash fre-
quency was any indication of future crash probability, then the Rietz hypothesis
would suggest that markets with more interruptions should boost a higher equity
premium. The opposite seems to be true empirically. In particular, the U.S. is
one of the few markets for which data is available without disruptions over the
last century, yet the magnitude of the U.S. equity premium seems to be the ex-
ception rather than the rule. They conclude that survival effects are important to
explain the equity premium puzzle. Their argument is that the high equity pre-
mium in the United States can be explained - at least partly - by conditioning
estimates upon the best performing market.
1 (2.47)
l'lc=-(r-8).
y
(2.48)
Hence, risk aversion and intertemporal substitution are closely tied together in
the representative consumer setup with power utility. The elasticity of intertem-
poral substitution \jI is the reciprocal of the coefficient of relative risk aversion
y. Hall (1988) argues that this linkage is not appropriate. The elasticity of in-
tertemporal substitution concerns the willingness of an investor to smooth con-
sumption between time periods (it is well defined even if there is no uncer-
tainty), whereas the coefficient of relative risk aversion concerns the willingness
of an investor to move consumption between states of the world (it is well de-
fined even in a one-period model with no time dimension). The non-expected
Weil's (1989) risk-free rate puzzle falls out immediately. The equity premium
puzzle posits that investors must be extremely risk-averse, so they must also be
extremely unwilling to substitute consumption intertemporally. Given positive
average consumption growth, a low riskless interest rate, and a high rate of time
preference, investors would have a strong desire to smooth consumption by bor-
rowing from the future. A low riskless rate is possible in equilibrium only if in-
vestors have a low or even negative time preference p that reduces their desire
to borrow. However, consumption growth in the United States over the 1947.02
to 1996.03 period was as high as 1.921 percent per annum. Recall, to produce
the historical equity premium of 7.852 percent a coefficient of relative risk aver-
sion of roughly 246 is necessary. Substituting into (2.45) and solving for the im-
plied rate of time preference, the time discount factor is roughly 1.12. This is
greater than one, corresponding to a negative rate of time preference. The rate of
time preference can be interpreted as the riskless real interest rate that would
prevail if consumption was known to be constant, with no growth and no vola-
tility. Intuitively, people should prefer current consumption to future consump-
tion, not the other way round.
Figure 2.1
The equity premium puzzle and the risk-free rate puzzle
Period: 1947.02 - 1996.03 (quarterly data)
12% ~------------------------------------~
10%
a::>
's.,
.....
8%
0.
....
·c 6%
Vl
C
'5
0"
r.I.l
4%
2%
0%
0% 50% 100% 150%
Risk-free rate of interest
The figure exhibits the equity premium puzzle by MehraJPrescolI (\985) and the risk-free rate puzzle
by Weil (1989). Quarterly U.S. consumption and stock market data over the period from 1947.02 to
1996.03 are taken from Campbell (1998). The coefficient of relative risk aversion y increases when
moving along the parabola. Given a reasonable coefficient of risk aversion (y<\O), it is impossible to
replicate simultaneously a high equity risk premium and a low risk-free rate of interest.
Figure 2.1 shows how the equity premium puzzle and the risk-free rate puzzle
relate to each other. Together, equations (2.45) and (2.46) describe a parabola in
a two-dimensional space spanned by the risk-free rate (on the horizontal axis)
and the equity premium (on the vertical axis). To make things as easy as possi-
34 Chapter 2
ble, the time discount factor ~ is set to one. The coefficient of relative risk aver-
sion y increases when moving along the parabola, and the equity premium rises
accordingly. Initially, the risk-free rate sharply increases due to the consumption
smoothing effect. However, at a risk aversion of about 160(!) the precautionary
saving effect starts to dominate, the parabola bends back and the risk-free rate of
interest declines. Hence, a trade-off emerges: given a reasonable coefficient of
risk aversion, it is not possible to produce the high historical equity premium
and the low risk-free rate of interest at the same time. The problem is even
worse, because a volatility puzzle becomes apparent. According to (2.45), a co-
efficient of risk aversion y = 246 implies that a country with a I percentage
point higher (lower) than normal consumption growth should have real interest
rates 246 percentage points higher (lower) than normal (ignoring prudence ef-
fects). This implies extremely volatile and often negative interest rates. Empiri-
cally, however, interest rates have been rather smooth in virtually all developed
countries.
The standard model assumes that there exists a representative agent who con-
sumes aggregate consumption. Unfortunately, the representative agent, complete
market model dramatically underpredicts the equity premium and overpredicts
the risk-free rate of return. Recently, a new strand of literature tries to explain
the equity premium puzzle and the risk-free rate puzzle by arguing that aggre-
gate consumption is not an adequate proxy for the consumption patterns of in-
vestors who participate in the stock market. The short survey of published re-
sults follows along Campbell (1998).35 First, in reality there might be two differ-
ent types of agents: constrained agents and unconstrained agents. While the lat-
ter group can fully participate in asset market trading, the former type of agents
is prevented from doing so and simply consumes the labor income each period.
While the consumption of constrained agents does not affect the determination
of equilibrium prices, it may nevertheless constitute a large fraction of aggregate
consumption. 36 The argument boils down to assuming that stocks are attractive,
but a variety of market frictions prevent people from buying them. Using U.S.
panel data, Mankiw/Zeldes (1991), Brav/Gezcy (1996), and Brav/Constantini-
des/Gezcy (1999) find that stockholders exhibit consumption patters that are
more volatile and more correlated with stock returns than those of non-
stockholders. Unfortunately, even for the group of market participants both the
volatility and the correlation are not sufficient to explain the equity premium
puzzle. It is only if those agents are highly risk averse that they are marginally
indifferent between stock and bonds.
37 Recall that according to (2.34) individual consumption streams are perfectly correlated
with each other and with per capita consumption.
36 Chapter 2
(2.49)
(2.50)
(2.51)
However, plugging into (2.46) reveals that the equity premium is completely
unaffected by idiosyncratic risk: 39
(2.52)
offers some prospect to better explain the stylized facts. Therefore, the starting
point of many recent papers has been to hand out idiosyncratic labor income and
close down the market for labor insurance. Telmer (1993) studies this type of
model, but his simulation results are disappointing. Agents can diversify almost
all of their idiosyncratic income risk by trading in a single risk-free discount
bond market. The incremental variability in the stochastic discount factor as a
result of market incompleteness is substantially lower than would be necessary
to explain the equity premium puzzle. The intuition is that nothing stops con-
strained consumers from saving. In Telmer's (1993) model idiosyncratic shocks
are transitory, allowing consumption to be smoothed over time by accumulating
financial assets after good shocks and selling assets after bad shocks. He con-
cludes that when markets are incomplete, the equilibrium allocation of risk can
be very close to first-best. Adding additional assets to the single bond further
enhances risk-sharing opportunities, thereby further reducing the volatility of the
stochastic discount factor.
(2.53)
(2.54)
Expression (2.54) posits that the stochastic discount factor is a function of the
cross-sectional mean and the cross-sectional variance of the change in log con-
sumption. Knowing the underlying preference structure, but being unaware of
the heterogeneity in the economy, one falsely constructs a representative agent
pricing kernel using aggregate consumption:
(2.55)
The second equality follows from the fact that ~Ck.'+1 is cross-sectionally uncor-
related with c k, . Subtracting (2.56) from (2.54) gives:
(2.57)
The time series of this difference can have a non-zero mean, helping to explain
the risk-free rate puzzle, while a non-zero variance could mitigate the equity
premium puzzle. If the cross-sectional variance of log consumption growth is
negatively correlated with market returns, idiosyncratic risk increases in bad
states of nature (i.e. in recessions). The true pricing kernel m;+1 becomes more
strongly countercyclical than in a representative agent model, thereby suggesting
a potential source for the equity premium. Clearly, this mechanism is assumed
away in the representative consumer economy. To my knowledge there is no
published work that examines the ConstantinideslDuffie (1996) framework em-
pirically. Preliminary results by Cochrane (1998) are discouraging. In a simpli-
fied setup he shows that the cross-sectional standard deviation of individual con-
sumption growth would have to be extremely sensitive to changes in the level of
the market return. In addition, assuming a plausible magnitude of risk aversion,
the persistence of idiosyncratic labor income shocks must be a lot higher to du-
plicate the historical equity premium than estimated by Heaton/Lucas (1996).
Epstein/Zin (1989, 1991) and Weil (1989) build on the theoretical framework of
KrepslPorteus (1978) to develop a more flexible version of the standard power
utility model. Recall, the model is restrictive in that it makes the elasticity of
intertemporal substitution 'V the reciprocal of the (constant) coefficient of rela-
tive risk aversion y. Theoretically, is it not clear whether these two concepts
should be linked so tightly. Hall (1988) notes that risk aversion describes the
individual's reluctance to substitute consumption across states of nature and is
meaningful even in an atemporal setting, whereas the elasticity of intertemporal
substitution describes the individual's willingness to smooth consumption across
dates and is meaningful even in a deterministic setting. A very high coefficient
of relative risk aversion allows to replicate the equity large risk premium. At the
same time, however, this is synonymous with a very low elasticity of intertem-
poral substitution and leads to the wrong prediction of an implausibly high risk-
free rate. The EpsteinlZinlWeil model breaks the link between these two con-
cepts. It is characterized by a constant elasticity of intertemporal substitution 'V
and a constant, but unrelated, coefficient of relative risk aversion y. Intuitively,
the former controls the magnitude of the risk-free rate, while the latter deter-
mines the risk premium. This might give the additional degree of freedom re-
quired to solve both puzzleS. 43
I-y 1) I~y
U, = ( (l-P)C,9 +P(E,U:~T)B , (2.58)
given the intertemporal budget constraint Wt+1 = Rm.t+1 (W, - C,), and where 8
is defined as (l-y)/(l-1/'V), C, denotes real per capita consumption a time t,
W, is individual wealth at time t, and Rm.'+1 is the gross rate of real return on
wealth (i.e. the 'market portfolio'). When Y= 1/'1', e= 1, the recursion in (2.58)
becomes linear and can be solved forward to obtain the standard time-separable
power utility model. 45 Using dynamic programming arguments, EpsteinlZin
(1989) show that a maximizing consumer will choose consumption such that the
following Euler-equation is satisfied: 46
(2.59)
For the special case of i.i.d. consumption growth, Kocherlakota (1990) proves
an irrelevance result. He shows that for this special case the equilibrium sto-
chastic discount factor is the same, regardless of whether the representative
agent has power utility or recursive utility.47 His empirical results show that 'I' is
irrelevant for the price determination in this case. Weil (1989) generalizes this
result. He solves (2.59) numerically for the equilibrium equity premium and
demonstrates two important results. 48 First, the risk premium depends - at least
for plausible calibrations of tastes and technology - almost exclusively on the
coefficient of relative risk aversion. The equity premium puzzle arises because
of economists' prior beliefs about reasonable levels of risk aversion. Disentan-
gling risk aversion from the intertemporal elasticity of substitution does not help
to duplicate the historical equity premium. Second, recursive utility helps to ex-
plain the risk-free rate puzzle. The equity premium can be replicated using very
high coefficients of risk aversion. With recursive utility, it is then possible to
explain the risk-free rate by allowing the intertemporal elasticity of substitution
to be high at the same time. Clearly, this is only a marginal improvement over
the standard power utility model.
prices and consumption are jointly lognonnal and allowing for heteroscadsticity
in both stock returns and consumption, the relationship between expected con-
sumption growth and expected returns is linear:
(2.60)
More important, Campbell (1993) shows that the equity premium on asset i can
be written as follows:
E. ( ) - rr.HI +
C..HI 2VII I = ao/
VIC. ( )
+ 1- a Vim .• , (2.61)
where lowercase letters again denote logs.49 In (2.60), Ilm .• is an intercept tenn
related to the second moments of consumption and market returns. In (2.61), the
V-tenns denote second moments of surprises. In particular, VII .• is defined as
var. (r;.HI - E. (r;.I+I))' Vim .• is cov. (r;,'+1 - E. (r;.'+I)' rm.HI - E. (rm.• +1)), and VIC .• is
cov. (rl,HI - E. (r;,'+1 ), t.C HI - E. (t.c.+ 1)). Hence, the risk premium on asset i is a
weighted combination of the asset's covariance with consumption growth (di-
vided by the elasticity of intertemporal substitution) and the covariance with the
return on wealth, This demonstrates that the EpsteinlZinlWeil model captures
both the consumption-based asset pricing model in (2.46) with power utility
(with a = 1 and, hence y = 1/V) and the traditional static capital asset pricing
model (with a = 0), The coefficient of relative risk aversion y is equal to the
sum of the prices attached to market and consumption risk, i.e, (1- a) + a/v = y,
The expression in (2.61) imposes cross-sectional restrictions that can be tested
empirically, 50 However, Campbell (1993) argues that this model does not ex-
plicitly account for the link between consumption and market returns via the in-
tertemporal budget constraint. He suggests substituting consumption out of the
model. Combining the log-linear Euler-equations with an approximate log-linear
version of the intertemporal budget constraint, he first shows that: 51
cov, (r..'+1 - E, ( r..HI ), CHI - E, ( C'+I ) ) == ~c., = V,m" + (1- 'l') V,h " ' (2.63)
as the covariance of unexpected returns on asset i with good news about future
returns on the market (i.e. upward revisions in expected future returns). The use
of 'h' is intended to recall the hedging demand in the spirit of Merton's (1973)
intertemporal capital asset pricing model (lCAPM).52 Substituting (2.63) into
(2.61) and using the definition of e in terms of the underlying parameters yand
'l', Campbell obtains a cross-sectional asset pricing condition that makes no ref-
erence to consumption: 53
V
E( ( r..HI ) - rf .(+1 + 2" = yV,m,( + (Y -1) V,h ,( . (2.64)
tertemporal substitution \If does not appear once consumption has been substi-
tuted out of the model.
The rate of return on the stock market is a very poor proxy for the rate of return
on wealth. Even if the stock index return captures the return on financial wealth,
as argued by Stambaugh (1982), it may not capture the return on human wealth.
This was forcefully demonstrated by JagannathanlWang (1996) only recently.
Campbell (1996) generalizes equation (2.64) to account for human wealth. 55 In
contrast, RestoylWeil (1998) argue that the rate of return on total wealth is not
only mismeasured, but also not measurable at all. They also use the link between
consumption and the rate of return on wealth in the budget constraint to derive:
1 (2.65)
E, (rm.'+1 ) = Jlc + \If E, (LlC'+1 ) ,
where Jlc is again an intercept term related to the second moments of consump-
tion and market returns. The conditional expected return on the market must be
higher, for a given conditional expected rate of consumption growth, the lower
the elasticity of intertemporal substitution. 56 They reverse Campbell's argument
and show that a consumer's total wealth can be reconstructed from consumption
data alone. Using the same log-linearization technique as in Campbell (1993,
1996), they derive a cross-sectional restriction similar to (2.64):
ance with current consumption growth and its covariance with revisions in ex-
pected future consumption growth.
The cross-sectional restrictions in (2.64) and (2.66) are central to our under-
standing of asset pricing. Looking at (2.64) first, an asset with Vim> 0 is an as-
set with high (low) returns when wealth is most (least) productive. Such an asset
makes it difficult to smooth consumption, implying that risk averse investors
(y>O) require a premium over the risk-free return to hold it. Second, when
y> 1, assets that do well when there is good news about future returns on the
market (Vlh > 0) have higher mean returns. On the one hand, such an asset is
desirable because it allows benefiting from improved investment opportunities.
On the other hand, however, this asset is undesirable because it fails to deliver
assets precisely when it is most valuable to the investor. In other words, it does
not provide a hedge against unfavorable shifts in the investment opportunity set
and, hence reduces the ability of an investor to smooth consumption patterns.
Assuming that the investor is more risk averse than implied by logarithmic util-
ity, the second effect dominates, and investors require a higher return to hold
this type of asset. 58 The interpretation of (2.66) follows accordingly, but there is
an additional term [-(l/'V)Vlg ] reflecting reluctance to substitute intertempo-
rally. This is in contrast to the Campbell model, which does not assign a role to
the elasticity of intertemporal substitution 'V in the determination of the equilib-
rium equity premium. In fact, disentangling the elasticity of intertemporal sub-
stitution from the coefficient of relative risk aversion was the original motive for
the use recursive utility. Mechanically, the more averse an investor is to substi-
tute intertemporally (the larger 1/'V), the more willing they are to hold an asset
with VI' > 0, and the smaller the excess return they require in equilibrium to
hold that asset. Intuitively, expressing excess returns as a mere function of co-
variances with market returns hides the elasticity of intertemporal substitution in
the covariance terms.
The previous discussion has exclusively focused on the relation between asset
prices and consumption, without asking how consumption is determined in rela-
tion to investment and production. A new line of research argues that asset
pricing should place more emphasis on the investment decisions of firms. Pro-
duction-based asset pricing models look at the 'other' Euler-equation, i.e. the
one characterizing the capital stock rental decision. In competitive markets rela-
tive prices represent, simultaneously, both the ability of producers to physically
transform one good into the other and the willingness of producers to substitute
one good for the other. Assuming profit and utility maximization, relative prices,
marginal rates of transformation, and marginal rates of substitution are all equal.
The consumption-based model relates asset returns to marginal rates of substitu-
tion, which are inferred from consumption data. The production based model,
originally proposed by Cochrane (l991b), ties asset returns to marginal rates of
transformation, which are inferred from data on investment through a production
function. Similar to the consumption-based model, this new approach implies
restrictions on the joint stochastic process of investment and asset returns. The
argument is that firms might be more responsive to anticipated changes in eco-
nomic conditions and, hence their investment decisions will be more closely re-
lated to asset market data than the consumption decisions of individuals. As-
suming that financial securities offer no additional spanning opportunities be-
yond those available by real, physical investment, the optimality conditions of
physical investment returns can be 'inverted' to express the stochastic discount
factor as a function of investment returns. Investment returns are the returns ob-
tained from buying a unit of installed capital (which becomes available next pe-
riod), using it to produce and sell output next period and then selling the unde-
preciated portion of the capital so as to keep future capital stock unchanged. 59 In
equilibrium, the investment return is the same as the return on the firm's share.
variation due to technology shocks as the models' driving force. However, the
problem is the absence of time variation in the price of capital. The standard
capital accumulation equation treats installed capital and uninstalled capital (in-
vestments) as perfect substitutes in making new capital. Therefore, they must
have the same price, that is, the price of stocks relative to consumption goods
must be exactly one. In this framework, Rouwenhorst (1995) shows that agents
can easily alter their production plans to reduce fluctuations in consumption,
which again makes it impossible to explain the equity premium puzzle. Hence, if
production-based asset pricing models are to generate volatile asset returns, they
must be modified to include adjustment costs on investments. One can drive a
wedge between the prices of installed and uninstalled capital by reducing output
during periods of high investment. In this case, changes in the demand for capi-
tal cause changes in the value of installed capital, or Tobin's q, rather than
changes in the quantity of capital. This makes it more costly to smooth con-
sumption trough changes in the capital stock, and agents are induced to take
more consumption risk. In equilibrium, investment returns are proportional to
changes in investment. Since investment is quite volatile, return expectations
should become more volatile as well. Empirical results by Cochrane (1996) for
u.S. data and Kasa (1997) for international data are encouraging. The model is
consistent with a good deal of the cyclical variation in investment and stock re-
turns.
Since production-based models account for the capital stock rental decision, the
pricing implications are derived from producer's first order conditions. It would
be desirable to combine the utility function and the production function to build
explicit macroeconomic models that replicate asset pricing quantities. Asset
prices mediate the consumption-investment decision, and this decision is at the
core of any dynamic, general equilibrium model. Jermann (1999) combines
habit formation with a neoclassical technology. To make sure that investors can-
not smooth consumption by adjusting their investment plans, thereby making it
impossible to explain the equity premium, he adds adjustment costs to the
model. As expected, this setup enables him to replicate the historical equity
premium. Unfortunately, he also reports a highly volatile risk-free rate of re-
48 Chapter 2
2.9 Linking the discount factor view with beta pricing models
The discussion so far has solely focused on the pricing implications of the Euler-
equation. However, much empirical work in finance has been done in terms of
beta representations for expected returns. This section outlines the connection
between the discount factor view and the more traditional beta pricing approach.
Fortunately, both approaches are closely linked with each other. The empirical
part of this book contains tests of both specifications. Again, the derivations in
this section draw on Cochrane (1999a).61
Equation (2.67) is the famous arbitrage pricing theory (APT), originally intro-
duced by Ross (1976).62 ~') is the amount of exposure of asset i to the risk of
factor j (j = 1, ... , K), and A) denotes the price of such risk-exposure, or the risk
premium. In a one factor model, all pairs of [E(R,), ~,J should line up on a line
with slope A). If there is a risk-free asset, its betas are all zero. Hence, the inter-
cept a in (2.67) is equal to the risk-free rate (or zero-beta rate):63
60 This is a well known problem for models with habit-formation. See section 7.2.2.
61 See Cochrane (1999a), p. 59f.
62 For the exact derivation and an overview of the different strands of the arbitrage theory
see Connor/Korajczyk (1995) and Oertmann (1997).
63 See Black (1972).
The theory of asset pricing 49
Rr =(X. (2.68)
Note that this regression is not about predicting returns using instrument vari-
ables. The objective in (2.69) is to determine an asset's risk exposures, and
whether returns are typically high in good or bad states of nature, i.e. whether an
asset is useful to smooth consumption. Second, given the sensitivities ~'J as the
explanatory variables, cross-sectional regressions are run to determine slope co-
efficients A.J and asset specific pricing errors (X,:
(2.70)
where the A.'S are risk premiums common to all assets. The model predicts that
(x, = O. Rather than estimating a zero-beta rate, one can examine a factor pricing
model using excess returns, which allows focusing on the central task of under-
standing risk premiums and risk corrections. Unfortunately, the second pass re-
gression in (2.70) suffers from an 'errors-in-variable' problem, that is, there are
measurement errors in the time series estimates of ~,.65 More recent papers use
the Generalized Method of Moments (GMM) to estimate both risk sensitivities
and risk premiums simultaneously.66
with a beta of one on factor j and zero betas on all other factors. Empirically, the
factors are projected on the vector of excess returns r. The result is itself an ex-
cess return, hence r] = proj(J; Ir) .68 The factor mimicking return r] then satisfies:
(2.71)
Hence, a set of excess returns on factor mimicking portfolios captures the pric-
ing implications of the original factors, so that expected excess returns obey the
following linear pricing condition:
The model in (2.72) explains expected excess returns as accurate as the original
factor pricing model in (2.67). However, note the difference to (2.67): E(r]):f. A]
and P'" :f. Pi]' Finally, note that the beta pricing model in (2.67) and the factor
regression model in (2.69) look very similar. It is tempting to take the expecta-
tion of the time series model to arrive at the cross-sectional pricing restriction.
However, this interpretation is incorrect. First, the time series regression in
(2.69) is not a license to regress anything on anything. Second, there is an asset
specific intercept a, in time series regressions, while the intercept a is the same
for all assets in the cross-sectional restriction in (2.67). The beta pricing equa-
tion is a condition for expected returns and, hence imposes a restriction on the
intercepts in time series regressions. In particular, if the factors are themselves
excess returns, all intercepts a, should be zero (e.g., Fama/French (1993)).
MUltiplying and dividing by the variance of the pricing kernel, and recalling that
a. == liE (m) yields:
E(R.)=a.+ (
cOV(m,RJ
var(m)
l( var(m) 1
- E(m) =a.+~.mAm· (2.75)
Beta pricing ¢:} linear discount factor models: The proof follows along Co-
chrane (1999a).70 Without loss of generality, take the mean of the factors b'E(f)
into the constant, so that f == f - E(f) are demeaned factors with means equal to
zero. Start with (2.76) and write:
fl, is the vector of ordinary least square regression coefficients in the time series
model (2.69),. defined as:
To make the proposition work, construct a relation between (a, A) and (a, b ):
Using (2.81), one can go backwards from the beta pricing representation to the
discount factor representation m = a + b'f .•
In a nutshell, given either (2.76) or (2.77), there is a model of the other form.
However, they are not unique. First, adding to m any random variable that is
orthogonal to returns or adding any risk factor with zero p and/or A does not al-
ter the pricing implications. Second, a multiple beta model can be expressed as a
single beta model with m = a + b'f as the single factor (or the return on the cor-
responding mimicking portfolio).71 Finally, expression (2.81) reveals that). de-
notes the price P of demeaned factors E(mf) brought forward at the risk-free
rate. To see this, write the premium J.. using the underlying factors r, such that
demeaned factors are f == r - E(r), in the following way:
This is a central result: J.. denotes the price of the factors less their risk-neutral
valuation, hence the factor risk premium. The negative sign comes in because
low asset prices imply high risk premiums, and vice versa.
(2.83)
Therefore, when the factors are orthogonal, each AJ is zero if and only if the
corresponding b J is zero. Indeed, most empirical tests of beta pricing models
have used orthogonalized factors. 73 The distinction between band J.. does not
matter in this case. If factors are correlated, however, things become more com-
plicated. In particular, AJ can tell whether factor fJ is priced. In contrast, b J
helps to assess whether factor 1; is marginally useful in pricing a cross-section
of assets, given all other factors. In other words, if b J is zero, assets can be cor-
rectly priced without incorporating factor 1;. AJ is proportional to the single re-
gression coefficient of the stochastic discount factor m on factor 1;, hence
AJ = -acov (m, 1;). Testing the null hypothesis that \ = 0 asks whether factor
1; is correlated with the true discount factor m. In contrast, according to (2.76),
bJ is a mUltiple regression coefficient of m on factor 1; in the presence of all
72 Recall that b are the regression coefficients of m on r , while p are the regression coef-
ficients of R, on r .
73 For example, see Oertmann (1997).
54 Chapter 2
other factors. Testing the null hypothesis that b J is zero can answer the question
whether an additional factor 1; should be included to an existing set of factors.
The empirical part of this book contains tests of both types. 74
(2.84)
The Euler-equation holds for any set of probabilities, ex-ante and ex-post. It can
refer to an agent's subjective probability distribution, objective population prob-
abilities, or to actual sample moments. If the law of one price holds, it is always
possible to form a portfolio X' = P'E(XX't X from sample moments such that
P(X) = E{X'X) holds exactly.78 In beta-pricing language, there exists an ex-
post mean-variance efficient portfolio such that sample average returns line up
74 See chapter 5 for a test in expected return-beta formulation, and chapter 6 for a test in
discount factor language.
75 See Stambaugh (1982) and Jagannathan/Wang (1996).
76 See section 2.4.5.
77 See section 2.l.
78 See section 2.4.5.
The theory of asset pricing 55
exactly with sample regression betas relative to that benchmark portfolio. Co-
chrane (1999a) warns that this is a danger in empirical work.79 Ad-hoc factor
pricing models will always work within a sample if little enough structure is
placed on the parameterization of the discount factor. But since these restrictions
are the entire content of the model they should better relate to the fundamental
sources of macroeconomic risk. Given that consumption is a good indicator for
bad states, the consumption-based asset pricing model is a natural candidate to
provide this link. In contrast, the arbitrage pricing theory (APT) and the in-
tertemporal capital asset pricing model (ICAPM) do not answer the question of
what factors to include. They are what Fama (1991) calls a 'fishing license,.8o It
is easy to add yet another ad-hoc factor and 'explain' (almost) any anomaly, but
this seems not a good idea at all from an economic standpoint. For example,
Carhart (1997) 'explains' momentum by adding an additional momentum factor
to the FamalFrench (1993, 1996) three-factor model. Even worse, nothing as-
sures that a portfolio that is ex-post mean variance efficient in one sample will
also do a good job in pricing assets - ex-ante or ex-post - in a subsequent sam-
ple. 8l In a nutshell, there is no theoretical gain in searching for spuriously good
in-sample pricing with lowest possible pricing error. The only empirical chal-
lenge is to identify discount factor models with compelling economic stories.
Even if they produce somewhat bigger pricing error, but are clearly related to
the business cycle, they are a lot more valuable from an economic standpoint
than any ad-hoc factor model. It is therefore too early to abandon the consump-
tion-based model yet. In the end, all factor pricing models are not alternatives to
the consumption-based model, but merely specializations of it. 82 One cannot be-
lieve in the capital asset pricing model (CAPM) or the arbitrage pricing theory
(APT), but discard the consumption asset pricing model at the same time.
Many empirical models test for the number of factors required to price a cross-
section of assets. This question is misleading per se without clearly showing the
assumptions of the particular model. Recall, m = a + b'f and E(R.) = ex + }.'P.
are equivalent, but not unique. For any multiple factor of multi-beta representa-
tion, a single beta representation does exist. Specifically, the single factor
m = a + b/f prices as well as the original factors f , as will X', the projection of
the space of payoffs K on the stochastic discount factor m. Furthermore, for any
multifactor expected return-beta representation there is an equivalent represen-
tation with a different number of factors. For example, to go from a 3-factor
model to a 2-factor representation in beta-pricing language, it holds that: 83
(2.86)
If payoffs and discount factors were i.i.d., the distinction between conditional
expectations and unconditional expectations would be irrelevant. One could
simply condition (2.86) down without any loss of pricing information to get:
(2.87)
In general, however, things are more complicated. Modeling the stochastic dis-
count factor as a linear combination of some benchmark portfolio return R~+"
such that mHI = a, + b,R~+" conditional asset pricing implies that the factor
weights a, and b, are time varying as the risk-free rate and means and variances
of risky returns vary over time. Using m'+1 = a, + b,R~+, and taking uncondi-
tional expectations of (2.86) yields a number of covariance terms that will - be-
cause both a, and b, are formed on conditional moments of returns - in general
be different from zero. 87 Therefore, simply conditioning down from (2.86) to
(2.87) is no longer appropriate. In discount factor language, expected returns,
betas, and risk premiums are likely to vary over time. However, even when ex-
pected returns are linear in betas for every time period, based on the available
information at time t, the relation between the unconditional expected return and
the unconditional beta could be 'flat'.
Stock I Stock 2
Market risk
CAPM risk CAPMrisk premium
Time Beta Beta
premium premium
The complete proof is extremely demanding, and the interested reader is referred
to HansenlRichard (1987). A simple example taken from a recent paper by Ja-
gannathaniWang (1996) can demonstrate the point. 88 Consider an economy with
two stocks that lasts for two periods only. Suppose that the capital asset pricing
model (CAPM) holds period by period, that is, holds conditionally on the rele-
vant set of information. Table 2.1 shows the risk characteristics of the two as-
sets, the market risk premiums, and the risk premiums assuming that the capital
asset pricing model holds. Average betas are 0.875 and 1.125 for the first and
second stock, respectively. If the capital asset pricing model holds in each pe-
riod, the expected risk premium on the first (second) stock will be 5 percent (15
percent) in the first period and 25 percent (15 percent) in the second period.
However, ignoring the fact that betas and risk premiums vary over time falsely
leads to a rejection of the model: two stocks with different average betas (0.875
versus 1.125) earn the same risk premium of 15 percent. In a nutshell, rejection
of unconditional mean-variance efficiency does not imply rejection of condi-
tional mean-variance efficiency.
A conditional factor model does not imply an unconditional factor model, but an
unconditional factor model implies a conditional factor model. This central re-
sult sheds negative light on the work by FamaiFrench (1992). On the basis of
unconditional multifactor models, they claim that 'beta is dead'. In other words,
the security market line is flat, beta risk is superimposed by 'risk' associated
with size and company fundamentals. This conclusion has been challenged in a
number of studies only recently.89 Formally, the example in table 2.1. demon-
strates that
E(R,)=a+pA. (2.89)
HansenlRichard (1987) show that the relation of the two properties is one of in-
clusion. For the same set of available portfolio returns, the set of conditionally
minimum-variance portfolios is (weakly) larger than the set of unconditionally
minimum-variance portfolios.
currency interest rate for some arbitrary currency. Unconditionally, it does not
lie on the efficient frontier. Even a Eurocurrency rate is a risky asset with non-
zero variance; it is surely dominated by an appropriately chosen portfolio. Con-
ditionally, however, the Eurocurrency rate is nominally risk-free. At the begin-
ning of each month an investor knows precisely what the Eurocurrency interest
rate will be over the following month. Hence, its conditional variance is zero;
the Eurocurrency rate is conditionally efficient, but not unconditionally. This is
sufficient to show that CMV -..fo UMV .•
This notion gives rise to the HansenlRichard (1987) critique, similar to Roll's
(1977) critique a decade earlier. Take the capital asset pricing model as an ex-
ample. The model originally implies a version of a conditional factor model
mt+1 = a, + b:f'+1 .93 However, the traditional approach is to test whether the
model can explain the cross-section of average returns using average risk. 94
Given the example in table 2.1, it is clear that the capital asset pricing model
may quite well be (conditionally) true, but fail these tests. The same is true for
unconditional tests of the consumption capital asset pricing model. These tests
strongly reject the model, but this should not discourage to explore the condi-
tional implications of the consumption-based model. Hence, a conditional ver-
sion of a (global) consumption-based asset pricing model in a beta-pricing
framework is tested in chapter 5. Unfortunately, the econometrician cannot ob-
serve the true information set fl,. Therefore, the Hansen-Richard critique posits
that conditional asset pricing models are not testable. Unobservable information
cannot simply be omitted in tests of conditional models. The best the financial
econometrician can do is to derive implications conditional on meaningful vari-
ables that are observable at the beginning of each period. Hence, even if the
wealth portfolio was observable (ignoring Roll's critique), the capital asset
pricing model cannot be tested because the full information set remains unob-
servable. The remaining question is how a (pragmatic) second best solution
looks like, that is, how to optimally incorporate conditioning information in em-
pirical tests of asset pricing models. There are two main strands in the literature:
(a) the instrumental variable technique (in an expected return-beta framework)
and (b) adding scaled payoffs and scaled factors (in a stochastic discount factor
framework). Both techniques are used in the empirical chapters of this book.95
Factor pricing models tie the discount factor m to some other data than con-
sumption. The standard approach in empirical research is to construct linear
factor pricing models, i.e. to replace the consumption-based expression for mar-
ginal utility growth with the following linear model:
(2.92)
where a and b are free parameters. The challenge is the choice of actual factors.
Ultimately, this is an empirical question, but from an economic point of view all
candidates must be good proxies for aggregate marginal utility growth, since:
A U/(C'+I) b/f
I-' u' (C,) == a + ,+1·
(2.93)
In other words, there must be a good economic story behind every chosen factor.
The most straightforward approach is to relate consumption to returns on indi-
ces, to interest rates, to growth in real activity, and to other macroeconomic
variables. All these variables measure the 'state of the world', with consumption
being the purest measure of bad states. Investors consume less in bad states and
more in good states. Low consumption reveals that this is a time at which in-
vestors would especially like portfolios not to do badly, and would be willing to
pay a premium to ensure that wish. In a nutshell, all factor models are a speciali-
zation of the consumption-based model. This has been overlooked in recent em-
pirical research, which mainly concentrated on beta pricing models. 96 All factor
models are the consumption-based model plus some additional assumptions that
95 Detailed discussions of both methods follow below. In particular, the tests in chapter 5
rely on the instrumental variable technique, while the tests in chapter 6 exploit the im-
plications of adding scaled returns in a discount factor framework.
96 An important exception is Cochrane (1996).
The theory of asset pricing 63
where Rm,'+1 denotes the return on the market or wealth portfolio. Standard deri-
vations of the capital asset pricing model assume quadratic utility or normally
distributed returns. 99 DybviglIngersoll (1982) were the first to derive a stochastic
discount factor representation of the capital asset pricing model. 1OO They start by
expanding the Euler-equation into the product of expectations plus the covari-
ance term. Rearranging gives:
(2.95)
Equating terms in (2.94) and (2.95) and some algebra reveals that the traditional
notation of the capital asset pricing model in (2.94) is equal to:
(2.96)
101 Many more asset pricing models are special cases of the Euler-equation, each defined
over a particular specification of the discount factor. For an overview, including non-
separability and habit, see FersonlJagannathan (1996). Empirical studies are by Dunnl
Singleton (1986), FersoniConstantinides (1991), FersonIHarvey (1992), and BraunlFer-
soniConstantinides (1993).
102 Fama (1970) shows that it is only if both the future consumption and investment op-
portunities are not state-dependent that the capital asset pricing model holds in an in-
tertemporal context.
The theory of asset pricing 65
agent will face, i.e. shifts in the investment opportunity set. Accordingly, the
intertemporal capital asset pricing model gives a tighter restriction on the choice
of factors: they must have some prediction power. This notion provides the natu-
rallink to tests of conditional asset pricing models with time varying compensa-
tions for exposure to fluctuations in the investment opportunity set.
To derive Merton's (1973) 'intertemporal security market line', note that opti-
mal consumption decisions are a function of an L-dimensional vector of state
variables s" hence C, = g (s,) .103 Substituting consumption out of the pricing
kernel (and omitting the time discount factor) results in:
u'(g( S'+I))
m =-'--'--":"':"" (2.97)
,+1 u'(g(s,))
E,(u'(C'+I)C~+I) ( )
E , ( r1,,+1 ) == I
L
E ('(C» cov '),1+1'
s r1,,+1 ' (2.98)
.FI ,u ,+1
where L denotes the number of states and C~ the partial derivative of the con-
sumption function with respect to the element sJ.l' The condition in (2.98) posits
that the conditional excess return on any asset i is linearly related to the condi-
tional covariances of the asset's return with the state variables. Further transfor-
mation leads to:
L
E,(u'(C,)C:) ( )
---'-----'-var s r
E, (u' (C,)) 'J.I+I' •. '+1 '
To elaborate the pricing result, it is assumed that there are portfolios with returns
perfectly correlated in a conditional sense with the L state variables, hence fac-
tor mimicking portfolios. I06 The excess returns on these 'hedge portfolios' are
collected in the L-dimensional vector rhp.I+ I ' The expected return on the j-th
hedge portfolio is:
E,(u'(CI+I)C:+I) ( ) ( ) ( )
E, ( rhpj.,+1 ) = E ( '(C )) p, sj.'+l'rhpJ.,+1 cr, SJ.'+I cr, rhPJ.'+1 ' (2.100)
,u ,+1
where cr, (.) denotes the conditional standard deviation and p, (.) is the condi-
tional correlation coefficient. Combining (2.98) with (2.1 00) yields:
Since coy, (s J.,+I, r,.'+1 ) = cov, (rhPJ.I+ I' r..'+1 )cr, (s ].1+1 )/ cr, (rhPJ.I+I)' expression (2.101)
can be transformed to arrive at the discrete version of Merton's (1973) intertem-
poral pricing condition in terms of the familiar beta coefficients:
L
Merton's model is often confused with Ross' (1976) arbitrage pricing theory
(APT), which has been briefly introduced in section 2.9.1. In empirical work,
the differences are negligible. I09 However, in theory the two models should be
clearly distinguished. In particular, factor structure can employ factor pricing (as
in the case of the arbitrage pricing theory), but factor pricing does not require a
factor structure (as in the case of the intertemporal capital asset pricing model).
In contrast to the arbitrage pricing theory, there is no assumption that factors fin
107 For related intertemporaJ asset pricing models see Long (1974), Brock (1982). Coxl
IngersolIIRoss (1985), Bossaerts/Green (1989), and DuffielZame (1989).
108 See CampbelUCochrane (l999b).
109 See Shanken (1985).
68 Chapter 2
m = b'f describe the covariance structure of returns in the derivation of the in-
tertemporal capital asset pricing model. To arrive at (2.102), the factors do not
even have to be orthogonal or i.i.d. High explanatory power in the time series
regression (2.69) may imply factor pricing, but it is not necessary. To derive the
intertemporal security market line in (2.102), the factor model can have as low
an R-square as one wishes. The reason is as follows: the arbitrage pricing theory
starts with a statistical analysis of the covariance matrix of returns and identifies
portfolios that characterize common movements, while the intertemporal capital
asset pricing model starts with identifying good proxies for marginal utility
growth, i.e. state variables that describe the conditional distribution of the in-
vestment and consumption opportunity sets. IIO The arbitrage pricing theory is
more of a 'fishing license'. In fact, the arbitrage pricing theory should hold natu-
rally for portfolios that have high explanatory power in the time-series regres-
sion. In other words, as R-squares go to one, the price of the idiosyncratic com-
ponent goes to zero. At the same time, as the number of primitive assets in-
creases, the R-square of well diversified portfolios increases to one. I II This is a
problem in empirical tests as demonstrated only recently by Ferson/Sarkissiani
Simin (1999). They find evidence for an 'alphabet effect'. Specifically, a well
diversified spread portfolio that is long in early-in-the-alphabet firms and short
in late-in-the-alphabet firms exhibits good cross-sectional explanatory power for
a sample of U.S. stock returns. This is highly suspicious, because the alphabet
attributes are completely unrelated to any source of macroeconomic risk. ll2 The
problem is less severe for the intertemporal capital asset pricing model. Factor
mimicking portfolios are projections of plausible state variables on the space of
returns. In other words, state variables have to forecast something.
llO Cochrane (l999a) argues that the work by ChenIRolIlRoss (1986) should be regarded as
a test of the intertemporal capital asset pricing model, while the three-factor model by
FamaiFrench (1993) is more a test of the arbitrage pricing theory. The FamaiFrench
factors are sorted on size and book-to-market just as the test assets, implying extremely
high R-squares in the time series. Therefore, much of the explanatory power is due to
'common movements' in assets captured by the factors. Nevertheless, ChenIRolIIRoss
motivate their paper as a test of the arbitrage pricing theory, while FamaiFrench inter-
pret their model as an intertemporal capital asset pricing model.
1 I 1 See Cochrane (1999a), p. 136. This refers to approximate arbitrage pricing in the spirit
of Chamberlain (1983), ChamberiainIRothschild (1983), Ingersoll (1984), and others.
112 See Daniel!Titman (1998) for similar evidence.
The theory of asset pricing 69
Asset pricing theory attempts to understand and predict the prices or values of
claims to uncertain payments. The two challenges are to properly account for
time and risk. A low price of an asset implies a high rate of return, so asset
pricing can equivalently be understood as determining a 'fair' expected return
on a risky asset. The most important concept in asset pricing theory is the trade-
off between risk and return. Hence, the central task of asset pricing is to under-
stand and measure the sources of macroeconomic risks that drive asset prices.
Conditional asset pricing allows for expected returns to vary across time and
across assets in ways that are linked to macroeconomic variables, or variables
that forecast macroeconomic events. A wide class of models posits that a 'reces-
sion' factor ultimately determines asset prices. Theoretically, consumption is the
purest measure of bad states. Breeden (1979) argues that "[ ... ] always, when the
value of an additional dollar payoff in a state is high, consumption is low in that
state, and when the value of additional investment is low, optimal consumption
is high. This is not always true for wealth, when investment opportunities are
uncertain".113 The implication is that asset are valued by their marginal contri-
bution to future consumption and not wealth.
While there are many differences among asset pricing models, there are three
main commonalties. First, the law of one price has to hold. Two claims with the
same future payoffs must sell for the same price. The second basic insight com-
mon to all asset pricing models is the no-arbitrage principle. Market forces tend
to align prices, making it impossible to incur a positive probability of a gain
with no chance of a loss by buying and selling different assets. Finally, the third
central concept is financial market equilibrium. A necessary condition for finan-
cial market equilibrium is that the first-order condition of an investor's optimi-
zation problem is satisfied. This approach - captured by the Euler-equation - is
the common language that allows to study pricing issues for basically all types
of assets: stocks, bonds, and derivatives. Intuitively, the equilibrium condition
requires that the investor is indifferent to small changes of his or her asset hold-
ings. Together with a market clearing condition, the Euler-equation stands as the
where P, is the price of a risky asset at time t, XI+I denotes the asset's uncertain
payoffs, and ml+ l is the stochastic discount factor (which indicates the valuation
of payoffs in different states of nature). The two equations are universal; in the-
ory they apply to stocks, bonds, and options. In the standard power-utility model
m'+l is related to the growth rate of aggregate consumption. While this model
does not perform very well when confronted to real data (this is the famous eq-
uity premium puzzle), it is inconsistent to simply discard the model and assume
that the traditional factor pricing stories holds, but the consumption-based model
as a class does not. Both the capital asset pricing model and its intertemporal
analogue are derived as specializations of the discount factor model, that is, by
choosing m'+l as a function of data and parameters. This explains the divergence
of this chapter from the usual order of textbook presentations. They start with
portfolio theory, mean-variance frontiers, spanning theorems, the capital asset
pricing model, the intertemporal capital asset pricing model, and finally the con-
sumption-based model. In other words, finance textbooks proceed in the order
these concepts were discovered and introduced into the literature. This chapter
follows Cochrane (1999a) and starts with the consumption-based model, main-
taining the stochastic discount factor framework throughout. It is then shown
that all other concepts can be traced back to specializations of the fundamental
pricing equation P = E(rnX). For example, starting from the Euler-equation,
one can still find mean-variance optimal portfolios, but this is no longer a central
issue.
71
Chapter 3
The global environment is characterized by several factors that are distinct from
a purely domestic setting. It is necessary to identifying these differences before
translating the pricing concepts introduced in the previous chapter into an inter-
national context. Most important, different countries have different commercial
and monetary policies. Barriers to the movements of goods, labor, and capital
are the results of national commercial policies. National monetary policies in-
troduce different currencies. Therefore, exchange rate risk and barriers to capital
flows are the two major aspects of international financial theory. These aspects
of the global environment will be referred to as differences in consumption and
investment opportunity sets throughout this chapter. Closely related is the styl-
ized fact that the purchasing power parity is grossly violated. Hence, people
from different countries have different appreciations for real returns. This chap-
ter presents the classical approaches to extend domestic asset pricing models to a
global context. As a by-product, international asset pricing models provide the
benchmark returns for tests of capital market integration. Finally, in an interna-
tional context, there is evidence that the correlation of consumption across
countries is extremely low, whereas it should be equal to unity in complete and
integrated markets. This 'consumption home bias' is related to the 'equity home
bias', the observation that domestic investors hold too few foreign assets relative
to what standard results of international portfolio theory suggest.
The remainder of this chapter is as follows. In section 3.1 both the consumption
home bias and the equity home bias are described. It is demonstrated that the
prediction of highly correlated consumption rates across countries, despite of
low correlations of output rates, is clearly rejected by international data. There is
reason to believe that the two biases are related with each other. Section 3.2
72 Chapter 3
starts with the notion that investors care about real returns, and that they want to
be compensated for their exposure to inflation risk. The main idea behind the
derivation of the international capital asset pricing model is the assumption that
purchasing power parity is violated. This implies that investors from different
countries do not have the same appreciation for the real return of a given asset.
In other words, the distinction between domestic and international asset pricing
models can equally be stated in nominal versus real tenns. The section continues
with important results from international portfolio theory. The simple prediction
of the domestic capital asset pricing model that investors simply hold a combi-
nation of the risk-free rate and the market portfolio does not translate into the
international context (unless very restrictive assumptions are imposed). Specifi-
cally, investors hold hedge portfolios to protect themselves against purchasing
power deviations. Aggregating across countries yields the pricing conditions of
the international capital asset pricing models by Solnik (1974b), GrauerILitzen-
berger/Stehle (1976), Sercu (1980), and Adler/Dumas (1983). Section 3.3 intro-
duces the most general model, the international consumption-based asset pricing
model by Stulz (1981). Assuming integrated global capital markets where in-
vestors can share their risks by buying and selling foreign securities, this model
falls out immediately from the results in chapter 2. Finally, the section briefly
demonstrates the translation of beta pricing models into a global framework.
Solnik (1983) and Ikeda (1993) develop an international arbitrage theory, de-
riving a beta pricing condition that holds irrespective of the numeraire currency.
Two puzzles have been proposed in a domestic context: the equity premium
puzzle by MehralPrescott (1985) and the risk-free rate puzzle by Weil (1989).1
Of course, these two puzzles remain unsolved even in a global context. Unfortu-
nately, two additional puzzles arise in an international framework. The most ap-
parent is the 'equity home bias, which refers to the observation that the propor-
tion of foreign stocks held by domestic investors is too small relative to the pre-
dictions of standard portfolio theory. The second puzzle is the 'consumption
home bias', which concerns the lack of risk sharing observed in common con-
sumption movements.
C (s, ) + C' (s, ) = y (s, ) + y' (s, ) for all s, ' (3.2)
where 'A. and (1- 'A.) are the social planner's weights on Home and Foreign. The
output levels of the tradable good in Home and Foreign are denoted with Y and
Y', respectively. 1t(s,) is the probability of state s,. The first order conditions
with respect to consumption are:
(3.5)
The prediction of the complete market model that different countries' consump-
tion rates should be highly correlated even if growth rates in output are not, per-
forms miserably when confronted with actual data. This is the consumption
home bias, the lack of risk sharing observed in consumption comovements
across countries. Panel A in table 3.1 shows the summary statistics of G7 (Can-
ada, France, Germany, Italy, Japan, the United Kingdom, and the United States)
real consumption rates of non-durable goods plus services over the period from
1973.01 to 1998.03. 5 Seasonally adjusted, quarterly consumption data is taken
from the database of the Organization for Economic Co-operation and Devel-
opment (OECD).6 It is evident that annual consumption growth rates have been
considerably high over the last three decades (which gives rise to the risk-free
rate puzzle). As discussed in chapter 2, consumption is a very smooth time se-
ries. Indeed, this seems to be a global phenomenon, as shown by the reported
standard deviations. 7 This observation is the ultimate rationale for the Mehra!
Prescott (1985) equity premium puzzle.
Table 3.1
Consumption growth
Consumption correlations
Panel A also reports the coefficients of skewness and kurtosis, as well as the re-
sults of a Jarque-Bera test under the null hypothesis of normality. There is clear
evidence for departures from normality in three of the G7 countries. Panel B of
table 3.1 exhibits pairs of correlation coefficients. The upper-right half of the
matrix reports the correlations between consumption growth rates, the lower-left
half shows the same correlations between industrial production growth rates.
Quarterly data for national industrial production indices are also taken from the
OECD database. Consumption correlations are low, typically less than 0.3. On
the other hand, industrial production correlations are significantly higher, well
above 0.3 in general. This is counterintuitive: if different countries experience
different output shocks, output should exhibit low correlations across countries.
However, investors should diversify (hedge) their domestic output risk by pur-
chasing and selling contingent claims on other countries' output. Consumption
growth rates would then be highly correlated even when output growth rates are
not. 8 The empirical evidence in table 3.1 shows that the opposite is true.
Several explanations have been suggested for the home bias in consumption.
Unfortunately, all of them seem inappropriate unless very restrictive assump-
tions are imposed. For example, one strand in the literature extends the above
model to allow for non-tradable goods. 9 Another line of research argues that the
gains from diversifying consumption shocks across countries are smaller than
the costs incurred from doing so. In fact, Lucas (1987) calculates that the gains
from international risk sharing might not be substantial. He derives a simple ex-
pression for the percent increase 't in annual consumption that has the same
positive utility effect as the total elimination of consumption uncertainty. This
equivalent variation is given by 't = (yvar( ~C) )/2, where y is the coefficient of
relative risk aversion. The average standard deviation of consumption growth in
table 3.1 is an annual 1.6 percent. Even with y =10, total elimination of con-
sumption variability would be worth only 0.13 percent of annual consumption to
a representative investor. If the gains from complete elimination of unpredict-
able consumption variability are so small, the gains from perfect international
risk pooling - which still leaves people to face systematic global consumption
risk - must be even smaller. 10
Possible explanations for the equity home bias are manifold. An obvious possi-
bility is that domestic equities provide a better hedge for risks that are specific to
the home country. Examples are hedges against domestic inflation and hedges
10 See von Wincoop (1994) and Tesar (1995) for further discussions.
11 See TesarlWerner (1995).
12 Admittedly, the argument is somewhat loose. Over the last decade, the Swiss stock
market has been a top performer compared to other markets. In fact, it is only hard to
show that it would have been beneficial for a Swiss investor to hold foreign stocks dur-
ing a more recent time window. Even with the long sample period in figure 3.1 starting
in 1980, the differences in mean returns are very small.
78 Chapter 3
against wealth that is not traded in capital markets, such as human capital. Ex-
ploring the role of non-hedgeable human capital, Baxter/lermann (1997) show
that it rather worsens the problem than helps to solve the equity home puzzle.
They argue that human wealth is highly correlated with domestic stock returns,
but not with foreign stock markets. Hence, domestic investors should not only
hold foreign shares, but also short the domestic market portfolio to invest an
even greater fraction of their wealth in foreign stocks. Another strand of the lit-
erature suggests to incorporate empirical uncertainty into the analysis. Measur-
ing the gains from international diversification crucially depends on accurate
measurement of expected returns. Given the highly volatile behavior of stock
returns, however, measuring average returns from past data is extremely unreli-
able. Therefore, the null hypothesis of equal mean returns (and standard devia-
tions) in different markets is only hard to reject with sufficient statistical preci-
sion. Tests in this direction are modem versions of so called 'spanning tests'. 13
There is an obvious link between the equity home bias and the consumption
bias. If investors hold too few claims on foreign assets, they will not optimally
share risks with foreigners. Then, as proposed in (3.5), marginal rates of substi-
tution are not equalized internationally. If domestic investors do not hold suffi-
cient claims on foreign output, country-specific consumption movements are
expected to be positively correlated with country-specific output movements as
well. In other words, the deviation of domestic consumption from world con-
sumption is likely to be positively correlated with the deviation of domestic out-
put from world output. Portfolios that are biased toward domestic equities gen-
erate consumption that is biased toward domestic output. While this intuition is
correct in principle, it can be misleading. In fact, a home bias in equities is nei-
ther sufficient nor necessary for a home bias in consumption. 14 To see that home
bias in equities is not sufficient, consider a counterexample with only interna-
tional borrowing and lending, but no trade in equities. Home bias in equity is
extreme in such a setup; there are no holdings of foreign equities at all in either
country. However, if output shocks across countries are transitory, consumption
growth rates should still be highly correlated. Telmer (1993) and HeatonlLucas
(1996) demonstrate that the desire to smooth consumption tends to mitigate the
consumption home bias. IS In particular, almost all of an investor's idiosyncratic
income risk can be diversified through trade in a single risk-free asset. In a nut-
shell, risk sharing in complete markets is largely duplicated under incomplete
markets, implying that the equity home bias is not sufficient to generate con-
sumption home bias.
Figure 3.1
The equity home bias
Period: 1980.0 I - 1998.08 (monthly data)
14.3%
14.2%
100% Home
14.0% (SwItzerland)
13.9%
17.0% 18.0% 19.0% 20.0% 21.0%
Standard deviation p.a.
The figure shows the properties of portfolios composed of Swiss and North American stocks over the
period from 1980.01 to 1998.08 in mean-standard deviation space. Monthly data is obtained from
Morgan Stanley Capital International (MSCI). Moving along the curve from 100% in Swiss stocks to
100% in North American stocks, the line exhibits mean returns and standard deviations from holding
an increased proportion of foreign stocks. All numbers are calculated in Swiss francs and are in per-
cent per year.
At the same time, an equity home bias does not necessarily imply a consumption
bias. Even if domestic investors are optimally invested in foreign equities, some
domestic output cannot be securitized if the space of output realizations is not
15 See the discussion on heterogeneous agents and incomplete markets in section 2.6.
80 Chapter 3
spanned by traded claims. Examples are claims on human capital and services.
In this case, domestic consumption might be correlated with country-specific
shocks to domestic output. A consumption bias can exist although there is no
measured equity home bias. These examples demonstrate that the strength of the
link between equity home bias and consumption home bias depends on the as-
sumptions about the structure of the economy. The link is strongest when mar-
kets are complete (that is, when equity is traded on all output in an economy)
and when there is a representative agent in every country.16 This type of model
predicts perfectly correlated consumption rates across countries, a nation that is
clearly rejected empirically, as shown in table 3.1. The more fundamental prob-
lem, however, is that consumption performs poorly explaining the patterns of
historical stock returns. 17 Therefore, as long as the equity premium puzzle re-
mains unsolved and a unified theory of consumption and equity prices is not
available, the link between the equity home bias and the consumption home bias
will remain weak as well.
3.2.1 The nominal and the real capital asset pricing model
The capital asset pricing model (CAPM) predicts that the expected real return r,
on asset i can be written in terms of the covariance of the asset with the real re-
turn on the wealth portfolio rm :20
(3.6)
where T] denotes the real risk-free rate of return and y is the representative in-
vestor's coefficient of relative risk aversion. The real and the nominal version of
the capital asset pricing model are strictly related. The real rate of return on asset
i is given as:
r. = 1+ R, -1 (3.7)
, 1+1t '
where R, denotes the nominal return on asset i, and 1t is the inflation rate. Ap-
plying the rules of stochastic calculus and substituting into (3.6) results in:
Rewriting the covariance term on the right-hand side and rearranging yields:
Applying the Fisher-equation and the Ito-Lemma, the first four terms of the
right-hand side in (3.9) sum to the nominally risk-free rate of return Rr .21 The
equation in (3.10) is the nominal capital asset pricing model. Investors care
about real returns and, hence risky inflation produces a separate premium on
nominal returns. Even if investors are risk averse, hence y =0, investors require
compensation for inflation risk. Because they do not suffer from money-illusion
and evaluate returns in real terms, the coefficient on the covariance with infla-
tion and the coefficient on the covariance with the market sum to unity. In con-
trast, if inflation is non-stochastic, the nominal capital asset pricing model in-
cludes the market risk premium only.
(3.11)
where 7t~~; is the Swiss rate of inflation measured in Swiss francs. The same
applies for the real return r;s on the same asset i from the perspective of a U.S.
investor, denoting e~~~ as the relative change of the spot rate measured in U.S.
dollars per Swiss franc:
USD = (1 + 7tCHF
1+ 7tUSD CHF )(1 + e CHF
USD) <=> r;CHF = r;USD (3.13)
In a nutshell, if purchasing power parity holds, all investors use the same price
index. This implies that the real returns on a security are the same for investors
in different countries. In contrast, violations of purchasing power parity create
heterogeneity across investors in the way real returns are appreciated in different
countries. Real returns on the same security then differ to the extent that the
price indices investors use to translate into the numeraire currency are different.
This also implies that the composition of optimal portfolios depends on an in-
vestor's residence. Finally, note that the results of this chapter could equally be
interpreted in terms of real versus nominal rather than international versus do-
mestic asset pricing models.
scribe expected returns. Without loss of generality, the Swiss franc is chosen as
the numeraire currency in the derivation. Each country or each national group of
consumer-investors 1 is assumed to have its own (possibly stochastic) rate of in-
flation. In contrast, if purchasing power parity holds, there is only a single world
rate of inflation. Denoting the rate of inflation of country 1 measured in the cur-
rency of country 1 with 1t:, and the rate of inflation of country 1 measured in
Swiss francs with 1tfHF, (3.13) can be rewritten as:
(3.14)
where e~HF is the relative change in the exchange rate (now measured as Swiss
francs per unit of currency 1). The capital asset pricing model in nominal terms
in (3.10) can now be translated into an international setting. Specifically, an in-
ternational capital asset pricing model (in Swiss francs) implies for each national
group I the following first-order condition for expected returns 23 :
where R f denotes the nominally risk-free interest rate (in Swiss francs), and R~
is the Swiss francs return on the portfolio held by the investors of country 1.
Specifically, R~ = I, w,'R, , with w,' being the weight allocated by residents of
country 1 to security i. Sercu (1980) shows that a change of the numeraire cur-
rency implies a mere change of the intercept term in (3.15): because investors do
not suffer from money illusion, the nominally risk-free interest rate on Swiss
francs Rr is replaced with the rate of interest on any currency one translates
into, denoted as R; .24 In fact, this invariance with respect to the reference cur-
rency ensures that a truly international pricing condition can be derived.
It is now possible to recover the vector of optimal portfolio weights {wi}. For
this purpose, define with f1 the Nxl vector of expected returns on risky assets.
This vector contains L foreign currency bills and M other risky assets, called
stocks. 25 Also denote with (In(CHF,') the Nxl vector of covariances between asset
returns R, and 1tfHF, and with 0 the NxN variance-variance matrix of nomi-
nal asset returns. Solving for the vector of portfolio weights Wi in expression
(3.15) yields:
where 1 denotes an Nxl vector of ones. Portfolio weights must always sum to
'unity, which implies that the amount invested in the nominally risk-free asset is
given as:
WI=( w;_
1-1 Wi
)=~( ~-'(p-Rr1)
l' 1-10-1 (p-Rr1)
)+(I-~X ~-IOlt(CHF.I)
l' 1-1 n-I01t(CHF.I)
). (3.18)
Equation (3.18) decomposes the optimal portfolio into two funds, with weights
IN and 1- (IN). The first set of weights:
O-'(p-Rrl) )
Wlog = ( I-t'O-'(p-Rrl) , (3.19)
is the asset allocation that would be selected by an investor with unit risk aver-
sion. This is an investor with logarithmic utility function?7 This class of inves-
tors does not care about prices and, hence the term O'lt(CHF.I) is not included in the
optimal weights. Most important, the log-portfolio is independent of 1, the coun-
try of residence. Hence, the weights in (3.19) denote the universal component of
the optimal portfolio. The second set of weights:
is the asset allocation an investor with infinite risk aversion would choose. In-
tuitively, in real terms such an investor only invests in the global minimum-
variance portfolio. Indeed, the weights in (3.20) support this notion. The expres-
sion n-'(J1t(CHF.,) denotes a vector of ordinary least square coefficients obtained
from a projection of the inflation rate 1tfHF on the nominal returns R, from the
N risky assets. This portfolio exhibits the maximum correlation with home in-
flation, thereby providing the best hedge against country 1 inflation (measured in
Swiss francs). In other words, w~ denotes the weights of an 'inflation mimick-
ing portfolio', or a global minimum-variance portfolio in real terms. 28 Note that
the domestic bond does not suffice for this purpose, it is not risk-free in real
terms. Expected returns f1 do not enter into (3.20), hence variance reduction is
really the only purpose of the hedge portfolio. Finally, because investors will
hold different hedge portfolios, the world market portfolio should no longer be
mean-variance efficient.
(lntCAPM).29 Accounting for deviations from purchasing power parity and sto-
chastic inflation, there are as many inflation premiums as national investor
groups. Intuitively, an asset whose return is correlated with deviations from pur-
chasing power parity earns a risk premium. Note that the sum is over all 1
(1 == 1,2, ... , L) , implying that Swiss investors add a term to the pricing condition
like a~y other foreign agent. This is because domestic inflation is stochastic in
this general framework. Finally, for the international capital asset pricing model
to be a viable theory, the pricing condition in (3.21) must be invariant to the
choice of the numeraire currency. Dumas (1994a) shows that this is indeed the
case. For any currency of denomination j the following pricing condition holds:
1
E(R;) == Rf + YLII ( y;--1
)WI
Wcov(R; ,1tI)+ycov(R;,R~). (3.22)
This pricing condition is identical to the one in (3.21), but with a different inter-
cept Rf (the nominal risk-free rate of il1terest in currency j) instead of Rr (the
risk-free rate of interest in Swiss francs).3o
The model in condition (3.21) has several testable implications. To see this, the
Adler/Dumas (1983) version of the international capital asset pricing model can
be written as follows:
L
E(R,) == a + LAI cov(R" 1t~HF) + 3cov(R"Rm), (3.23)
1=1
where the L A'S measure the prices of both domestic inflation risk and exchange
rate risk (note from (3.14) that the various inflation rates 1t(HF do not only reflect
the random fluctuations in exchange rates but also the randomness in local cur-
rency inflation rates), and 0 denotes the price of market risk. The testable re-
strictions of this model are: (1) the intercept term a should be equal to the
nominally risk-free rate in the numeraire currency, (2) under the assumption of
risk aversion, 3 should always be positive, and (3) the sum LIAI + 3 has to be
equal to one. 31 Recall from section 2.10 that an unconditional pricing condition
29 See the discussion on aggregation problems in AdlerlDumas (1983), p. 949f, and their
resulting pricing condition (16), p. 95l.
30 For the derivation see Dumas (1994a), p. 309.
31 See AdlerlDumas (1983), p. 950.
88 Chapter 3
implies a conditional one, but not vice versa. Hence, to arrive at a fully condi-
tional factor model all moments in (3.21) can be conditioned on information
available to the investor, that is, E, (.) and cov, (.) are given time-subscripts.
Similarly, the prices of inflation risk 1..1.'+1 (1 = 1,2, ... ,L) and the price of market
risk 0'+1 are allowed to vary over time. Models with time varying currency and
market risk premiums have been tested empirically only recently. Dumas/Solnik
(1995) apply the Generalized Method of Moments (GMM) in the instrumental
variable framework originally introduced by Harvey (1991a). This technique
allows both first and second moments to change over time, but only the first
moments are explicitly parameterized. In contrast, the time evolution of vari-
ances and covariances cannot be extracted. The multivariate GARCH-in-the-
mean technique applied by De Santis/Gerard (1998) specifically accounts for
changing conditional second moments. 32 However, note that the AdlerlDumas
(1983) version of the international capital asset pricing model only accounts for
changes in the consumption opportunity set, but does not incorporate any antici-
pated variation in the investm~nt opportunity set in the spirit of Merton (1973).33
Accordingly, conditional covariances with the relevant state variables should be
included into the pricing condition in (3.21). Empirically, this seems impossible
to accomplish given today's estimation techniques. 34 Stulz (1984) argues that
additional premiums might be captured by the fact that deviations from pur-
chasing power parity are correlated with changes in other variables affecting the
investor's expected lifetime utility.35
Sercu (1980) generalizes this model by relaxing some of Solnik's restrictive as-
sumptions. 36 In the Solnik-Sercu framework the returns on the home currency
deposit are risk-free in real terms. If asset demands are computed in a common
numeraire currency, all investors have a perfect hedge against purchasing power
deviations. In fact, the hedge portfolio collapses to 100 percent of the investor's
home currency bil1. 37 To verify this intuition, adjust the weights in (3.20) for the
assumption that domestic inflation is equal to zero:
I _( n-l(Je(CHF.I) )
Wh - (3.24)
1-1'nI
- (Je(CHF.I)
'
where <J e (CHF.1) is the N x 1 vector of covariances between asset returns and the
exchange rate. Recall that the vector of N assets also includes Swiss franc re-
turns on the bills denominated in foreign currencies. Let rd(l) be the real return
on the deposit of country 1, measured in currency 1, and let R~nr be the nominal
return on the same deposit, but measured in Swiss francs. By definition:
I (l+R~nn(l+e~HF)
rd(l) = I 1. (3.25)
1+ Itl
Since rd = 0 , it also holds that:
1 + rd(l) =(1 + R~nr) (1 + e~HF ) . (3.26)
The weights in (3.24) are obtained by regressing the changes in the exchange
rate e~HF on all securities, including R~(~r. According to (3.26), the random
component of R~nr is equal to e~HF' hence the fit of this regression is perfect in
the sense that the coefficient on R~nr will equal unity. All remaining assets do
not load on the exchange rate changes, so their regression coefficients are zero. 38
As a result, in the absence of home inflation, the hedge portfolio reduces to the
home deposit. The optimal weights in (3.18) can then be written as:
(3.27)
36 For example, Solnik (1974) assumes that stock returns in their own currency are uncor-
related with exchange rates.
37 See the derivation in lorion/Khouri (1996).
38 This intuition is taken from De Santis (1997).
90 Chapter 3
In a nutshell, the Solnik-Sercu special case posits that the optimal portfolio can
be decomposed into a universal log-portfolio containing all risky assets and the
domestic risk-free bill (in real terms) as a hedge against purchasing power risk.
The relative amount invested in each component again reflects the investor's
individual risk aversion. 39
This version of the model also demonstrates the dual role of currencies. Note
that forward contracts are redundant assets because a long (short) forward posi-
tion can be replicated by a long (short) position in the foreign bill financed with
a short (long) position in the domestic bill. 4o Starting at (3.27), Sercu (1980)
shows that the optimal currency positions depend on the optimal stock positions,
which themselves are affected by the presence of (non-zero) expected excess
return on currencies in the portfolio. 4J This implies that the log-portfolio can
further be decomposed into a portfolio of stocks optimally hedged against cur-
rency risk and a portfolio of speculative positions in bills. The demand for for-
eign bills can be decomposed into a part which serves to hedge the stock posi-
tions, and a part which serves to maximize the portfolio's risk adjusted return
(i.e. the demand which would obtain if stocks were already hedged). The de-
mand for the domestic bill can be decomposed into the demand that would ob-
tain if stocks were already hedged, and a part which represents the proceeds of
the (likely) short positions in foreign bills that were contracted to hedge the
stocks. 42 This gives rise to a generalized 'three' -fund theorem: all investors hold
(i) their country's riskless bond, (ii) a portfolio of hedged stocks, and (iii) a port-
folio of bonds of all currency denominations. 43
There are two important practical implications. First, full hedging neglects the
speculative component of currencies and, hence is not optimal from a theoretical
point of view (see Glen/Jorion (1993) for empirical evidence).44 Second, curren-
cies should be regarded as an independent asset class, which offers a specific
risk-return spectrum to an investor. AdlerlPrasad (1992) and Jorion (1994) argue
that stock and currency positions should be optimized simultaneously. In par-
ticular, 'overlay hedges' - where a core portfolio of equities is selected first, and
currency risk is managed separately in a second step - are only second-best so-
lutions because the core portfolio is chosen without reference to currencies.
Third, the determination of universal (equilibrium) hedge-ratios requires very
strong assumptions. Assuming that all investors have the same risk aversion,
Y = y, and that each national wealth is exactly equal to the value of each stock
market, Black (1990b) shows that the optimal hedge ratio reduces to a universal
value 1- (1/y). Given that reasonable levels of risk aversion are between 2 and
4, the universal hedge ratio should be somewhere between 50 percent and 75
percent. 45
cov(R rrCHF
I
I' ) = cov(R e CHF
I )
I' . (3.28)
44 For an opposite view see PeroldiSchulman (1988). They claim that hedging is a 'free
lunch', reducing the volatility ofreturns without affecting average returns. For empirical
support see LevylLim (1994).
45 For a critique see Solnik (1993a) and AdlerlPrasad (1992).
46 See table 4 in AdlerlDumas (1983), p. 946.
47 For a critique see Solnik (1993a) and AdlerlPrasad (1992).
92 Chapter 3
In this special case, the covariance terms in (3.21) denote exposure to currency
risk only, and the A coefficients can accordingly be interpreted as premiums for
currency risk. 48 Swiss investors do no longer contribute a term to the pricing
condition denominated in Swiss francs, because domestic inflation is zero. The
corresponding Swiss covariance term drops out, leaving only L -1 premiums
for currency risk. Hence, the restriction that the sum IIAI + 0 has to be equal to
one in (3.23) is no longer testable. There is evidence that exchange rate risk is
priced into international equity returns both conditionally (see Dumas/Solnik
(1995) and De Santis/Gerard (1998» and unconditionally (see Vassalou (1997».
Alternatively (and somewhat pragmatic), Harvey (1995) constructs a single cur-
rency factor by weighting returns on positions in foreign bills. Finally, Sercu
(1980) shows that this version of the model is equivalent to a decomposition of
the total risk premium into (i) the expected cost (benefit) of hedging stocks
against exchange risk and (ii) a risk premium for uncertainty not associated with
exchange rate changes similar to the domestic capital asset pricing model, i.e. a
risk premium on the hedged market portfolio. 49 This, of course, is just another
representation of what accounts for the primary driving force in international
asset pricing models, namely the demand for hedging stocks against purchasing
power risk. Similarly, AdlerlDumas (1983) show that for international equities
the capital asset pricing model applies using returns that are hedged against ex-
change rate risks. 50
chasing power parity holds. Real exchange rate risk disappears. Prices can be
evaluated in real terms all over the world because country inflation rates col-
lapse into a single inflation rate when expressed in a cornmon reference cur-
rency. In a nutshell, an asset's expected real return on any asset is the same for
all investors. This gives rise to a capital asset pricing model in real terms with a
simple pricing condition as follows:
A E()
E( r. ) == I-',m .h A cov(r.,rm)
rm == (3.29)
WIt I-',m
var (rm) '
where r. now denotes the one-period real return on asset i in excess of the risk-
free rate, and rw is the real excess return on the world market portfoliO. 51 ~,w
denotes the real market beta, i.e. the sensitivity of the real return on asset i to the
real return on the world market portfolio. The model in (3.29) does not carry any
country subscripts, hence it applies to whatever numeraire currency chosen. As
with the domestic capital asset pricing model (but under admittedly unrealistic
assumptions), a single-beta relation is implied in an international context.
Hence, the condition in (3.29) is referred to as an international version of the
capital asset pricing model. In nominal terms, the GrauerlLitzenbergerlStehle
(1976) framework collapses the AdlerlDumas (1983) model into:
(3.30)
which holds for any numeraire currency. All variables are defined above. The
special feature of the pricing condition in (3.30) is that it contains - in addition
to the covariance of nominal returns on asset i with the market portfolio - only a
single covariance term of the returns on asset i with the global inflation rate 1t.
Although this model holds only under unrealistic assumptions, it has the merit of
showing that (i) exchange rate risk is related to deviations from purchasing
power parity and (ii) the distinction between domestic and international asset
pricing models is qualitatively equal to stating the same models in real versus
nominal terms.
51 See GrauerlLitzenbergerlStehie (1976), p. 240, and Stulz (1984), equation (3), p. 57.
94 Chapter 3
The international capital asset pricing models proposed by Solnik (1974), Sercu
(1980), and AdlerlDumas (1983) require many critical assumptions. In a domes-
tic as well as an international context, the consumption-based model stands as
the fundamental approach to price risky assets. All other models are merely spe-
cializations of consumption-based asset pricing.52 Stulz (1981) shows how to
accomplish global aggregation under deviations from purchasing power parity to
derive an international consumption capital asset pricing model (IntCCAPM).
As another specialization of this model, Solnik (1983) and Ikeda (1991) moti-
vate international beta pricing models. In particular, they show that a given fac-
tor structure holds even after a change of the numeraire currency, implying an
international arbitrage pricing theory (IntAPT).
where c' denotes the real aggregate consumption rate. A translation of the pric-
ing condition from real terms into nominal terms yields:
E(R.) - E( 1t) + var( 1t) - COV(Ri, 1t) = Tl + ycov(R. -1t,C -1t), (3.32)
where Ri again denotes the nominal return on risky asset i, and c is the nominal
consumption rate. Using the Fisher equation and summing terms to the nominal
interest rate Rr gives an expression that is similar to (3.10):
This model holds under very general assumptions about exchange rate dynamics
and from the point of view of any national group of investors. 53 Stulz (1981) di-
rectly works in real terms and examines the aggregation of (3.31) over all na-
tional investor groups. Specifically, he shows that the expected real excess re-
turn on risky asset i is proportional to the covariance of the real returns on that
asset with changes in the world real consumption rate:
cov(r,,~) ( ) ()
E ()
r, = () E To = ~ICE ~ , (3.34)
var ~
where 1C denotes the real excess return on a portfolio whose real return is
maximally correlated with the growth rate of real world (aggregate) consump-
tion that is consumed by the global representative investor. 54 In complete mar-
kets, this 'consumption mimicking portfolio' does exist. ~IC denotes the associ-
ated proportionality coefficient, that is, the consumption beta. Technically, sys-
tematic risk in this model is measured relative to changes in the level of real ag-
gregate world consumption. The underlying rationale for the model in (3.34) is
intuitive. Assuming complete markets implies perfect risk sharing and, hence
perfect correlation of consumption rates across countries (as shown in (3.5».
But this is equivalent to saying that capital markets are fully integrated. 55 Hence,
the pricing condition in (3.34) holds for all investors in all numeraire currencies,
even under deviations from purchasing power parity.
Finally, it is important to note that the model in (3.34) is the most general inter-
national asset pricing model. While AdlerlDumas (1983) assume a constant in-
vestment opportunity set, the Stulz (1981) framework allows for both the con-
sumption and the investment opportunity sets to vary over time. Again, every-
thing boils down to the Euler-equation, E( mR) =1. Asset prices must adjust
until the investor is indifferent with respect to investing or consuming one addi-
tional money unit. In fact, as argued throughout chapter 2, this first-order condi-
tion for the optimal investment-consumption choice holds in a multi-period set-
ting, even if investment and consumption opportunity sets change over time. It
only requires that the investor has access to the securities to which the condition
is applied, implying perfect risk sharing. It has not been assumed at any stage of
the derivation, however, that consumption and investment opportunity sets must
be constant. 56 In fact, given perfect risk sharing in complete (integrated) global
asset markets, the model in (3.34) falls out immediately from what has been said
in chapter 2 about consumption-based asset pricing in a domestic framework.
where R, denotes the return on asset i denominated in Swiss francs, and the as-
sociated expectation is E(R j) . .t; is the random change of the j-th (mean-zero)
factor, ~'J denotes the loading of security i on factor j (j = 1,2, ... ,K). Hence, the
model in (3.35) assumes that there are K pervasive factors, whose changes drive
the returns on risky assets and capture all systematic risk. Finally, £, is an error
term. The arbitrage pricing theory exploits the notion that in large capital mar-
kets it is possible to construct arbitrage portfolios that do neither exhibit system-
atic nor idiosyncratic risk in the sense of a factor structure. Ross (1976) uses this
arbitrage argument to show that the expected return on asset i must be a linear
combination of K factor loadings:
K
E(R,)=a+ L~'JAJ' (3.36)
j=l
where a denotes the nominally risk-free rate in Swiss francs, and AJ is the risk
premiums on thejth pervasive factor, with j=I,2, ... ,K. 57 The arbitrage pricing
theory can be extended to the international environment if it is invariant to the
56 See Stulz (1984), p. 65, and Wheatley (1988), pp. 178 and 184.
57 See expression (2.76).
Theory of international asset pricing 97
choice of the numeraire currency. The problem with the pricing condition in
(3.36), however, is that the presence of exchange rate risk in an international
context introduces an additional country-specific term into the analysis which
might adversely affect the arbitrage activities in capital markets. 58 Basically, the
following conditions are required for international beta pricing models to hold:
(i) the risk arising from exchange rate shifts must be diversifiable like any other
unsystematic risk, (ii) an arbitrage portfolio that is riskless in any given currency
must be riskless in any other currency, and (iii) the factor structure must be in-
variant to the choice of a currency. 59 While the conditions (i) and (ii) are some-
what technical prerequisites, the condition (iii) must be satisfied for the interna-
tional arbitrage pricing theory to be a viable theory.
Assume that asset markets are perfectly integrated. For the special case in which
exchange rate fluctuations have the same K-factor structure as the asset returns,
Solnik (1983) shows that an arbitrage portfolio that is nominally riskless in one
currency is riskless for all investors, irrespective of their numeraire currency.
Consequently, the factor structure is also invariant to the currency of denomina-
tion, but the coefficients in (3.36) vary with the numeraire currency.60 Ikeda
(1991) relaxes the assumption of a K-factor structure for exchange rate changes
and allows for a random process. In this case, expected returns are not invariant
to the currency of denomination unless they are adjusted for by the cost of ex-
change risk hedging. He suggests constructing an arbitrage portfolio of invest-
ments in the risky assets that are fully financed by taking opposite positions in
national bonds. From interest rate parity, this portfolio is fully hedged against
exchange rate risks. The usual no-arbitrage condition can then be invoked for
hedged portfolios. Hence, the pricing condition of the international arbitrage
pricing theory for the expected return on asset i from country I, measured in
Swiss francs, implies: 61
K
E(R,) -{R~ + E( efHF)} =: ~)"J~:J . (3.37)
J=I
R~ denotes the interest rate on the risk-free bond in the country 1, measured in
the currency of this country. efHF is the exchange rate movement of country I's
currency in terms of the Swiss franc. A] is the premium a Swiss investor can
expect per unit of exposure on a hedged foreign investment in country 1 with
respect to the j-th source of risk. In contrast to the Solnik (1983) specification,
the factor exposures ~~] in Ikeda's pricing condition are generally measured in
terms of the currency of country 1. This pricing condition holds for any nu-
meraire currency. This property guarantees that condition (iii) as stated is satis-
fied, i.e. the arbitrage pricing theory can be extended to an international context
and, hence represents a viable theory. Conditions (i) and (ii) hold implicitly in
Ikeda's (1991) derivation. Note that the left-hand side of (3.37) represents the
expected return in Swiss francs on a risky asset for which the exchange risk is
hedged by foreign borrowing and lending. The cost of hedging is given by the
second term on the left-hand side of the pricing condition in {.} .62
To sum up, for a cross-section of international assets whose returns are fully
hedged against exchange risk by taking short positions in local bills, the interna-
tional arbitrage pricing theory provides a pricing condition that holds for all in-
vestors, irrespective of their residence. In fact, the demand for exchange risk
hedging in Ikeda's model is very similar to the mechanism in the Solnik-Sercu
special case of the international capital asset pricing model. 63 It is also important
to note that the international arbitrage pricing theory is subject to the same factor
identification problems as its domestic version. It might be hard to identify in-
ternational risk factors that simultaneously affect both asset and currency re-
turns. Empirically, testable specifications of the international arbitrage pricing
theory look very similar to the AdlerlDumas (1983) pricing condition in (3.23).
In fact, both the international arbitrage pricing theory and the international
capital asset pricing model are only hard to distinguish empirically.
62 In fact. this is the same mechanism also at work in Sercu (1980). equation (11). p. 104.
See the discussion in the final part of section 3.2.5.
63 See section 3.2.5.
Theory of international asset pricing 99
International asset pricing models differ with respect to the assumptions made
on consumption and investment opportunity sets in different countries. The
translation of domestic asset pricing models to an international context is com-
plicated by the fact that purchasing power parity does not hold. This implies that
investors from different countries have different appreciations for the real return
on a given asset. Consequently, this leads to different portfolio compositions in
different countries. The main streams in international asset pricing theory are as
follows:
Chapter 4
Following FaIna's (l970a) dictum, asset markets are said to be efficient if secu-
rity prices fully reflect all available information. Of course, this strong form of
efficiency has been subject to considerable critique from two main strands. First,
trading and gathering information are not costless. Grossman/Stiglitz (1980) ar-
gue that efficient markets should reflect relevant information only to the point
where the marginal benefits of acting on information do not exceed the marginal
cost. If markets are perfectly efficient, the return on getting information is nil.
There would be little reason to trade and markets would eventually collapse. In
other words, the degree of market inefficiency determines the effort investors
are willing to expend to gather and trade on information. A non-degenerate mar-
ket equilibrium will arise only when there are sufficient profit opportunities, i.e.
inefficiencies to compensate investors for the cost of trading and information
gathering. Second, for a long time financial economists used to think that market
efficiency implies that returns follow a random walk. Indeed, the old market ef-
ficiency-constant expected return model seemed to perform well in the early lit-
erature. 1 However, the argument has come under attack both theoretically as
well as empirically more recently. In particular, it is only under the strong as-
sumption of constant expected returns that the random walk model is inevitably
implied by the traditional efficient market hypothesis. First opposing empirical
evidence by FamalFrench (l988a) and PoterbalSummers (1988) suggested that
stock returns measured at long horizons exhibit strong negative autocorrelation.
More recent work shows that returns are predictable on the basis of instrument
variables related to real activity, such as term spreads, default spreads, industrial
production, and the dividend yield, among others. As LolMacKinalay (1999) put
it, "[ ... ] financial markets are predictable to some degree, but far from being a
symptom of inefficiency or irrationality, predictability is the oil that lubricates
the gears of capitalism".2
Two explanations have been asserted in the literature. First, predictability could
be evidence of market inefficiency due to some form of general or limited irra-
tionality, such as fads, speculative bubbles, or noise trading. Return predictabil-
ity reflects irrational deviations of prices from their fundamental values. Shiller
(1981) and Campbell/Shiller (1988b) argue that stock returns exhibit too much
variation to be rationally explained. LakonishoklShleiferNishny (1994) find that
value strategies yield higher returns because these strategies exploit suboptimal
behavior of the representative investor and not because they are fundamentally
riskier. They argue that investors extrapolate past growth rates too far into the
future, without accounting for mean reversion. Although the literature on 'irra-
tionality' is substantial, there is no parsimonious and empirically robust general
equilibrium model supporting the inefficiency argument. 3 Second, a large num-
ber of studies argue that expected returns vary with the business cycle. 4 As Fama
(1991) concludes at one point in his survey: "If the variation through time in ex-
pected returns is rational, driven by shocks to taste or technology, then the
variation in expected returns should be related to variation in consumption, in-
vestment, and savings".5 A number of empirical studies documents that time
variation in economic risk premiums can explain a substantial fraction of return
predictability.6 In fact, intertemporal asset pricing models posit that predictabil-
ity is perfectly consistent with the concept of market efficiency.7 Balvers/Cosi-
manolMcDonald (1990) and CheccettilLamlMark (1990) show that stock prices
need not follow a random walk in a rational valuation model. They argue that
equilibrium stock returns can be predicted to the extent that there is predictabil-
ity in the endowment process. Evidence of trend reversion in aggregate output
This chapter does not yet focus directly on asset pricing issues. Instead, an ad-
hoc attempt is taken to deepen the understanding of the link between time varia-
tion in expected returns and business conditions, as encouraged by Fama (1990).
In particular, I try to shed new light on simple instrumental forecasting models.
Past studies have demonstrated that appropriate information variables can serve
as indicators of recent and future real activity. Return predictability on the basis
of these variables is interpreted as evidence that expected stock returns are
closely related to the business cycle. 9 However, regressing returns on lagged in-
strument variables might not reveal the whole story about a possibly complex
relationship. Critics argue that there are data mining biases in any detected rela-
tionship. IO In other words, predictive regressions could simply capture noise
without any meaningful economic interpretation. Therefore, it is interesting to
examine whether the relationship between expected returns and instrument vari-
ables is driven by the slow-moving, medium frequency business cycle compo-
nents of the instrument variables. The main contribution of this chapter is to ex-
tract and isolate low-frequency, medium-frequency, and high-frequency compo-
nents of stock price and global economic instrument series to study the links by
frequency bands. If Fama's (1990) business cycle story has some merits, strong
common movements should be found in the cyclical frequencies, i.e. the time
series components associated with the business cycle.
This chapter is organized as follows. Section 4.1 first presents the standard in-
strumental variable forecasting model and discusses its economic interpreta-
tions. The last part of this section introduces the Baxter/King (1998) filter-
technique, designed to extract business cycle components from macroeconomic
time series. Section 4.2 gives a guide on how to choose proper economic in-
8 See chapter 7.
9 For contrasting empirical evidence see Eleswaraguffiwari (1996).
10 For example, Lo/MacKinley (1990) and FosterlSmithlWhaley (1997).
104 Chapter 4
In standard financial valuation models stock price volatility stems from a com-
bination of (i) shocks to expected cash flows, (ii) predictable variation due to
variation in discount rates, and (iii) shocks to expected returns. The financial
econometrician must be concerned with exploring these three sources of return
variation. In fact, many studies find that a large fraction of stock return variation
is associated with forecasts of real gross national product, industrial production,
and investment that determine cash flows to firms. II There is also evidence that
expected returns vary through time. Therefore, asset pricing models must cap-
ture any time variation in the discount rates that price expected cash flows. 12 Fi-
nally, some empirical studies demonstrate that part of the variation in stock re-
turns is caused by shocks to the discount rates that generate opposite shocks to
prices.13 I will ignore this last effect, but explore the first two sources of stock
return variation in this chapter within the framework originally introduced by
Fama (1990) and Schwert (1990). Their approach allows determining whether
stock price volatility is 'rational' in the sense that available information is re-
flected in the valuation at each point in time. 14 Both studies first examine how
well lagged (real) stock returns forecast current industrial production, and then
invert this equation to see how returns are influenced by various leads of indus-
trial production (interpreted as perfect rational expectation forecasts of real ac-
II For example, see Fama (1981), GeskelRoll (1983), Kaul (1987), Barro (1990), and Lee
(1992).
12 Pioneering studies in this direction are Keirn/Stambaugh (1986), FersonlMerrick (1987),
Campbell (1987), CampbelUShiller (1988a), and FamalFrench (1989).
13 See French/SchwertlStambaugh (1987) and GlostenlJagannathanlRunkle (1993).
14 See also Chen (1991) for a similar approach.
Time varying expected returns and the business cycle 105
(4.1)
where RP.~~i denotes the risk premium in country i, R~~~ the total stock return
in country i, and r~HF is the risk-free rate of interest, all denominated in Swiss
francs. The expectation operator E is conditioned on the set of available infor-
mation at the beginning of each period t, denoted as <Ill. The Swiss franc risk
premium on a foreign asset is approximately equal to the national market risk
premium (denominated in domestic currency) plus the respective currency risk
premium. In contrast, the realized (ex-post) Swiss franc excess return X~~~ in
country i is given by:
(4.2)
Together, equation (4.1) and (4.2) imply that period-by-period stock returns are
predictable up to a mean zero random term. This is perfectly consistent with ra-
tional valuation, implying that stock prices reflect all available information at
any point in time.
(4.3)
where the Swiss franc superscript is omitted for notational convenience. Asset
pricing models that generate linear conditional expectations are parametric in the
sense that some distributional characteristics are either explicitly or implicitly
assumed. To test whether the approximation in (4.3) is appropriate, Harvey
(199Ic) experiments with nonparametric density estimation. Intuitively, if there
was only a slightly nonlinear relation between stock returns and information
variables, kernel regressions should lead to an increase in explanatory power.
Interestingly, Harvey reports that - compared to the standard linear model in
(4.3) - forecasts of stock market returns are not improved with nonparametric
density estimation techniques. He concludes that linear conditional expectations
can be used as a reasonable approximation in conditional asset pricing research.
Therefore, using the linear specification in (4.3) in empirical work is not as re-
strictive as it might look in the first place.
risk-free rate if the asset provides a hedge against adverse changes in their con-
sumption. Technically spoken, excess returns and the stochastic discount factor
(i.e. the marginal rate of substitution) need not be negatively correlated in all
states of the world, implying that a positive ex-ante risk premium is not a neces-
sary condition for capital market equilibrium. Tauchen/Hussey (1991) demon-
strate that this outcome is not very plausible using reasonable assumptions and
parameter values, but it remains an empirical question as to whether this condi-
tion actually occurs at some points in time. Asness (1991) finds that there are
only a few periods with negative conditional consumption betas in the United
States over the last decades. Specifically, these periods are brief and coincide
with business cycle peaks. Using a technique to incorporate inequality con-
straints in the Euler-equation framework, BoudoukhlRichardsonlSmith (1993)
show that the U.S. market risk premium has indeed been negative in some states
of the world. These states are characterized by high expected inflation rates
(high T-bill rates) and inverted yield curves. 15 Ostdiek (1998) replicates these
results for the ex-ante world market risk premium. She provides evidence that
the global market risk premium can be negative, both denominated in U.S. dol-
lars and local currency.16 A final concern is that any reasonable 'fit' of the linear
regression model might come as the result of data snooping or data mining
rather than a close approximation of the true conditional expectation. However,
the cross-country evidence demonstrated below seems to invalidate this cri-
tique.17
of the economy. If the economy is doing well, an agent's welfare will be greater
than if it is doing badly, even if his of her level of wealth is constant. Accord-
ingly, the intertemporal capital asset pricing model (ICAPM) posits that the de-
mand for risky assets is made up of two components: (i) the mean-variance
component as in the static portfolio optimization problem and (ii) a demand to
hedge adverse shocks to the investment opportunity set. The resulting pricing
equation contains multiple betas, with the number of betas equal to one plus the
number of state variables that drive the investment opportunity set over time. 19
In other words, the intertemporal capital asset pricing model is a linear factor
model with wealth and state variables that forecast changes in the distribution of
future returns or income. Unfortunately, the analysis remains silent about the
identity of the state variables, except that they have to forecast something. As
argued in chapter 2, there is a crucial link between return predictability and the
hedging motives of investors. Recently, Campbell (1996) argues that priced
factors should neither be found by running a factor analysis on the covariance
matrix of returns nor by selecting macroeconomic variables. Instead, the eco-
nometrician must look at the time series behavior of stock returns, that is, the
factors' importance in forecasting future returns. Campbell concludes that the
appropriate instrument variables (as proxies for the true underlying state vari-
ables) must be closely associated with changes in forecasts of real activity to
yield rational predictions for stock returns. Different sensitivities of asset returns
to the underlying state variables that generate time varying market returns cause
risk premiums on the assets to differ. Therefore, my general working hypothesis
follows along Fama/French (1989) and Chen (1991) and consists of two parts:
(i) state variables that are positively (negatively) related to the recent growth of
the economy are negatively (positively) related to expected excess market re-
turns, and (ii) state variables that are positively (negatively) related to the future
growth of the economy are positively (negatively) related to expected excess
returns.
models, the standard power utility function can be maintained. Second, the life
cycle-permanent income hypothesis stands as the mechanism that links aggre-
gate savings to transitory changes in aggregate output/income. 22 Supposing that
rational investors anticipate high future levels of endowment, both the usual in-
tertemporal relative price effect and the substitution effect are at work simulta-
neously. Assuming that the utility function is more concave than implied by log-
utility, the latter effect dominates. In particular, agents want to smooth con-
sumption and try to increase current consumption by selling part of their asset
holding. This results in instantaneously falling stock prices and higher expected
returns. The other way round, investors anticipating lower future output in the
next period will attempt to smooth consumption by transferring wealth to this
anticipated period of scarcity, thereby accepting a lower rate of return. This is
only consistent with higher current stock prices. The Lucas-type exchange econ-
omy with a perishable consumption good ultimately depends on this link be-
tween consumption smoothing and risk-correction?3 My second working hy-
pothesis follows exactly along this intuition: state variables that are positively
(negatively) related to future growth rates of the economy are positively (nega-
tively) related to expected excess stock market returns. Most important, Bal-
vers/CosimanolMcDonald (1990) emphasize that predictability cannot induce
agents to exploit it and gain excess profits. To do so, they would have to give up
current consumption to invest in stocks during a recession, that is, at a time
when both the risk premium and the marginal utility of consumption are high.
This, however, can be shown to be inconsistent with utility maximizing behav-
ior. Any possible advantage from higher returns is exactly offset by the disad-
vantage of a less smooth consumption pattern. In equilibrium, expected stock
returns adjust in a way to guarantee that the average investor just holds the mar-
ket. In other words, multiple factors and return predictability cannot have any
portfolio implications for the average investor.
To sum up, the general notion is that variations in expected returns are due to
changes in the business cycle as reflected in the state variables. Stock returns are
closely tied to fluctuations in aggregate output/consumption. In the next chapter
22 See ModiglianifBrumberg (1955), Friedman (1957), and the discussion in section 2.2.
23 See chapter 7 for an attempt to explicitly model this mechanism.
Time varying expected returns and the business cycle III
The choice of proper instrument variables follows along two dimensions. First,
instrument variables have to be correlated with the business cycle. Second, de-
pending on whether a segmented or integrated view of the global stock markets
is appropriate, the choice is among local and global instrument variables.
nomic variables are not available right at the end of a sampling period. In addi-
tion, many macro-series are seasonally filtered using centered moving averages,
invalidating a rational expectation forecast of stock returns. Even worse, macro-
data are subject to substantial revisions. Most important, economic news is fil-
tered by investors when forming expectations, that is, depending on the market
consensus expectation, similar innovations in macro-variables can imply 'bad'
or 'good' news at different points in time. Therefore, I follow the traditional ap-
proach in the literature and use variables that are 'internal' to the stock market.
ces reflect a degree of internationalization that is not typical for the average
stock in the index. LombardIRouletiSolnik (1999) construct a purely domestic
index that is free of 'internationalization biases' and run horse-races against
global factors. Interestingly, for a sample of Swiss stocks they find that the dif-
ferentiation between domestic and multinational companies is very pronounced.
While domestic betas are high for companies with a local focus in their activities
and low for multinationals, international exposures are low for the smaller do-
mestic firms and very high for multinational companies. Cross-sectional regres-
sion results provide further evidence that stock price movements heavily depend
on the degree of a company's international activities. Therefore, assuming that
stock markets are globally integrated, I choose to work with a set of global in-
strument variables only. Specifically, the world dividend yield, a G7 term
spread, the TED spread, a G7 short-term interest rate, and the lagged world mar-
ket return to capture any residual predictable part of returns are used in the em-
pirical analysis that follows.
where lowercase letters denote the logarithm of a variable. The last tenn on the
right side is a nonlinear function of the log dividend price ratio. A first-order
Taylor approximation around the average log dividend price ratio (d - p) gives:
where P == 1/(1 + exp( d - p)) and k == -log(p) - (1- p) log (1 / P -1) are parameters
of linearization. The Taylor approximation in (4.5) replaces the log of the sum
of the stock price and the dividend in the exact relation with a weighted average
of the log stock price and the log dividend with p and (1- p) as the weights.
Equation (4.5) is a linear difference equation in the log stock price. Assuming
the transversality condition lim j-t_ pip,+ J =0 to rule out 'rational bubbles', using
expectation operators, and solving forward one obtains:
and finally,
(4.7)
Equations (4.6) and (4.7) describe a dynamic version of the well known Gordon
growth model (Gordon/Shapiro (1954)). It is not a theory, it is a pure accounting
identity. A high current stock price must necessarily be followed by high future
dividend growth L1d, and/or a low expected return r. Equivalently, the log divi-
dend yield is high if expected dividend growth is low and/or expected return is
high. Most important, dividend price ratios can only vary if they forecast
changing dividend growth and/or changing expected returns. Otherwise, the
dividend yield would have to be constant. Cochrane (1999b) argues that an ap-
parent difference in yields should give rise to an offsetting movement, but this
does not seem to be the case. Something should be predictable, so that returns
are not predictable. In particular, a high dividend yield should mean that divi-
dends grow more slowly over time. However, this tendency seems to be com-
pletely absent. Using a variance decomposition argument, Cochrane (1991c)
shows that almost all variation in price dividend ratios has been due to changing
return forecasts. Dividend/price ratios do not seem to forecast dividend growth.
118 Chapter 4
In addition, FamalFrench (1989) observe that the dividend/price ratio takes its
highest values during recessions, though the major swings seem to be beyond
the business cycle. They argue that any forecasting ability of the dividend/price
ratio "reflects time variation in expected returns in response to business condi-
tions that tend to persist beyond measured cycles".33 Similarly, projecting the
dividend yield on several leads and lags of quarterly growth rates of gross na-
tional product, Chen (1991) concludes that the dividend/price ratio is an indica-
tor of the current health of the economy (high in recessions and low in booms),
but does not contain information about the future health of the economy. Ac-
cording to my working hypothesis, a positive relationship between expected re-
turns and the dividend/price ratio can be expected in predictive regressions.
G7 term spread: The economic story to justify the term spread as an appropriate
instrument variable builds explicitly on the life cycle hypothesis of Modiglianil
Brumberg (1954) and the permanent income hypothesis of Friedman (1957). In
fact, it is a natural candidate to proxy for the underlying state variables. Recall,
investors prefer a smooth consumption stream rather than very high consump-
tion at one stage of the business cycle and very low consumption at another
stage. Consumption smoothing drives the demand for insurance or hedging. A
natural way to do so is to substitute bonds of different maturities, as proposed by
Harvey (1991 b). 34 If the economy is in a growth stage, but a general slowdown
is expected, people will hedge by buying assets that deliver high payoffs during
the future economic downturn. For example, an investor could purchase long-
term government bonds and simultaneously sell short-term bonds to hedge. If
many investors follow this strategy, the price of long term bonds increases, im-
plying decreasing yields. On the other hand, the selling pressure for short term
bonds will drive down prices and increase yields. As a result, the term structure
flattens or even becomes inverted. Indeed, Harvey (1991c) finds that the term
spread is an excellent predictor of future business cycle stages. Similarly, Chen
(1991) reports that an above average term spread forecasts that the gross na-
tional product will continue to increase over the next four to six quarters, and
vice versa. 35 FamaiFrench (1989) report empirical evidence that the term spread
is one of the best variables to forecast stock as well as bond returns. The term
spread is low near business cycle peaks and high near business cycle troughs.
Return forecasts should be high at the bottom of a recession (when the term
spread is high and predicts an economic upturn) and low at the top of a boom
(when the term spread is low and predicts an economic slowdown). According
to my working hypothesis, therefore, a positive relationship between expected
stock returns and the term spread can be expected in the predictive regressions. 36
TED spread: The TED spread is defined as the difference between the 3-months
Eurodollar rate and the yield on the 90-days U.S. Treasury bill. The TED spread
is affected by three major factors: (i) world political stability, (ii) balance of
trade, and (iii) fiscal policy in the United States. When political uncertainty is
high and the risk of disruption in the global financial system increases, the yield
differential widens. When the balance of trade is decreasing, the TED spread
should rise as well. The TED spread can be assumed to be an indicator of the
current health of the economy. The yield differential should be higher during
phases of economic recessions (when investors are seeking safer assets), and
decreases during expansionary phases. This is confirmed by regressing growth
rates of industrial production on various lags of the TED spread. Coefficients are
negative up to five lags. However, the TED spread has no significant forecasting
power beyond the second quarter. 37 Given my working hypothesis, there should
be a positive relationship between expected excess returns and the TED spread.
FamalFrench (1989) observe that interest rates tend to fall during periods of
economic downturns and rise during expansionary phases. Chen (1991) finds
that the current Treasury bill rate can forecast changes in the growth rate of the
gross national product. An interest rate below average forecasts that the level of
the gross national product will continue to increase over the next 4-6 quarters. 41
Return variation due to shocks to expected cash flows: Fama (1990) and
Schwert (1990) report evidence that stock returns are highly correlated with fu-
ture growth rates of industrial production. They first measure how well lagged
real stock returns forecast current industrial production growth and then invert
the equation to see whether leads of industrial production are able to explain
stock returns. 42 Given that expectations are rational, information about the pro-
duction of a given period is spread across proceeding periods and so affects the
stock returns of proceeding periods. Recall that financial securities are claims
against future output. The relation between current stock returns and growth
rates of future industrial production reflects information about future cash flows
that are discounted in today's stock prices. A measure of the future productivity
of capital - such as the expected growth rate of the aggregate economy - should
therefore be positively correlated with the expected stock market return accord-
ing to my working hypothesis.
Most studies in this direction are limited to U.S. data, international evidence is
scarce so far. 43 Following Fama (1990) and Schwert (1990), I use leads of quar-
terly (logarithmic) growth rates of global industrial production over the coming
four quarters as rational expectation proxies for expected future cash flows. The
following regression is estimated for each national stock market:
given period is spread across several preceding periods, also affecting stock re-
turn expectations. Leads of industrial production are used as rational expectation
proxies for future cash flows to companies and, hence to explain expected stock
returns.
Of course, the relationship between stock returns and economic activity is possi-
bly a lot more complicated. In fact, interpreting future production as a proxy for
changing expectations about future cash flows might be only one side of the
complete story. FisherlMerton (1984) note that corporate profits are an impor-
tant part of gross national product and are also likely to be positively correlated
with other components of gross national product. The forward-looking property
of stock prices would therefore qualify the stock market as a predictor of the
business cycle. Indeed, they show that the stock market generally starts falling in
the quarter before a recession, as determined by the National Bureau of Eco-
nomic Research (NBER).44 FischerlMerton (1984) conclude that the stock mar-
ket is the single best leading indicator. Barro (1990) also finds that lagged
changes in real stock market prices have great explanatory power for the growth
rate of investment and dramatically outperforms Tobin's q.45 Finally, stock re-
turns might even actively cause changes in real activity. Gains on the stock mar-
ket increase wealth, which in tum is likely to increase the demand for consump-
tion and/or investment goods. ShleiferNishny (1986) present a model where the
information in stock returns influences companies' investment decisions. How-
ever, distinguishing between the stock market as a mere predictor of the busi-
ness cycle and the stock market as a causal factor in the business cycle seems
impossible to accomplish empirically. Therefore, as already stressed above, it is
not the goal of this chapter to disentangle the exact cause and effect in the rela-
tion between stock returns and real activity. Rather, the goal of the empirical
analysis is to better understand how agents form their expectations about the
stock market. Fama (1990) takes a pragmatic approach and argues that "[ ... ] as
long as the return variation that results from the relations between stock returns
and real activity is rational, it is a legitimate part of the story for rational varia-
tion in stock returns".46
Figure 4.1a
Frequency components of instrument variables
45
40
35
30
2.5
1
76 78 80 82 84 86 88 90 92 94
08
02
04
00
-U4
-U8
76 78 80 82 84 S6 8S 90 92 94 -U 47"J.6"1"'1 'T'"f
78",,",S"'0""S"'2"'""84'"'"'S"'6""S"'S"'""90"""'9""2""94
1998.03 are used. The indices are all value weighted aggregates calculated on a
total return basis, including the dividends paid. I postpone a detailed description
of stock returns until chapter 6, when emerging stock markets are also included
into the analysis. Following previous work by FersonIHarvey (1993) and Oert-
mann (1997), three of the four instrument variables incorporate globally aggre-
gated information. To construct global variables if they are not directly available
from Datastream, I aggregate national data on the instrument variables across
the G7 countries (Canada, France, Germany, Italy, Japan, the United Kingdom,
and the United States), using the countries' relative shares of total G7 gross do-
mestic product (in real terms) as the weights. The weights are updated annually.
Figure 4.1b
Continued
G7 tenn spread
0.010 0.004
0.005 0.003
0.002
0.000
0.001
.(l.005
0.000
.(l.010
.(l.001
.(l.015 .(l.002
76 78 80 82 84 86 88 90 92 94 76 78 80 82 84 86 88 90 92 94
0.006 0.006
0.004 0.004
0.002 0.002
0.000 0.000
.{).002 .{).002
.{).004 .{).004
'{).006 .{).006
76 78 80 82 84 86 88 90 92 94 76 78 80 82 84 86 88 90 92 94
Figure4.1c
Continued
TED spread
0014 0006
0012 0.005
0010
0004
0008
0003
0.006
0002
0004
0002 0001
0000 0000
76 78 80 82 84 86 88 90 92 94
0003 0006
0002
0.004
0001
0002
0000
0000
-0001
-0002 -0002
-0003 -0004
76 78 80 82 84 86 88 90 92 94 76 78 80 82 84 86 88 90 92 94
The four panels of figure 4.la-d show the time evolution of both the raw instru-
ment series and the frequency bands as defined above: 'trend', 'business cycle',
and 'irregular' components. The first instrument is the market capitalization
weighted dividend/price ratio, as provided directly by Datastream. The time t
ratio is calculated as the average value of dividends paid over the last 12 months
on the Datastream world market index, divided by the index value at time t. 47 As
shown in table 4.1, the average value of this instrument is 3.249 percent per an-
47 The dividend/price ratio has been used as a forecasting instrument by Ball (l978), Ro-
zeff (l984), CampbeIVShiller (l988a), FamalFrench (l988b, 1989), Chen (l991), Fer-
sonlHarvey (1991), Cochrane (1992), and Evans (l994). A global dividend yield is used
in Harvey (1991, 1995), FersonlHarvey (l993), and Oertmann (1997).
126 Chapter 4
num, with a low standard deviation of only 1.038 percent. The second instru-
ment is a G7 term spread. It is defined as the difference between the yield on
long term government bonds (with maturity of at least five years) and the 1-
month Eurocurrency interest rate in each country, weighted by the time varying
national shares of total G7 gross domestic product. 48
Figure 4.1d
Continued
0.05 0035
004 0030
0.03 0025
0.02 0020
001 0015
000 0010
76 78 80 82 84 86 88 90 92 94 76 78 80 82 84 86 88 90 92 94
0010 0006
0008
0004
0006
0002
0004
0002 0000
0000
-0002
-0002
-0.004
-0004
-0006 -0006
76 78 80 82 84 86 88 90 92 94 76 78 80 82 84 86 88 90 92 94
The figures display the time evolution of the raw series and the different frequency bands of global
instrument variables. The band-pass filter suggested by Baxter/King (1998) is used to identify three
frequencies: 'trend', 'business cycle', and 'irregular' components. By construction, three years of data
are lost at the beginning and at the end of the original sample period.
48 The U.S. term spread as an instrument variable was originally used by KandeVStam-
baugh (1986), ChenIRolllRoss (1986), and Campbell (1987). More recent studies are by
FamalFrench (1989), Harvey (1991a), CampbeIVAmmer (1993), FersonlHarvey (1993),
BrownlOtsuki (1993), Evans (1994), llmanen (1995), and Oertmann (1997). Many of
these studies use global term spreads.
Time varying expected returns and the business cycle 127
The data for the yield on long term government bonds is taken from the database
of the international Monetary Fund (IMF). The time series for Eurocurrency in-
terest rate are from the Bank for International Settlement (BIS). For some G7
countries the series do not cover the entire sample period. For the very early
years the spread is constructed by adjusting the gross domestic product weights
for those countries with available data on interest rates. As shown in table 4.1,
the average value of this instrument is 0.652 percent annually, with a standard
deviation of 0.734 percent. Third, the TED spread is the difference between the
3-months Eurodollar interest rate and the yield on the 90-days U.S. Treasury
bill. Again, data is taken from the Bank for International Settlement (BIS) and
the database of the International Monetary Fund (IMF). As shown in table 4.1,
the average value of this instrument is a yearly 1.200 percent, with a standard
deviation of 0.526 percent. 49
Table 4.1
Descriptive statistics of the instrument variables
Period: 1973.01- 1998.03 (quarterly data)
Level Autocorrelation
Mean in % S.D.in%
PI p, p, P4 p, P. PI2 p"
annual annual
iDYWL 3.249 1.038 0.948 0.885 0.835 0.787 0.737 0.697 0.548 0.147
iTSG7 0.652 0.707 0.541 0.302 0.218 0.051 -0.011 0.048 0.052 0.017
iTED 1.200 0.526 0.683 0.620 0.594 0.461 0.493 0.363 0.111 0.169
iSRG7 7.808 1.506 0.915 0.850 0.786 0.679 0.604 0.529 0.106 -0.044
The table reports the means, the standard deviations (both in percent per year), and the autocorrela-
tions up to lags of 24 quarters of the instrument variables. The instrument variables are the world divi-
dend yield (iDYWL), a G7 term spread (iTSG7), the TED spread (iTED), and a G7 short-term interest
rate (iSRG7).
(BIS) and aggregate them in the way described above. 50 The average value of
this instrument is as high as 7.808 percent per annum, with a standard deviation
of 1.506 percent. Table 4.1 also shows that all instruments are serially corre-
lated. The largest first order autocorrelations are documented for the world divi-
dend yield (0.948) and the G7 short rate (0.915). If daily returns are only slightly
predictable by a slow-moving instrument variable, that predictability adds up
over long horizons. CampbelllLo/MacKinlay (1997) demonstrate that R-squares
in regressions of stock returns on dividend/price ratios increase with the time
horizon 51 . Most previous studies use monthly data, so it is be interesting to see
the amount of predictability in quarterly data.
Table 4.2
Test of the stationarity of instrument variables
Dickey-Fuller model:
LlZh' = a hO + ahIZt,I_1 + a h2 .6Zh ,t-I + ChI
Phillips-Peron model:
DZ" = a hO + a.,Z.,._1 + E h.; for Zh. = 1,2,3,4 (instruments)
iTSG7 0.545
The instrument variables are defined in the text. For the Dickey-Fuller model, t-statistics are given in
italics below each estimated coefficient. The Dickey-Fuller test statistic corresponds to the signifi-
cance level of the first slope coefficient in the Dickey-Fuller test equation. The Phillips-Perron test
statistic is a transformation of the first slope coefficient in the Phillips-Perron test equation, and it is
reported in italics in the last column of the table. *1**1*** implies a rejection of the null hypothesis of
a unit root (non-stationarity) in the time series of the respective instrument. The lower panel of the
table shows the correlations of the stationary series of instrument variables.
130 Chapter 4
This section presents the empirical results. The order of sequence is reversed.
First, I present estimation results of equation (4.8), examining the relation be-
tween future growth rates of industrial production and stock market risk premi-
ums. Second, I show the results of predictive regressions in (4.3). Finally, I ex-
amine the correlations between stock prices and instrument variables by fre-
quency bands. This allows assessing - at least roughly - how much of the pre-
diction power in linear regressions can be attributed to cyclical movements in
both series, as suggested by economic theory. In a nutshell, is the forecasting
power in linear instrumental variable regressions for real or do we merely inter-
pret noise within a pseudo-economic framework?
To give additional insights, table 4.4 reports the results of standard Wald-tests.
Even though the coefficients are not significant in single country regressions,
they could still exhibit significant power to explain the cross-sectional differ-
Time varying expected returns and the business cycle 131
ences in market risk premiums. This possibility can be examined for each of the
four leads separately as well as for all of them together. The two hypothesis to
be tested imply the following two types of cross-sectional restrictions on the
sensitivity coefficients in a system of seemingly unrelated regressions (SUR): (i)
Ho :b j ==0, for j==I,2,3,4 leads of OEeD world industrial production growth
rates across all i == 1, ... ,18 stock markets ('The sensitivities on lead j are jointly
equal to zero across all markets'), and (ii) Ho :b,) ==0, for all i==I, ... ,18 stock
markets and for all j == 1,2,3,4 leads of OEeD world industrial production
growth rates simultaneously ('The sensitivities are jointly equal to zero across
all markets and for all leads of industrial production'). The resulting Wald-test
statistic is chi-square distributed with degrees of freedom equal to the number of
restrictions. The tests are run for the entire sample period and two subperiods of
equal length. For the full sample period and the early subperiod from 1973.01 to
1985.12 only the first two leading rates tum out statistically significant for the
cross-section of expected excess returns. In contrast, the null hypothesis that the
sensitivities on the industrial production rates with leads of three and four quar-
ters are jointly equal to zero over all markets cannot be rejected. For the more
recent period from 1986.01 to 1998.03 all leads of industrial production growth
rates exhibit significant explanatory power for the cross-sectional differences in
expected excess returns. This might indicate an increasing importance of a
'global' business cycle during the last decade. The second null hypothesis, that
all growth rates of industrial production together have no explanatory power for
the cross-section of expected excess returns, is clearly rejected at the 1 percent
level of significance for both the total sample and the two subperiods.
Japan -0.019*
c;
2.330*** -0.381 1.780 0.722 0.
0.111 0.073
-1.332 2.289 -0.283 1.130 1.023
Netherlands -0.003 0.506 1.787** 0.541 1.102** ~
0.156 0.119 3Vl
-0.304 0.668 1.829 0.646 1.639 I>'
Norway -0.015 2.671 ** 2.268 0.230 -1.280 ::s
0.120 0.083 0.
-1.030 1.684 1.235 0.123 -0.823 :;-
n>
Singapore -0.028* -0.079 4.293*** -0.176 0.900 0-
0.104 0.066 t:
-1.137 -0.064 3.323 -0.119 0.659 Vl
S·
Spain -0.021 4.126*** 0.336 0.499 1.199 n>
0.035 -0.006 Vl
en
-1.104 2./63 0./06 0.128 0.464
g;
Sweden 0.002 0.912 0.912 0.507 0.353 0.
0.038 -0.003
0./68 0.970 0.843 0.736 0.314 :;-
n>
Switzerland -0.008 0.059 3.074*** -1.103 1.401 ** 0-
0.143 0.106 t:
-0.888 0.093 3.070 -1.149 1.781 Vl
S·
United Kingdom -0.018 0.027 1.820* 1.496 1.581** n>
Vl
0.133 0.096 Vl
-1.099 0.022 1.474 1.225 1.918 (")
'<
(")
United States -0.012 0.062 1.527 1.212 0.470
0.086 0.047 0-
-0.870 0.083 0.964 1.11 1 0.568
The table reports the coefficient estimates for the regression model in 4.8. The t-statistics reported under the coefficients (in italics) are adjusted for heterosce-
dasticity and autocorrelation. *1**1*** denotes a coefficient estimated on the 20%110%/5% level of significance. The R-square denotes the coefficient of de-
termination, adj. R-square is adjusted for degrees of freedom. 'IP' denotes quarterly logarithmic growth rates of seasonally adjusted world industrial produc-
tion, as reported in the OECD Main Economic Indicator Database.
......
t.;.l
t.;.l
134 Chapter 4
Table 4.4
Testing the cross·sectional influence of leads of industrial production rates
Wald-test hypothesis:
Hyp.l: 'The sensitivities on lead j are jointly equal to zero across all markets.'
Hyp.2: 'The sensitivities are jointly equal to zero across all markets and all leads
of industrial production.'
Chi-square statistics
1973.01 - 1998.03
Subperiods
1973.01 - 1985.12
The world dividend yield (iDYWL) has a strong positive (and statistically sig-
nificant) influence on the stock market risk premium in the majority of coun-
tries. This is consistent with the working hypothesis laid down above: (i) the
mechanical argument from standard valuation models and (ii) the dividend yield
as an indicator of the recent health of the economy. Chen (1991) reports that
above average term spreads forecasts higher growth rates in gross national prod-
uct for the next five quarters. Since the term spread is not related to the recent
136 Chapter 4
(4.9)
where SR' denotes the maximum unconditional Sharpe ratio, and SRo the (an-
nual) unconditional buy-and-hold Sharpe ratio. k refers to the prediction hori-
zon, i.e. 0.25 for quarterly data. Expression (4.9) allows assessing the potential
benefits of market timing on the basis of global instrument variables, but with-
out actually simulating a specific market timing strategy. The historical buy-
and-hold Sharpe ratio for the world stock market was roughly 0.50 on an annual
basis. 59 Plugging both the historical Sharpe ratio and the average cross-country
prediction power (0.143) into (4.9), it turns out that a market timing strategy on
the basis of linear instrumental predictions would have been beneficial for in-
vestors. The maximum unconditional Sharpe ratio increases to an impressive
0.92. Holding portfolio volatility constant, this implies that market timing on the
basis of the model in (4.3) can almost double average annual returns.
However, these results should not be taken for granted. There is a clear tendency
to discover spurious relationships when applying tests that are inspired by evi-
dence from previous visits to the data. Merton (1987) explicitly asks whether it
is "[ ... ] reasonable to use standard t-statistics as a valid measure of significance
when the test is conducted on the same data used by many earlier studies whose
results influenced the choice of theory to be tested?".60 Lo/MacKinley (1990)
investigate data-snooping biases and suspect that grouping stocks into portfolios
might induce a bias in statistical tests. In particular, the selection of securities to
be included in a given portfolio is almost never random, but is usually based on
some of the stocks characteristics. The country indices provided by Morgan
Stanley Capital International (MSCI) are clearly no exception. Given that the
empirical evidence is very similar for different global stock markets, the data-
mining critique is somewhat invalidated. Alternatively, FosterlSmithlWhaley
(1997) adjust standard test statistics for overfitting. Their approach allows to
assess whether the findings in table 4.5 are for real from an economic point of
view, or indistinguishable from merely searching all possible regression combi-
nations given a sufficiently large set of potential regressors. Predictive regres-
sions interpreted on the basis of traditional test statistics assume that a certain
test is performed only once on a particular set of data. While careful studies (in-
cluding the present one) do not search all possible regression specifications and
report the best one only, they do rely on the results of previous studies. Hence,
standard R-squares are biased because the temptation to select the five 'best'
regressors is ignored. FosterlSmithlWhaley (1997) derive an approximate bound
for the joint distribution function of the maximum R-square. 61 Intuitively, the
cutoff R-square increases. Unfortunately, their analysis sheds negative light on
the results in table 4.5. Given the size of my sample (101 quarterly data points)
and assuming ten potential regressors only, the regression R-square must exceed
0.224 to make sure that an exhaustive search of all possible regression combina-
tions would not identify a model with higher prediction power.62 With an even
greater number of potential regressors, the 95 percent confidence interval for the
R-square further increases. In contrast, BossaertslHillion (1999) find that over-
fitting should not be a major concern. They apply various model selection crite-
ria and conclude that there is ample evidence of in-sample predictability. Most
important, the results are consistent for every selection criterion across all stock
markets. Unfortunately, the out-of-sample forecasting power of even the 'best'
models is nil. They conclude that linear instrumental variable forecasting models
are inherently non-stationary. A related problem is addressed by FersoniSar-
kassianiSimin (1999b). They examine the sampling properties of return regres-
sions with persistent regressors. The dependent variables in (4.3) are stock re-
turns, which are not highly persistent. However, returns are assumed to be the
sum of unobserved expected returns plus white noise. If the 'true' expected re-
turns are persistent, but an unrelated instrument is chosen, the danger of a spuri-
ous regression bias is apparent. Given the number of quarterly observations, the
R-squares, and the autocorrelation of the instrument variables (smaller than 0.95
in all series), their simulation results indicate that standard t-statistics should be
reasonably well behaved.63 In other words, the great number of significant t-
Table 4.7 combines the regression models in (4.3) and (4.8). Excess returns are
simultaneously regressed on lagged global instrument variables and the four
leads of global industrial production rates. As expected, the R-squares increase
compared to both tables 4.3 and 4.5. In particular, the average R-square rises
from 0.143 (in table 4.5) to 0.227. Given that the random walk alternative im-
plies an R-square of zero, the fraction of return variability explained by this
simple model is substantial. The signs of the sensitivity coefficients on future
production growth rates turn out mixed, and they are insignificant most of the
time. Only the growth rate of industrial production in the immediate future
maintains its significant (positive) influence on excess stock returns. Most im-
portant and contrary to Fama (1990), none of the global instrument variables is
driven out of the model by leads of industrial production in the combined re-
gression model. Rather, the global instruments seem to absorb most of the in-
dustrial production leads' explanatory power. This can be interpreted as evi-
dence for my working hypothesis. A considerable fraction of the predictable
variation in stock returns is related to the business cycle, but clearly not all of it.
......
~
Table 4.5 10
Singapore 0.073*** 0.095 -29.954*** -11.444* -28.102*** -0.118 2.164 0.141 0.096 ()
'"'"
2.543 1.147 -4.515 -1.468 -3.776 0.756 '<
()
Sweden 0.057*** 0.056* -13.231 *** -7.044* -10.158* 0.100 2.077 0.054 0.004 ~
2.221 1.370 -1.992 -1.267 -1.453 0.682
Switzerland 0.068*** 0.098*** -8.516** -14.159*** -10.848*** 0.0265 2.084 0.281 0.243
5.201 3.165 -2.222 -4.965 -2.602 0.414
UK 0.049*** 0.129** -9.899* -8.957* -13.884*** 0.194* 2.066 0.157 0.113
2.031 1.934 -1.566 -1.473 -2.035 1.562
USA 0.027* 0.115*** -3.413 -6.164 -6.684* 0.232*** 1.862 0.153 0.108
1.378 3.263 -0.709 -1.218 -1.263 2.030
The instruments are defined in the text. The t-statistics reported underneath the coefficients (in italics) are adjusted for heteroscedasticity. *1**1*** denotes a
coefficient estimated on the 20%110%15% level of significance. R-square denotes the coefficient of determination, adj. R2 is adjusted for degrees of freedom.
D.W. is the Durbin-Watson statistic.
.....
.j:>.
.....
142 Chapter 4
Table 4.6
Testing the cross-sectional influence of global instrument variables
Wald-test hypothesis:
Hyp.l: 'The sensitivities on instrument j are equal to zero across all the markets
Hyp.2: 'The sensitivities are jointly equal to zero across all markets and for
all global instruments'
Chi-square statistics
Global instruments
1973.01 - 1998.03
Regressions of excess returns on global instrument variables and future growth rates of industrial production ~
S·
oq
Instrument regressions: n>
><
't:1
r.,t+1 = bill + bL1Z 1,\ + .,. + b,5Z5.1 + b1fiIP,,1+1 + ... + b,';lIPI+3,1+4 + e.,I+l n>
(")
0
Q.
Period: 1973.01- 1998.03 (quarterly data) nl
Sensitivity to global instrument variables Sensitivity to future production growth rates
a
a
Stock markets Intercept iDYWL iTSG7 iTED iSRG7 iWRL IP(O,1) IP(I,2) IP(2,3) IP(3,4) R2 ~
'"
::l
Q.
Australia 0.009 0.127'" -16.855'" -2.519 -14.947**' 0.153 1.169 4.381'" -0.118 -1.477 0.248
So
n>
0.287 1.719 -2.197 -0.356 -1.734 1.077 0.616 2.214 -0.068 -0.835 (0.171)
0-
Austria 0.014 0.028 -9.126' -3.263 -6.466 0.033 1.689* 0.504 -1.234* 0.697 0.050 c:
0.515 0.623 -1.592 -0.664 -1.031 0.338 1.292 0.397 -1.348 0.694 (-0.048) '"S·
n>
Belgium 0.052'** 0.098*** -14.993*** -9.626*** -13.587'" -0.124 1.412 1.003 -0.084 -0.005 0.276 (")
'"'"
2.611 2.216 -3.116 2.29 -2.453 -0.834 1.102 0.837 -0.091 -0.005 (0.202) '<
(")
Canada -0.004 0.111 *** -14.028*** -0.633 -10.736*** 0.142** 2.579** 0.692 2.231* -1.607* 0.217 0-
-0.133 2.305 -2.364 -0.107 -1.489 1.092 1.721 0.456 1.558 -1.318 (0.137)
Denmark -0.025 0.033 -6.005 -0.033 -3.387 0.184** 3.742*** 0.805 0.555 1.667** 0.235
-1.187 0.902 -1.102 -0.010 0.666 1.662 2.760 0.664 0.555 1.754 (0.157)
France 0.041* 0.170*** -18.403*** -6.231 -20.229"* 0.079 2.412" -0.056 0.453 -0.847 0.260
1.508 3.180 -3.016 -1.066 -2.823 0.858 1.776 -0.041 0.405 -0.789 (0.184)
Germany 0.021 0.119*'* -7.134* -4.343 -7.983* 0.073 1.436* 0.760 -0.563 -0.133 0.163
0.910 2.550 -1.316 -1.115 -1.326 0.872 1.313 0.023 -0.629 -0.122 (0.077)
Hong Kong 0.026 0.235'* -19.659** -6.524 -18.120* 0.229 8.269*** -2.167 -0.161 -0.482 0.151
0.547 1.920 1.220 -0.569 -1.368 0.958 3.223 -0.631 -0.049 -0.158 (0.064)
Intercept iDYWL iTSG7 iTED iSRG7 iWRL IPCO,I) IP(l,2) IPC2,3) IPC3,4) R2
Stock markets
Italy 0.045 0.111** ·28.004*** -8.563 -21.056*** 0.171* 4.027** -0.715 0.813 -1.699 0.183
1.155 1.902 -8.629 -1.105 -2.408 1.341 1.867 -0.351 0.441 -1.172 (0.100)
Japan 0.004 0.144*** -21.381*** 1.422 -19.186*** 0.094 5.276*** -1.731 1.310 -0.157 0.249
0.139 3.032 -3.597 0.237 -2.634 0.625 3.801 -1.032 0.996 -0.147 (0.172)
Netherlands 0.029* 0.136*** -10.862*** -6.143* -7.889* 0.084 3.294*** 0.341 0.270 0.191 0.334
1.599 3.289 -2.607 -1.638 -1.577 1.005 3.081 0.309 0.283 0.212 (0.267)
Norway 0.048* 0.068 -23.700*** -8.713* -18.414** 0.085 4.231 *** 0.634 -0.153 -1.661 0.187
1.316 0.856 -2.713 -1.604 -1.718 0.457 1.966 0.285 -0.079 -0.661 (0.103)
Spain -0.012 0.012 -3.908 -2.034 1.893 0.040 3.167** 0.707 -1.013 0.791 0.045
-0.286 0.215 -0.476 -0.313 0.177 0.244 1.724 0.289 -0.456 0.552 (-0.052)
Singapore 0.015 0.111* -29.331*** -5.617 -15.785* 0.023 4.163** 2.945* -0.441 0.899 0.214
0.409 1.322 -3.911 -0.621 -1.588 0.153 1.996 1.518 -0.261 0.547 (0.133)
Sweden 0.010 0.089* -15.566*** -2.244 -2.677 0.028 4.967*** 0.354 0.433 0.655 0.142
0.312 1.575 -2.189 -0.356 -0.341 0.202 3.112 0.239 0.317 0.512 (0.054)
Switzerland 0.044*** 0.135*** -10.439*** -10.600*** -9.428** 0.011 2.282*** 1.325** -1.305* 0.140 0.348
2.489 3.405 -2.614 -3.127 -1.886 0.192 2.505 1.615 -1.620 0.168 (0.282)
United Kingdom 0.002 0.128** -10.133* -4.204 -5.903 0.114 2.971 0.606 1.223 0.931 0.214
0.083 1.901 -1.402 -0.676 -0.730 0.935 1.306 0.315 0.816 0.744 (0.133)
United States -0.009 0.141*** -4.720 -1.869 -1.891 0.175* 3.108*** 0.464 0.907 -0.227 0.214
-0.339 2.707 -0.883 -0.336 -0.295 1.501 2.356 0.284 0.678 -0.177 (0.136)
The instruments and the measures of future real activity are defined in the text. The t-statistics reported under the coefficients (in italics) are adjusted for het-
eroscedasticity. *1**1*** denotes a coefficient estimated on the 20%110%15% level of significance. R-square denotes the coefficient of determination, ad-
justed R-squares are reported underneath in brackets.
(1)
Ii
.,
~
Time varying expected returns and the business cycle 145
To sum up, using quarterly data to investigate the relationship between global
forecasting variables and excess returns on international stock markets, my re-
sults are qualitatively similar to those in studies based on monthly data. 64 How-
ever, the effects are slightly more pronounced. An average R-square of 0.224
raises the question whether explaining almost 23 percent of quarterly return
variance is good or bad news for rational asset pricing stories. Of course, die-
hard critics will still argue that predictable variation in expected returns is irra-
tional. However, the more convincing story from an economic point of view
follows along the notion that a considerable fraction of time variation in ex-
pected returns is closely related to the cyclical movements/forecasts of real ac-
tivity. The observation that the global instruments drive out future growth rates
of global industrial production in joint regression can be interpreted as evidence
for this kind of explanation. Fama (1990) goes one step further arguing "[ ... Jit is
possible that, if the variables and functional forms that drive the rational varia-
tion in stock prices were somehow revealed, we would find that the in-sample
R-squares obtained here understate the rational proportion of the variation in
returns.,,65 Therefore, the next section attempts to isolate and extract the pure
business cycle components of stock prices and global instruments. The pre-
sumed relationships can then be examined by frequency bands.
cally, the correlations between stock prices and instruments are examined for
each frequency band separately. From an economic standpoint, to preserve the
concept of market efficiency, the bulk of the average 22.3 percent prediction
power should be attributed to the time variation in expected returns associated
with the business cycle frequencies.
The time evolution of the global instruments on the different frequency bands
has been shown in figures 4.1a-4.1d. Unfortunately, the setup must be slightly
changed. First, computing returns from levels is to apply a filter on the original
price series already, putting too much weight on the noise components and hid-
ing the cyclical movements. To avoid this problem, I use the original stock price
series and directly apply the Baxter/King (1998) framework to extract three
components: trend (fluctuations which exceed thirty-two quarters in duration),
business cycle (fluctuations of six to thirty-two quarters in duration), and noise
components (movements with frequency two to five quarters). Intuitively, this
changes the stories put forth in section 4.2 by reversing the signs of the relation-
ship between stock prices and global instruments (at least it should on the busi-
ness cycle frequencies). Expected returns can only be high when prices are rela-
tively low, and vice versa. Second, a regression analysis according to (4.3) is not
appropriate. By construction, the band-pass filter is a two-sided moving average
that maintains certain business cycle properties. Therefore, using business cycle
frequencies and running country-by-country regressions cannot be interpreted as
rational expectation forecasts of excess stock returns. Computing contempora-
neous correlations between filtered price and instrument series seems to be the
simplest alternative.
Results are shown in table 4.8. The time period for calculating correlations is
constrained from 1976.01 to 1995.03, as twelve quarters are lost at the begin-
ning and the end of a series when applying the band-pass filter. Overall, the re-
sults are encouraging for my working hypothesis. Theory implies a positive re-
lationship between expected excess returns and the dividend/price ratio yield.
Recall, it is an indicator of the current health of the economy. High (low) levels
of the ratio's business cycle frequencies should imply high (low) expected ex-
cess returns and, hence low (high) stock prices. Indeed, there is considerable
negative correlation between the business cycle components of stock prices and
Time varying expected returns and the business cycle 147
the world dividend/price ratio. Specifically, the average correlation is -0.31. Ta-
ble 4.8 also reveals a strong influence of the noise component. In contrast, the
G7 term spread is assumed to forecast the future health of the economy. The ar-
gument follows along the life-cycle hypothesis of ModiglianilBrumberg (1954)
and the permanent income hypothesis of Friedman (1960). The correlations
between stock prices and the term spread on the business cycle frequencies are
high in many of the countries. Equally important, the noise component of the G7
term spread seems to account for only a small fraction of the explanatory power
in predictive regressions. Recall, the sensitivity coefficient of the G7 term
spread was expected to be positive, but it was estimated with the reverse sign in
multivariate regression models. In table 4.8, however, the negative sign of the
correlation between stock prices and instrument variables on the cyclical fre-
quencies perfectly corresponds to the consumption smoothing story. The aver-
age correlation is -0.19, but it is considerable higher for several for the larger
stock markets; -0.38 for Canada, -0.35 for the United States, -0.32 for the Neth-
erlands, and -0.27 for the United Kingdom, among other countries with correla-
tions below -0.20. In particular, there is a tendency for high correlations to occur
in countries with high R-squares in linear regression models. The correlation on
the irregular band is above 0.1 only for Spain, and the average 'noise' correla-
tion is a mere 0.03. The TED spread had a significantly negative impact on ex-
cess returns in most stock markets, even though economic theory argues for a
positive relationship. Again, the multivariate analysis seems to hide important
economic effects. The results in table 4.8. show that the irregular components of
the TED spread do not explain a substantial fraction of common movements. On
the other hand, the correlation between cyclical movements of stock prices and
the TED spread is of the right sign and considerably high for Denmark (-0.33),
Germany (-0.22), and the United States (-0.34). For all other markets, there is no
clear picture as to whether the correlation is economically important, in a posi-
tive or negative direction. Overall, the yield differential between the 90-days
Eurocurrency interest rate and the Treasury bill rate seems to be a bad proxy for
a truly global state variable. The puzzle remains why the U.S. sensitivity on the
TED spread in table 4.5 was not estimated with sufficient statistical precision.
Again, the true relationship seems to be superimposed in a multivariate regres-
sion analysis, and it is only revealed when looking at the three frequency bands
148 Chapter 4
separately. The results for the G7 short-term interest rate are somewhat disap-
pointing. The correlations reported in table 4.8 are very low on both the noise
and the cyclical frequencies. This is surprising, given (i) that the relation has
been estimated highly significant in table 4.5, and (ii) the common belief across
financial practitioners that high (low) interest rates are bad (good) news for the
stock market. The latter argument seems to be valid only in long run, as indi-
cated by highly negative correlations on the trend frequencies.
Overall, the results are a first and encouraging step to identify those movements
in economic time series which drive the predictable variation in stock returns, as
suggested in Fama (1990). In fact, business cycle movements of the world divi-
dend yield, the G7 term spread, and - to a lesser extent - the TED spread seem
to explain a considerable fraction of the time variation in risk premiums on
global stock markets. Unfortunately, the approach is unable to capture the exact
percentage of regression R-squares explained by either trend, cyclical, or ir-
regular frequencies.
>-i
s·
(b
Table 4.8 <!
Correlations between stock prices and global instrument variables on different frequency bands ~
S·
Period: 1973.01 - 1998.03 (quarterly data) (JQ
(b
'0
><
World dividend yield G7 term spread TED spread G7 short-term interest rate (b
0
~
Stock market Trend Cycle Irregular Trend Cycle Irregular Trend Cycle Irregular Trend Cycle Irregular 0-
d
Australia -0.92 -0.47 -0.28 -0.64 -0.26 0.10 -0.89 0.18 0.06 -0.68 0.18 -0.05 2
3
Austria -0.92 -0.03 -0.09 -0.74 -0.05 0.09 -0.90 -0.13 -0.02 -0.73 0.21 -0.02 '"I»::s
0-
Belgium -0.93 -0.43 -0.38 -0.74 -0.22 0.01 -0.90 -0.02 0.07 -0.76 -0.02 -0.09 go
(b
c:r
c:
Canada -0.90 -0.38 -0.37 -0.43 -0.38 0.00 -0.87 -0.10 0.10 -0.55 0.17 -0.09
'"
S·
Denmark -0.91 -0.09 -0.08 -0.67 -0.21 0.04 -0.91 -0.33 0.04 -0.70 0.07 -0.04 I~
'"0
France -0.93 -0.34 -0.43 -0.71 -0.24 0.02 -0.91 -0.02 0.03 -0.73 0.06 -0.04
I~
Germany -0.95 -0.27 -0.22 -0.69 -0.25 0.05 -0.93 -0.22 0.03 -0.74 -0.03 0.01
Hong Kong -0.77 -0.26 -0.18 -0.74 -0.16 0.00 -0.79 0.13 0.10 -0.71 -0.13 -0.03
Italy -0.97 -0.38 -0.35 -0.51 -0.12 om -0.89 0.06 0.12 -0.64 0.12 -0.01
Japan -0.97 -0.48 -0.51 -0.53 -0.04 -0.01 -0.88 0.11 0.11 -0.63 -0.04 -0.04
Netherlands -0.88 -0.40 -0.38 -0.72 -0.32 0.03 -0.88 -0.16 0.09 -0.74 -0.12 -0.05
Norway -0.96 -0.13 -0.29 -0.63 -0.01 0.06 -0.92 0.08 0.20 -0.54 0.41 -0.08
World dividend yield G7 term spread TED spread G7 short-term interest rate
Stock market Trend Cycle Irregular Trend Cycle Irregular Trend Cycle Irregular Trend Cycle Irregular
Singapore -0.75 -0.25 -0.24 -0.52 -0.10 -0.05 -0.77 0.06 0.15 -0.69 0.07 -0.10
Spain -0.97 -0.31 -0.34 -0.76 -0.04 0.15 -0.92 0.10 0.10 -0.81 0.15 0.02
Sweden -0.93 -0.31 -0.40 -0.68 -0.15 0.05 -0.91 -0.20 0.13 -0.72 0.05 -0.07
Switzerland -0.86 -0.44 -0.41 -0.73 -0.27 -0.05 -0.86 -0.07 0.09 -0.75 -0.14 -0.03
United Kingdom -0.92 -0.33 -0.41 -0.66 -0.27 0.01 -0.90 -0.05 0.09 -0.70 -0.07 -0.08
United States -0.89 -0.26 -0.39 -0.66 -0.35 0.08 -0.90 -0.34 0.07 -0.72 -0.12 -0.02
Average -0.9/ -0.3/ -0.32 -0.65 -0./9 0.03 -0.89 -0.05 0.09 -0.70 0.05 -0.04
The table reports the contemporaneous correlations between quarterly stock prices (all denominated in Swiss francs) and global instrument variables on dif-
ferent frequency bands. Frequency bands are identified applying the band-pass filter proposed by Baxter/King (1998). 'Trend' components are fluctuations
in a time series which exceed 32 quarters in duration, 'business cycle' fluctuations are cycles of six to thirty-two quarters in length, and 'irregular' move-
ments are those with frequency two to five quarters. The band-pass filter passes through components of stock prices and instrument variables with cycles 01
the specified length, using twelve leads and lags of the data to construct moving averages and retaining the major features of business cycles.
(')
:r
.§
g
.j>.
Time varying expected returns and the business cycle 151
This chapter has documented time series evidence for the relationship between
global economic conditions and expected excess returns on international stock
markets. In intertemporal asset pricing models a la Merton (1973) and Campbell
(1996) the priced state variables are those that can forecast changes in the in-
vestment and consumption opportunity sets. In the theoretical part of the chapter
I have argued that simple linear prediction models are consistent with (more
complex) rational asset pricing stories. The working hypothesis that underlies
the empirical work consisted of two parts: (i) state variable that are positively
(negatively) related to the recent growth of the economy are negatively (posi-
tively) related to expected excess market returns, and (ii) state variables that are
positively (negatively) related to the future growth of the economy are positively
(negatively) related to expected excess market returns. To isolate and extract the
components of global instrument variables that ultimately drive the rational
variation in excess stock returns, a band-pass filter technique is applied to ex-
amine the interrelation between stock returns and instruments by frequency
bands, i.e. for trend, business cycle, and irregular components. The major results
of this chapter can be summarized as follows:
spread and the TED spread the presumed direction of the functional relation
with future stock returns is superimposed in multivariate regressions, i.e. the
signs of the estimated regression coefficients are opposite to what economic
intuition suggests. Standard Wald-tests show that the sensitivities on all five
global instrument variables possess explanatory power for the cross-sectional
differences in expected excess returns.
• The correlations between stock prices and the business cycle frequencies of
global instruments are considerably high for the world dividend yield and the
G7 term spread. The average correlation between stock prices and the world
dividend/price ratio on the business cycle band is -0.31, but a similar average
correlation coefficient of -0.32 on the irregular frequencies reveals a strong
Time varying expected returns and the business cycle 153
• The results are less clear-cut for the other two instruments. The TED spread
seems to be a merely U.S.-related instrument variable rather than a global
one. Significant business cycle correlations are only found for the United
States, Denmark, and Germany. Even more disturbing, the signs of the cor-
relations vary across markets. Similarly, no clear patterns emerge for the G7
short-term rate of interest. Of course, in the long run there is a clear negative
relationship between stock prices and interest rates, but the business cycle
correlations vary dramatically in both sign and magnitude across countries.
Surprisingly, the noise component of the short-term interest rate is almost
negligible for the time variation in stock returns.
155
Chapter 5
The last chapter has provided evidence that stock returns are predictable (at least
to some extent), and that the time variation of expected excess returns is linked
to the business cycle. In this chapter I will further explore the relationship be-
tween stock returns and real activity within a formal beta pricing framework.
Merton's (1973) pricing equation includes the covariances of each asset's return
with both the market portfolio and all state variables. His model specifies a mul-
tivariate linear proxy for marginal utility. In equilibrium, investors hold a port-
folio that is optimally hedged against changes in the investment opportunity set.
Breeden's (1979) consumption capital asset pricing model (CCAPM) collapses
the state variables into a single source of risk, the consumption beta. His pricing
equation includes the covariance of the asset's return with aggregate consump-
tion only. Stulz (1981) developed an international version of this model, where
expected returns are linear in a measure of real world consumption risk.
Recall, an asset whose payoff covaries positively with the discount factor has its
price raised, and vice versa. Hence, in the consumption-based model an asset's
price is lowered if its expected payoff covaries positively with expected con-
sumption, and vice versa. The driving force behind the consumption-based ap-
proach to asset pricing is the life cycle-permanent income theory, which posits
that investors do not like uncertainty about consumption. Buying an asset whose
payoff covaries positively with consumption makes an investor's consumption
stream more volatile. This is because such an asset delivers high payoffs when
the utility of an extra unit of consumption is low, but pays off badly when
wealth is most valuable to the investor. In short, such an asset is perceived as
risky and the investor requires a low price or, equivalently, a sufficiently high
156 Chapter 5
expected return to hold it. On the other hand, buying an asset whose payoff co-
varies negatively with consumption helps to smooth consumption. Hence, it is
more valuable for an investor, implying a higher price or, equivalently, a lower
expected rate of return.
Asset pricing theories link average returns to average risk. At the same time,
conditional asset pricing models must also capture the time variation in expected
returns. Market efficiency and the random walk hypothesis of stock returns have
long been treated as one and the same. In the context of intertemporal models,
however, predictability is not inconsistent with the concept of efficient markets.
The papers by Balvers/CosimanofMcDonald (1990), CheccetilLamfMark (1990,
1994), and Black (1990) show that market efficiency does not imply that stock
returns must follow a random walk. Their argument is that stock returns can be
predicted to the extent that there is predictability in the endowment process.
Extending the classical Lucas (1978) framework, they build models that demon-
strate how the interplay between consumption smoothing and risk correction can
generate mean-reverting behavior of stock returns. I Transitory changes in ag-
gregate output/income are related to aggregate savings. Cochrane (1988), among
others, documents some degree of trend reversion in aggregate output, and the
Balvers/CosimanofMcDonald model posits predictability in aggregate savings
on the basis of predictable changes in aggregate output. To maximize expected
utility, rational investors attempt to smooth consumption, thereby adjusting the
required rate of return on financial assets. Consequently, stock returns become
predictable as well. Any attempt to exploit this predictability increases variation
in consumption decreases expected utility. I will explore a model of this type in
chapter 7. In this chapter it is argued that rational beta pricing models must si-
multaneously explain the differences in international expected returns (i.e. the
cross-sectional restriction) and the predictability of excess returns on the basis of
instrument variables related to the business cycle (i.e. the time series restriction).
Indeed, Ferson/Harvey (1993) and Ferson/Korajczyk (1995) show that a large
fraction of the predictable variation in security returns for different investment
horizons is captured by conditional beta pricing models.
Evidence for mean reversion has first been provided for u.s. data by PoterbalSummers
(1988) and FamaIFrench (1988a).
Testing a conditional version of the consumption-based model 157
This chapter is organized as follows. There are two main working hypothesis: (i)
time varying risk premiums on a small number of common risk factors can ex-
plain a considerable fraction of excess return predictability on global stock mar-
kets, and (ii) some conditional moment implications of the consumption-based
asset pricing model can explain the global cross-section of expected stock re-
turns. Section 5.1 introduces different versions of so called latent variable mod-
els. This type of model posits useful time series and cross-sectional restrictions
on expected stock returns. The latent variable setup can be extended to explore
whether the growth rates of different measures of global consumption qualify as
proxies for the (possibly many) latent state variables. Since latent variable mod-
els also allow assessing whether stock markets are globally integrated or seg-
mented, the theoretical concept of capital market integration is presented in sec-
tion 5.2. Section 5.3 presents the empirical results for all versions of the model
introduced in this chapter.
Many economic models yield predictions about the correlation between unob-
served expectations of two series. For example, Obstfeld (1989) works with a
model in which fully integrated world capital markets imply that expected con-
sumption growth is perfectly correlated across countries. Cumby/Huizinga
(1992) argue that the hypothesis that real exchange rate follows a random walk
implies that expected changes in nominal exchange rates and expected inflation
differentials are equal. Taking another look at the permanent income hypothesis,
CampbelllMankiew (1989) argue that a perfect correlation between expected
changes in disposable income and expected changes in consumption can be
taken as evidence of 'rule of thumb' consumers. Most important, in their semi-
nal paper Hansen/Singleton (1983) show how the consumption capital asset
pricing model (CCAPM) can yield the prediction that expected consumption
growth and expected real asset returns are perfectly correlated. Using a sample
of international stock returns, this argument can be reversed to design a con-
sumption-based test of market integration?
Latent variable models allow for time variation in risk premiums, but maintain
the assumption that the conditional betas are fixed parameters. In an interna-
tional context, the time variation of expected stock returns is assumed to be a
function of time varying risk premiums on a small number of global factors,
with constant conditional beta factors as the coefficients of proportionality.
Therefore, the major restriction of the model is that expected returns are per-
fectly correlated across countries. Latent variable models remain agnostic about
the underlying risk factors, they are treated as being unobservable. A simple ver-
sion of a latent variable model was originally introduced and tested using for-
eign exchange market data by HansenIHodrick (1983). GibbonslFerson (1985),
Cumby (1990), CampbelllHamao (1992), and Harvey/SolniklZhou (1995) apply
more sophisticated models to global stock market data. Campbell/Clarida (1987)
and Jorion (1992) estimate latent variable models with Eurodeposit returns, and
BekaertlHodrick (1992) include both stock and foreign exchange market data in
their analysis.
version of the capital asset pricing model (CAPM) asked whether cross-sectional
differences in average risk could explain differences in average returns. No time
series predictions other than constant expected returns are implied. s This uncon-
ditional setup, however, fails to simultaneously explain the cross-sectional dif-
ferences in expected stock returns and their predictable time variation. Uncondi-
tional models do not account for new information that periodically becomes
available to investors, who then adjust their investment strategies accordingly. In
contrast, conditional asset pricing models assume that investors form time
varying expectations about the joint probability distribution of future returns.
The cross-sectional implications remain unchanged, differences in conditional
expected returns are determined by differences in conditional risk. In addition,
however, conditional pricing models also contain meaningful implications for
the predictable time variation in expected returns.
The essence of conditional beta pricing models is that conditional expected re-
turns can vary with changes in both conditional risk and fluctuations in market-
wide risk premiums. A number of different methods have been suggested to in-
corporate information available at time t to forecast the return over the period
from t to t+1. Most frequently, instrument variables are used. Recall, instrument
variables proxy for the total set of (unobservable) information used by investors
to form return expectations. Ideally, these variables should be associated with
the business cycle. Following the procedure in chapter 4, I model expectations
as linear combinations of a small number of instruments. In particular, the re-
turns on asset i, R,.l+l' are projected on the vector of predetermined variables,
again collected in the vector Zt. The fitted values 0iZt are taken as unrestricted
estimates for the conditional expected returns on asset i, written as E(R'.'+lIZ,).
0, represents the vector of this asset's fixed sensitivities against the predeter-
mined set of instruments. The intuition behind latent variable models is to im-
pose cross-sectional restrictions on these sensitivity coefficients, allowing to
detect reduced dimensionality in the predictable time variation of returns across
global stock markets.
5 In an international setting, such models have been tested by Solnik (I 974c), Stehle
(1977), CholEunlSenbet (1986), Wheatley (1988), KorajczyklViallet (1990), and Gulte-
kinlGultekinlPenati (1989), among others.
160 Chapter 5
Consider the conditional (linear) beta pricing model of the familiar form:
(5.1)
which posits the existence of K market-wide conditional risk premiums (or ex-
cess returns on factor mimicking portfolios), denoted as A. j (Zt), for j = O, ... ,K.6
If there is a portfolio with betas equal to zero on all K factors, then A. o (Zt) is
the expected return on this 'zero-beta' portfolio (see Black (1972». The coeffi-
cients b 'l , ••• , b ,K denote the conditional betas of asset i relative to the K under-
lying risk factors. Conditional on the set of information, E(Rj,t+dZt) describes
the expected return on asset i. In general, the betas will also be time varying
functions of Zt. The primitive latent variable models restrict beta pricing mod-
els by assuming that the conditional betas are fixed parameters. To mitigate this
limitation, I will show how to extend latent variable models to incorporate time
varying risk exposures. Most important, however, risk premiums vary over time
and in correspondence to the business cycle, depending on the time evolution of
the instrument variables. The best known example of such models is the capital
asset pricing model (CAPM), where K = 1 and the market portfolio as the state
variable.
(5.2)
(5.3)
where rt+l denotes the N dimensional vector of excess returns, and fJ (with ele-
ments b l! - bo ) is a K x N matrix of betas for excess returns. Ferson (1995)
notes that tests that use excess returns do not exploit all restrictions of a beta
pricing model. In particular, it is not possible to determine whether the model in
(5.3) has to be interpreted as a 'zero beta' model or rather as a model which as-
sumes that a riskless asset exists. In fact, latent variable models that are defined
in terms of raw returns impose additional restrictions on the vector of sensitivity
coefficients.9 To derive a latent variable model, take the vector of excess returns
and partition it as:
(5.5)
where the dimensions of PI and P2 are KxK and [Kx(N - K)], respectively.
8 As has been shown in chapter 3, international asset pricing models are always stated in
terms of real returns. Working with excess returns is convenient because it does not re-
quire specifying an appropriate price deflator. Empirically, real returns and excess
nominal returns are only hard to distinguish.
9 See FersonIFoersterlKeim (1993).
10 This terminology has been originally introduced by Ferson (1990).
162 Chapter 5
With this notation, the multibeta model in (5.1) can be split and written as:
E(rl.I+IIZt) = At+lfll
(5.6)
E(r2,t+II Z t) = At+1fl2,
where At+1 is the 1x K vector of time varying risk premiums, common across all
stock markets. The difficulty in empirical work is the choice of economically
meaningful factors. The true underlying factors are not observable for the finan-
cial econometrician. However, combining the two parts in equation (5.6) and
eliminating the benchmark portfolio returns is a way to avoid the problem. Ma-
nipulating the first part of equation (5.6) yields:
(5.7)
Substituting this expression into the second part of (5.6) results in a cross-
sectional restriction on the expected returns of the N - K test assets:
(5.8)
In other words, the conditional expected returns on N - K test assets are pro-
portional to the conditional expected returns on K reference assets. Constant
conditional betas serve as the coefficients of proportionality. The key ingredient
of latent variable models is to condition the expected reward for exposure to a
unit of (unspecified) global risk on a set of global instrument variables, which
are observable at the beginning of each period, i.e. when asset prices are set at
time t. To keep the model as tractable as possible from an econometric point of
view, the relationship between global instruments and the risk premium on any
latent factor is assumed to be linear:
(5.9)
where Zm,t (with m = 1,2, ... ,L) denotes the level of global instrument variable
m at time t. The slope coefficients Ojrn (with m = 1, 2, ... ,L) capture the influ-
ence of these variables on the j-th risk premium (j =1,2, ... ,K). 0)0 is a constant
term and denotes the part of the j-th risk premium that does not vary over time.
As in chapter 4, global instrument variables with a clear association to the busi-
Testing a conditional version of the consumption-based model 163
ness cycle are used in my empirical investigation: the world dividend yield, a G7
term spread, the TED spread, and a G7 short-term interest rate. Hence, the di-
mension of Zt is lx5 (including the constant). The results in chapter 4 pro-
vided evidence that these variables do have an ex-ante ability to proxy for the
current and/or future states of the world.
Harvey (1991a) suggests a straightforward setup that allows to test the pricing
restriction in equation (5.8). A possible specification of a latent variable model
is as follows:
(5.10)
(5.11)
15 Harvey/SolniklZhou (1995) and Zhou (1994) construct analytic tests of latent variable
models, i.e. tests with closed form solutions for parameter estimates.
Testing a conditional version of the consumption-based model 165
16 This would only be true if returns and factors were i.i.d. See Cochrane (l999a), p. 104.
17 See sections 2.4.5 and 2.9.5.
166 Chapter 5
formulation of Roll's (1976) critique. He emphasizes that the capital asset pric-
ing model (CAPM) cannot be rejected if the true market portfolio was used.
From a theoretical point of view, therefore, the bottom line is that multiple fac-
tor models make only sense if one imposes some additional, economically
meaningful restrictions on the factors.
Wheatley (1989) criticizes latent variable models, arguing that the constant beta
part of the hypothesis cannot be examined when the state variables are not speci-
fied. On the other hand, Connor/Korajczyk (1989) show that this simplification
is not excessively restrictive. Given some additional assumptions, their central
result is that a unique sequence of competitive equilibrium returns according to
the conditional multibeta pricing relationship in (5.1) must exist, where the
vector of beta coefficients is constant through time. 18 In addition, Ferson/Harvey
(1993), Evans (1994), and FersonlKorajczyk (1995) report that a large fraction
of predictable time variation in stock returns is captured by time variation in risk
premiums, while the direct effect of time varying betas accounts for only a small
fraction of predictability. 19 Nevertheless, apart from the perspective of modeling
I
the time series variation in expected returns, time varying betas may be impor-
tant for modeling the cross-sectional variation in expected returns. The results
by JagannathanlWang (1996) indicate that the common time variation in both
betas and risk premiums influences the cross-sectional implications of the con-
ditional capital asset pricing model (CAPM) for its unconditional version. 20
Latent variable models are not precise tests of a particular equilibrium theory of
international asset pricing. Three interpretations should be considered instead.
First, tests of latent variable models can be motivated by existing asset pricing
theories with additional restrictions. As has been demonstrated above, the sys-
tem in (5.11) can easily be justified on the basis of conditional beta pricing
models with time varying risk premiums, but constant conditional betas. Second,
using additional assumptions, Campbell (1987) shows how latent variable mod-
els can be related to Merton's (1973) intertemporal version of the capital asset
pricing model (ICAPM).21 Third, Campbell (1987) and CampbelUClarida (1987)
argue that latent variable models could simply be viewed as data description, i.e.
as parsimonious characterizations for time varying predictions of excess returns,
ignoring all pricing implications. Conditional on the set of instrument variables,
expected returns are restricted to move proportionately in response to move-
ments in common risk factors, with constant conditional betas as the proportion-
ality coefficients. Expected stock returns on different markets are assumed to be
perfectly correlated. More structured implications from an asset pricing point of
view can easily be incorporated into the model, as will be shown in section 5.1.4
below.
(5.12)
where' ®' denotes a Kronecker product of vectors. Given N asset returns, L in-
struments and K state variables, the model implies N x L orthogonality condi-
tions with [K x (N - K) + (Lx K)] parameters to estimate. Hence, the qui-square
test statistic for the model's goodness-of-fit contains (L - K) x (N - K) overi-
dentified restrictions.
(5.13)
where 1 is a K x 1 vector of ones. However, note that with the original set of L
instrument variables the dimension of C* becomes [Lx(N - K)], implying that
the system is no longer identified. Therefore, a specific subset of the original set
of instrument variables with dimension L* < L, denoted by Z;, must be chosen.
Given the discussion in the previous section, it is clear that a single latent vari-
able model with time varying betas must not be rejected on theoretical grounds.
Empirically, however, given that the fraction of return variance that is explained
by lagged instrument variables is very small, this would not be surprising. Latent
variable models rely fundamentally on this predictable variation. In addition,
Ghysels (1998) finds that time varying betas are inherently misspecified. He
demonstrates that in many cases the pricing errors in models with constant betas
are smaller than in specifications with time varying betas. Ghysels even con-
cludes that the misspecification of time varying conditional betas appears to be
serious enough that they do not help but actually hurt, especially in linear mod-
els of the stochastic discount factor. 22
22 See Ghysels (1998), p. 550. However, Ghysels also demonstrates that time varying con-
ditional betas in combination with non-linear models of the stochastic discount factor
lead to significantly reduced pricing errors (e.g., BansallHsiehlViswanathan (1993».
23 See RollIRoss (1980) and Connor/Korajczyk (1988), and KorajczyklViallet (1990).
Testing a conditional version of the consumption-based model 169
agent's welfare does not only depend on his own wealth, but also on the state of
the economy. His pricing equation includes the covariance of each asset's return
with both the market portfolio and the full set of state variables. Breeden (1979)
extends this analysis showing that preferences must be defined over consump-
tion and not over wealth when investment opportunities are time varying. Con-
sumption is the best measure of bad states, implying that in the consumption
capital asset pricing model (CCAPM) an aggregate consumption rate is the sole
state variable. In other words, Merton's (1973) intertemporal multibeta model
collapses into a single-beta model, where the betas are computed with respect to
real aggregate consumption rates. Stulz (1981) extends the model to an interna-
tional context. 24
This section demonstrates that the latent variable framework can be extended to
capture a more structured asset pricing hypothesis. In particular, the primitive
latent variable framework is restated to test a conditional consumption-based
linear factor mode1. 25 In their seminal analysis, Hansen/Singleton (1983) were
the first to test conditional moment restrictions of the model, but they work di-
rectly in a stochastic discount factor framework. Cumby (1990) interprets the
model in (5.11) as a test of whether international stock returns are consistent
with the consumption-based model of international asset pricing, loosely inter-
preting the latent factor as associated with consumption risk. Ferson (1990) and
Ilmanen (1995) both develop models that allow to examine directly whether
consumption rates qualify as proxies for the latent state variables. Their ap-
proaches require actual consumption data as a model input. I apply their frame-
works and empirically test the proposition that the premiums on the latent fac-
tors in (5.l1) are a compensation for real global consumption risk. In a nutshell,
I explore whether or not the conditional beta pricing model with multiple risk
factors collapses into a model with a single consumption beta. 26
Ferson (1990) extends (5.11) by adding an additional equation to model the un-
anticipated part of an observable proxy for the true underlying risk factor, de-
noted as fr.+!. The econometric framework is as follows: 27
The first equation in (5.14) defines the unanticipated part T11+1 of the single state
variable fr.+I, defined as the residual of a linear projection on the vector of global
instrument variables Zt (including a constant), with fixed coefficient vector y.
The second equation defines a conditional beta ~ as the regression coefficient of
unexpected returns on the factor innovation T1t+I' The vector £t+1 denotes the
idiosyncratic part of unexpected returns. Both equations, again, have little em-
pirical content unless one specifies the exact parameterization of expected re-
turns. Therefore, the third equation defines a linear regression model for predic-
tions about the state variable, conditional on the L dimensional vector of instru-
ments Zt. Combining the restrictions in (5.8) with the system in (5.14) produces
an empirically testable, conditional version of a consumption-based asset pricing
model. In particular, using the restriction for the time variation in expected re-
turns derived above,
(4.15)
and the statistical assumptions in (5.14) yields the following linear model with a
single risk factor fr.+I:
fr.+1 == Zt Y + T1t+1
rl,t+1 == ZtOI + (fr.+1 - Zt y) PI + £1,1+1 (5.16)
r2,t+1 == ZtOI (P1 IP2) + (fr.+1 - Zty)P2 + £2,t+I'
Again, the first equation in (5.16) specifies the conditional mean of the state
variable kl' In particular, growth rates of different components of G7 aggre-
gate consumption are used as the sole state variable. The innovations in the risk
factor, denoted by 111+1, represent the relevant type of risk. 28 Any predictable
behavior of the risk factor that is captured by the instrument variables will not
affect the model. Therefore, y is a vector of free parameters, and only the series
of disturbances 111+1 affects the conditional beta pricing model in the second and
the third line of (5.16). In particular, the regression model for the K reference
assets is contained in the second equation, defining the vector PI as the associ-
ated conditional consumption betas. The third equation represents the regression
model for the N - K test assets, with P2 as their conditional consumption betas.
Note again that only the unanticipated parts of stock market returns and the state
variable are reflected in the estimates of constant conditional consumption betas.
The cross-sectional restriction embedded in (5.16) is:
(5.17)
(5.18)
which are also estimated using the Generalized Method of Moments (GMM).
Given the N excess returns, the L instruments, and the K state variables, there
are [(N + K)xL] orthogonality conditions. Hence, the number of free parame-
ters to estimate in (0 1, y,PI,P2) is [Kx (2L+ N)]. Overidentification restricts the
number of assets, and it must hold that N > [(KxL)/(N - K)]. In the spirit of
Breeden's (1979) original analysis, only one state variable with K = 1 is used in
the empirical section.
(5.19)
In this section I argue that failure to reject latent variable models is consistent
with integrated global stock markets, i.e. markets in which the riskiness of an
asset is determined by the covariance of its returns with global sources of sys-
tematic risk. In less formal words, integration implies that there is a small num-
ber of common value drivers sufficient to capture the co-movements of expected
returns across global stock markets.
financial assets with the same risk characteristics have identical expected re-
turns, irrespective of the market to which they belong. 3D More fundamentally,
perfect risk sharing in complete global capital markets leads to perfectly corre-
lated consumption rates. 3l Assuming that consumption is the best measure of
bad states (and, hence the correct state variable), expected returns should also be
perfectly correlated across integrated markets. 32 This is perfectly consistent with
the prediction of latent variable models.
35 See BekaertlHarvey (1995), p. 405. For an overview of different methods used among
practitioners to measure the cost of capital see Harvey (1997).
36 For example, see Beckers/Connor/Curds (1996), BekaertlHarvey (1997), Hargis/Ma-
loney (1997), and BrackerlDockinglKoch (1999).
Testing a conditional version of the consumption-based model 175
a country decides to open up, investors can transfer some part of these risks to
foreign investors, while taking some foreign risk in exchange. Stulz (1995) ar-
gues that each country's cost of capital is reduced by making risks diversifiable
that would not be otherwise diversifiable. 38 Obstfeld (1994) uses the risk sharing
argument to show that financial diversification may itself promote resource allo-
cations favorable to long-term economic growth. Investors shift funds from low-
return, safe investments toward high-return, risky investments they would not
otherwise undertake. He concludes that growth depends on the availability of an
ever-increasing array of specialized, hence inherently risky, production inputs.
Stulz (1999) derives a necessary and sufficient condition for a reduction in the
risk premium is small capital markets?9 Denoting the volatility of stock returns
in small markets as <Js ' the volatility of world market returns as <J w , and the
correlation between the two as Ps,w' the condition is as follows:
(5.20)
The risk premium depends on the covariance of a country's stock return with the
return on the global market portfolio. A country is risky if the covariance is
high, implying that a large risk premium is required from a global perspective.
On the other hand, if the return on the market portfolio of the small country is
uncorrelated with the return on the world stock market, then the small country's
market portfolio is not expected to earn a risk premium provided that it is inte-
grated in the global market. Hence, the intuition behind (5.20) is that as long as
the correlation between the small country's market portfolio return and the re-
turn on the world market portfolio is not too high, or if the volatility of small
country's market portfolio is not too low, the small country's risk premium falls
when it opens up toward the global equity market. In the most relevant case,
where the volatility of world market returns is lower than that in the small coun-
try, globalization necessarily decreases the cost of capital for investments in a
small country, irrespective of the correlation coefficient.40
40 Stulz (1999) identifies a second source of pressure on the local cost of capital. Global-
ization increases the monitoring by investors, thereby reducing the discount rate they
charge.
41 This follows from simple static valuation models such as Gordon/Shapiro (1956) or the
dynamic version in CampbelUShiller (1988a,b) and Campbell (1991). See also the dis-
cussion in Bekaert/Harvey (2000).
42 See Wheatley (1988a).
178 Chapter 5
tion of stock market integration can have two reasons, which cannot be disen-
tangled: (i) either the markets examined are segmented and/or (ii) the underlying
pricing model is incorrect.
Another difficulty with the international capital asset pricing model (ICAPM) is
that the econometrician has to identify an appropriate world market portfolio. In
his seminal paper Roll (1977) shows that tests of the capital asset pricing model
(CAPM) really test the mean-variance efficiency of the market portfolio, and
that the model must hold tautologically if the true market portfolio is used. 43
Latent variable models provide an elegant way to circumvent this problem. The
choice of a specific benchmark portfolio is avoided by the very nature of the
model. The proposition is that expected returns vary through time in a perfectly
correlated fashion. All stocks are assumed to be driven by the changing prices
for a small number of unobservable (latent) sources of risk. Therefore, expected
returns have a latent variable representation when markets are integrated. Nev-
ertheless, the latent variable model is a very general representation of traditional
beta pricing models, and as such invariably suffers from the joint-hypothesis
problem. One may falsely reject the hypothesis of integrated stock markets if
they are in fact integrated, but the assumptions about the unobservable bench-
mark are inappropriate. For example, if Swiss and U.S. stocks are exposed to
different sources of risk, and if the prices of these risks move independently, ex-
pected excess returns move independently even if prices are set in a single world
capital market.
One possible classification of empirical tests for global stock market integration
is whether conditional or unconditional moment implications of returns are ex-
amined. A more appropriate way is to refer to the underlying degree of integra-
tion: segmented markets, integrated markets, and partially segmented markets.
The first tests of the capital asset pricing model (CAPM) all assumed that mar-
kets were completely segmented. The second class of studies assumes that world
capital markets are perfectly integrated. For example, Solnik (1974c), Stehle
(1977), lorion/Schwartz (1986), CholEun/Senbet (1986), and Gultekin/Gultekin!
43 See section 2.9.5, but also Stambaugh (1982), Shanken (1987), KandeUStambaugh
(1987,1995), and RolU Ross (1994).
Testing a conditional version of the consumption-based model 179
Penati (1989) test simple models of this type with constant expected returns.
Wheatley (1988a) tests an unconditional version of the consumption based asset
pricing model using international stock market data. The papers by Camp-
belVHamao (1992), FersonIHarvey (1993), Harvey/SolniklZhou (1995), Fersonl
Korajczyk (1995), and OertmannlZimmermann (1997) can also be interpreted as
tests for perfect stock market integration with time varying risk premiums. 44
This chapter also falls into this category. A final strand of literature falls in be-
tween segmentation and integration, the so called mild segmentation model. Er-
runzalLosq (1985) model restricted market access fer a specific class of inves-
tors, but leave the degree of segmentation fixed through time in their model.
Bekaert/Harvey (1995) specify a regime switching model, allowing the degree
of market integration to change over time.
There are three working hypotheses to be tested empirically. First, it has been
argued that a considerable fraction of the predictable time variation in excess
returns can be explained by the time variation of risk premiums on a small num-
ber of global latent risk factors. Second, it is assumed that stock markets are in-
tegrated, i.e. that the time variation in expected returns is (almost) perfectly cor-
related across countries. Third, different measure of world consumption can be
used to proxy for the latent state variables. In other words, a linear factor model
with a single conditional beta relative to consumption is assumed to hold. This
section reports the empirical results.
asset. Therefore, the beta of country i is c, times as large as the Swiss beta. The
left panel of table 5.1 shows the estimation results for a single latent variable
model with constant conditional betas, as described in (5.11). The estimates re-
veal intuitively reasonable patterns. All country betas are formed as positive
combinations of the Swiss market index. Most of the c, estimates are significant
at conventional levels of statistical precision, as shown by the t-ratios under-
neath the estimated factor exposure (in italics). The coefficients on the global
instrument variables, captured in the 0 vector, are also estimated with statistical
precision. As in the previous chapter, the following four instruments with their
close association to the business cycle are applied: the world dividend yield
(iDYWL), a G7 term spread (iTSG7), the TED spread (iTED), and a G7 short-
term interest rate (iSRG7). The lagged world market return (iWRL) is added to
possibly capture any remaining predictability. However, the associated sensitiv-
ity coefficient is not significant, indicating that the global instruments (plus a
constant) suffice to match the time series restriction of a single latent variable
model. This could be interpreted as evidence that the relative beta on a single
source of risk, in combination with a time varying risk premium, explains the
cross-sectional variation in expected excess returns.
Nevertheless, the GMM chi-square test statistic rejects a single latent variable
model, as indicated by the p-ratio of 0.017. This is in line with the results of Fer-
son (1990) for U.S. stock market data and Campbell/Hamao (1992) for a com-
mon U.S. and Japanese data set. Consequently, conditional beta pricing models
are consistent with either (i) a single time varying conditional beta, or (ii) more
than one constant conditional beta for each asset. I will examine these alterna-
tive specifications later on. For the moment, however, it is instructive to assess
the performance of the single latent variable model in more detail. In particular,
I look at the volatilities of the unrestricted and the restricted forecasts of excess
returns, as well as their correlations across markets to explore why the model
fails.
sions in chapter 4. The intuition of this test is similar to the variance ratio tests in
Ferson/Harvey (1993) and FersoniKorajczyk (1995). If the restricted volatility is
much smaller than the unrestricted one for a particular market, the cross-
sectional restrictions of the single latent variable model are causing serious dete-
riorations in prediction power. Table 5.2 reports both restricted and unrestricted
volatilities, together with the p-values of an F-statistic testing the null hypothesis
that both expected excess return series have equal variances. It is interesting to
note that there are three countries (Austria, Denmark, and Germany) where the
restricted volatility is higher than the unrestricted one. Given a rather low ex-
planatory power of simple instrumental variable regressions for these countries -
especially for Austria and Denmark, where it is virtually nil - this is not sur-
prising. With only a few exceptions, the null hypothesis of equal variances is
rejected by an F-statistic, which has 100 nominator degrees of freedom and 100
denominator degrees of freedom. 45 This indicates that the model's cross-
sectional restrictions adversely affect the prediction power of the global instru-
ments. Preserving this informational content of the instruments for the time se-
ries behavior of stock returns, however, is essential for the cross-sectional re-
strictions in the latent variable framework to be empirically meaningful at all.
45 The number of degrees of freedom is determined by the 101 observations in the sample.
182 Chapter 5
As a final check, table 5.3 reports the cross-country correlation coefficients be-
tween the series of unrestricted excess return predictions. The restricted series
are perfectly correlated by construction, with a single relative beta coefficient as
the factor of proportionality. Rejecting the model implies rejecting (almost) per-
fect correlation of excess return forecasts. What the p-values do not tell, how-
ever, is how far the predictions are away from perfect correlation and whether
the divergence of the estimates from the null hypothesis is economically impor-
tant. The correlation coefficients of unrestricted excess return predictions can
give some insight. In general, the correlations are high, usually well above 0.8,
and often 0.9. However, some smaller countries display only a modest degree of
common movement with the other countries. As expected, these are the stock
markets with little explanatory power in instrumental regressions, namely Aus-
tria, Denmark, Italy, Norway, and Spain. The correlations fall below 0.3 in some
cases, although the results in table 4.5 have shown that the signs of the sensitiv-
ity coefficients on the instrument variables are very similar across countries. 46
only a few exceptions - not significant. This confirms Ghysels' (1998) results,
reporting that beta risk is often misspecified in conditional asset pricing frame-
works. Somewhat pragmatic, no further attempt is made to capture the dynamics
of beta risk. Rather, I interpret the empirical results for the augmented specifica-
tion in (5.13) as evidence for the primitive latent variable specification in (5.11).
Although the restrictions of a single latent variable specification have been re-
jected, the model seems to hold reasonably well for many countries in the sam-
ple (especially for the larger ones). Hence, a natural approach is to test whether a
model with more than one constant relative beta for each asset fits the data more
appropriately.47
47 Using factor analysis. CholEunlSenbet (1986) and KorajczyklViallet (1990) find evi-
dence for a small number of unobservable factors in international stock returns.
48 See Ferson (1990), p. 409.
184 Chapter 5
Table 5.1
Asset pricing models with a single unspecified state variable
System of non-linear equations for GMM estimation:
[r",., - Z,o,], 1
10,., =(u",., u",., e, .. >' =[ [r",., - z,o,],. .'
[Z,o,-Z,o,(l®Z,)C]
Constant factor
Time varying factor exposure e~
exposure e'l
Relative factor Coefficients of linear combination of factor exposure
Table 5.1
Continued
Constant factor
Time varying factor exposure c~
exposure e'l
Table 5.2
Comparison of unrestricted and restricted volatilities
Period: 1973.01- 1998.03 (quarterly data)
Unrestricted Restricted
Country F-statictic
volatility p.a. volatility p.a.
32.875"'**
Australia 0.119 0.021
0.000
4.921***
Austria 0.039 0.088
0.000
2.761 ***
Belgium 0.102 0.061
0.000
5.844***
Canada 0.087 0.036
0.000
2.514***
Denmark 0.066 0.104
0.000
2.506***
France 0.117 0.074
0.000
2.107***
Germany 0.077 0.111
0.000
6.494***
Hong Kong 0.134 0.053
0.000
2.673***
Italy 0.107 0.066
0.000
1.134
Japan 0.102 0.096
0.529
1.024
Netherlands 0.097 0.096
0.907
6.744***
Norway 0.112 0.043
0.000
2.768***
Singapore 0.127 0.076
0.000
27.038***
Spain 0.039 0.007
0.000
1.461 *
Sweden 0.062 0.052
0.060
1.482*
Switzerland 0.104 0.086
0.051
1.188
United Kingdom 0.108 0.099
0.391
1.304
United States 0.091 0.079
0.187
The table shows annualized volatilities of unrestricted and restricted predictions of excess stock returns.
Restricted estimates are from the single latent variable model using GMM point estimates of the pa-
rameters, unrestricted estimates are the fitted values from the least square instrumental predictions in
chapter 4. The last column reports an F-statistic for the null hypothesis of equal variances; p-values are
reported underneath (in italics). The F-statistic has 100 numerator degrees of freedom and 100 denomi-
nator degrees of freedom. *1**1*** denotes statistical significance at the 10%15%11 % level.
Table 5.3 ;'
Correlation matrix of unrestricted excess returns predictions '"g.
OQ
Period: 1973.01 - 1998.03 (quarterly data)
n
'"0
AUS AUT BEL CAN DEN FRA GER HK ITA lAP NET NOR SGP SPA SWE SWI UK USA :;:I
e:
g.
AUS 1.00
:;:I
AUT 0.86 1.00 e;.
-<
Cl>
BEL 0.80 0.77 1.00 ...
CAN 0.99 0.81 0.74 1.00 '"o·
:;:I
Figure 5.1
Time evolution of expected excess returns
02 0.15
0.10
0.1
0.05
00 000
·005
·01
·0.10
·015
75 80 85 90 95 75 80 85 90 95
Australia Austria
015 015
010
005
0.05
0.00
·0.05
'{)IO
·015
·020 ·0.15
75 80 85 90 95 75 80 85 90 95
Belgium Canada
015 02
010
01
75 80 85 90 95 75 80 85 90 95
Denmark France
Figure 5.1
Continued
0.2 02
0.1
0.1
00
0.0
-0.1
-0.1
-02
-02 -03
75 80 85 90 95 75 80 85 90 95
0.2 020
0.15
0.1
0.10
0.05
00
0.00
-0 I -005
-010
-02 -015
75 80 85 90 95 75 80 85 90 95
Italy Japan
015 02
010
01
005
000 0.0
-005
-0 I
-010
-015 -02
75 80 85 90 95 75 80 85 90 95
Netherlands Norway
FigureS.1
Continued
________________,
02~----------
o04 ~--------------------------,
01 002
0.00
0.0
-0.02
-01
-0.04
-0.2 -006
75 80 85 90 95 75 80 85 90 95
Singapore Spain
015~---------- ______________-, 02~-------------------------,
0.10
0.1
0.05
0.0
0.00
-01
75 80 85 90 95 75 80 85 90 95
Sweden Switzerland
02~----------------------__-, 015~-------------------------,
010
01
005
00 000
-005
-0.1
-0.10
75 80 85 90 95 75 80 85 90 95
The figure shows the time evolution of expected excess returns. The solid line exhibits unrestricted
excess return predictions, computed as the fitted values from the instrumental least square regressions
in chapter 4. The dashed line displays the restricted return predictions from the single latent variable
model in (5.11), using the GMM estimates to produce the fitted values. By construction, this line is
identical for all countries, differences in scale occur because of differences in the relative risk expo-
sures. All returns are reported on a quarterly basis.
Testing a conditional version of the consumption-based model 191
Tables 5.4 and 5.5 summarize the empirical results for latent variable models
with K = 2 and K =3 factors. Simulation analysis by Ferson/Foerster (1994)
shows that the GMM methodology tends to reject a true null hypothesis too of-
ten, especially in large models. In order to keep the system tractable and pre-
serve the power of the tests, results are reported for two separate subsamples: G7
stock markets in table 5.4 and non-G7 stock markets in table 5.5. Switzerland,
the United States (and Germany when K:= 3) are used as reference countries in
the G7 set, and Switzerland, the Netherlands (and Hong Kong when K:= 3) in
the non-G7 sample. It can be shown that p-values and pricing errors are invari-
ant to this choice of reference assets. 49 Many coefficients in C are negative, but
never significant. The latent risk premiums significantly load on the global in-
struments most of the time, as shown by the estimates for the «') vector of sensi-
tivity coefficients. Most important, the p-values indicate that neither model can
be rejected this time. These results can be interpreted as follows. First, there is
evidence that time varying risk premiums on K:= 2 or K:= 3 factors account for
a large fraction of the time variation in expected excess returns (i.e. the time se-
ries restriction of the model). Second, the estimates support the working hy-
pothesis of global stock market integration. Integration implies that a low di-
mensional factor structure suffices to describe a high (in fact, near perfect) de-
gree of co-movement in expected excess returns across countries (i.e. the cross-
sectional restriction of the model). In a nutshell, it seems that (i) two or three
risk factors in combination with constant conditional betas suffice to character-
ize the cross-sectional variation in expected excess returns, and (ii) that the time
varying risk premiums on these factors capture a large fraction of the predictable
time variation in global stock returns.
Figure 5.2 provides an assessment of how well multiple latent variable models
fit the data on the basis of mean pricing errors. A pricing error is defined for
each country as the difference between the restricted predictions of excess re-
turns for a given model (either K:= 2 or K:= 3), using the GMM point estimates
of the model parameters. In particular, it is equal to the sum of the model error
in fitting the expected return, plus the deviation of the return from the 'true' ex-
pected return. To the extent that the sample mean return approximates the aver-
age expected return, the mean pricing error approximates the average model er-
ror. If the pricing error is positive (negative), the actual return is on average
higher (lower) than the expected return given the level of risk. Numbers are ex-
pressed in percent on a quarterly basis. The largest mean pricing error in the G7
sample is less than 2 percent per year, but pricing errors are a lot higher in the
non-G7 sample. Overall, the results compare favorably to the average rnispricing
in Connor/Korajczyk (1988), Ferson (1990), and Oertmann (1997). Surprisingly,
there is no evidence that the mean pricing errors are systematically smaller in a
three-factor latent variable model (K = 3) than in a two-variable model (K = 2),
implying that that the third latent variable does not help to fit global stock mar-
ket data. Rather, the third factor seems to increase the mean pricing error for the
set of non-G7 test assets, which is only hard to explain.
Figure 5.2
Pricing errors of latent variable models
2,-------------------------------------------____________-,
G7 test assets Non-G7 test assets
1.5
·0.5
·1 L -______________________________________________________- J
OK=2. K=3
The figure shows the mean pricing errors for the two subsamples of test assets in latent variable mod-
els with K=2 and K=3 factors. A pricing error is defined for each stock market as the difference be-
tween actual excess returns and restricted predictions for excess returns. expressed in per cent on a
quarterly basis. For a given number of latent factors in (5.11). GMM point estimates of the model pa-
rameters are used to generate the restricted predictions for excess stock returns.
Testing a conditional version of the consumption-based model 193
Table 5.4
Asset pricing models with two and three unspecified state variables
System of non-linear equations for GMM estimation:
_ ( UI,t+1
£1+1 - U 2.1+ 1 e HI )' -_ ( [1j.H' - Z, /), j' , )
[r,.,+, - Z,/),C]
Period: 1973.01- 1998.03 (quarterly data)
K=2 K=3
Table 5.4
Continued
df 24 15
Estimation results for the highly structured linear factor model in system (5.16)
with K = 1 state variable (denoted as 1.+1) are reported in table 5.6 and table 5.7.
Growth rates of the five different measures for global consumption compete as
Testing a conditional version of the consumption-based model 195
Table 5.5
Asset pricing models with two and three unspecified state variables
System of non-linear equations for GMM estimation:
£,+\:::
( )' ([r, 1+' - Z,O, j' )
U I,I+1 U 2.1+ 1 e \ + \ : : : ' ,
[r,.'+1 - Z,O,C]
Period: 1973.01-1998.03 (quarterly data)
K=2 K=3
Table 5.5
Continued
K=2 K=3
Reference assets: Switzerland Netherlands Switzerland Netherlands Hong Kong
df 36 24
50 See section 3.1 for a description of national consumption data. FersonIHarvey (1992)
use unadjusted consumption data, which is more volatile. However, their results indi-
cate that this cannot help even in models with time-separable preferences. See also sec-
tion 6.1.4.
Testing a conditional version of the consumption-based model 197
basis. In addition to the global instrument variables (plus a constant and the
lagged return on the world market), consumption rates with lags of two, three,
and four quarters are used to produce testable orthogonality conditions. In order
to avoid any bias from temporal aggregation, I deliberately choose not to use
consumption growth at lag one. 51 Again, the total sample is split into two sub-
samples to preserve the power of the OMM methodology: 07 countries and non-
07 countries. Swiss stock returns are used as the reference asset in both cases.
This specification of the model produces 9 x 8 == 72 orthogonality conditions
and requires 18 parameters to estimate for the 07 subsample, leaving 54 overi-
dentified restrictions. There are 12 x 8 == 96 orthogonality conditions and 21 pa-
rameters to estimate for the non-07 subsample, so that the OMM qui-square test
statistic for the goodness-of-fit has 75 degrees of freedom.
To sum up, the empirical evidence for a conditional version of the consumption
capital asset pricing model (CCAPM) is rather disappointing, but not disastrous.
Clearly, the results are a manifestation of the equity premium puzzle. Aggregate
G7 consumption rates are too smooth to explain the level of risk premiums on
global stock markets. Although the full restrictions of the model have been re-
jected, there is a consistent positive relationship between expected stock returns
and expected consumption growth, as shown by the positive estimates for the
constant conditional consumption betas. Admittedly, this is only a weak indica-
tion for the link between consumption smoothing and risk-correction. Earlier
tests of the consumption-based approach were not very favorable to the model
neither. However, most of these tests analyzed unconditional moment implica-
tions, while the present framework allows testing some of the conditional mo-
ment implications of consumption-based asset pricing. Recall, HansenlRichard
(1987) show that conditional mean-variance efficiency does not go through to
imply unconditional mean-variance efficiency, while the converse is true. 54
Equivalently, a conditional beta pricing model does not imply an unconditional
one. Therefore, testing some conditional moment restrictions of the consump-
tion-based asset pricing model for a global set of data is an important empirical
question.
Table 5.6
Asset pricing models with a single consumption beta
G7 countries: ~, ~, ~i ~, ~,
Switzerland 2.762** 2.149** 2.520** 0.142 1.693
2.407 2.125 2.197 1.471 1.501
Canada 2.958** 2.191 * 2.742** 0.123 1.452
2.226 1.923 2.068 1.405 1.413
France 3.823** 2.671** 3.831 ** 0.177 2.087
2.525 2.145 2.384 1.469 1.492
Germany 1.412 1.005 1.439 0.097 1.099
1.535 1.406 1.510 1.354 1.380
Italy 2.972* 2.169 2.944* 0.143 1.669
1.944 1.731* 1.869 1.365 1.375
Japan 3.799*** 2.762 3.715** 0.139 1.688
2.617 2.201** 2.483 1.460 1.470
United Kingdom 3.050** 2.270 2.812** 0.132 1.689
2.149 1.916* 1.971 1.394 1.430
United States 1.818* 1.407 1.559 0.084 1.106
1.706 1.600 1.487 1.290 1.349
GMM chi-square 73.613 74.460 68.125 103.119 70.203
df 54 54 54 54 54
p-value 0.039** 0.034** 0.094* 0.000*** 0.068*
The instrument variables are described in the text. The reported coefficients - the constant conditional
consumption betas - are estimated simultaneously using GMM; the t-statistics reported underneath the
coefficients (in italics) are heteroscedasticity consistent. *1**1*** denotes statistical significance at the
10%15%11 % level. The GMM test statistic for the goodness-of-fit is chi-square distributed with 54 de-
grees offreedom (df).
200 Chapter 5
TableS.7
Asset pricing models with a single consumption beta
For these reasons it seems worthwhile to construct mimicking portfolios for dif-
ferent measures of G7 consumption as a plausibility check. The composition of
the factor mimicking portfolio is determined by regressing consumption growth
rates on a group of asset returns. 58 The slope coefficients of such regressions are
used as the set of fixed portfolio weights. To avoid overfitting, I apply fresh data
to construct the maximum correlation portfolios. Specifically, the 38 industry
Empirical results for tests that use consumption mimicking portfolios as the state
variable in specification (5.16) are documented in table 5.8. Only the GMM qui-
square test statistics are reported, because the estimates for the constant condi-
tional betas are similar to those in tables 5.6 and 5.7. The p-values strongly force
me to reject the model. Taking a look at the results for the non-G7 subsample,
59 For a more in depth description of the specification of these industries and their exact
construction principles see 'Morgan Stanley Capital International Perspective', which is
published quarterly by Morgan Stanley & Co. Inc., New York.
60 See Chamberlain (1983), p. 131Of.
Testing a conditional version of the consumption-based model 203
the suspicion that the p-values reported in table 5.7 might suffer from low power
due to a proliferation of orthogonality conditions proves correct. Curiously, du-
rable goods again seem to be the most 'appropriate' single proxy for the latent
variables.
Table 5.8
Asset pricing models with consumption mimicking portfolios
System of non-linear equations for GMM estimation:
f'.1 =Z,1+11
rl.,.1 = Z,GI + (f'.1 -Z,1)~1 +&1.'.1
r,.'.1 = Z,GI (~II~,)+(f'.1 -Z,1)~, +&,.,.1
Period: 1973.01 - 1998.03 (quarterly data)
G7 countries:
df 54 54 54 54 54
Non-G7 countries:
df 75 75 75 75 75
p-value 0.000 0.000 0.001 0.098 0.000
The instrument variables are described in the text. The GMM test statistics for the goodness-of-fit are
chi-square distributed with 54 and 75 degrees of freedom for G7 and non-G7 stock markets, respec-
tively. *1**1*** denotes statistical significance at the 10%15%11 % level.
single system. Hence, it is possible to assess whether the previous results have
been influenced by arbitrarily splitting the entire sample into two subsamples.
Using a constant and the four global instrument variables, the system yields
18 x 5 = 90 orthogonality conditions. Given that there are only 18 parameters to
estimate, the GMM test statistic is qui-square distributed with 72 degrees of
freedom. Again, the pricing tests are conducted using both reported consumption
data and the associated mimicking portfolios. The results in table 5.9 are quali-
tatively similar to those in the previous tables. All models are rejected, and
mimicking portfolios again exhibit worse pricing properties than consumption
rates. The last column in the table contains the average of 18 country specific R-
squares from regressions of the residuals on the global instruments. There is
clear evidence of residual predictability, especially when mimicking portfolios
are used. If the model properly described the cross section of expected stock re-
turns, there should be no additional information in the residuals.
Table 5.9
Asset pricing models with a single consumption beta
System of non-linear equations for GMM estimation:
r,+. - PI.+. = 0
Period: 1973.01 - 1998.03 (quarterly data)
Consumption data:
The aim of this chapter has been to explore three important issues of interna-
tional asset pricing. First, I investigate whether time variation in expected re-
turns can be explained on the basis of time varying risk premiums, rather than
interpreted as signs of market inefficiencies. Second, testing the cross-sectional
restrictions of latent variable models, the concept of global stock market inte-
gration is addressed. This question directly affects the cost of capital for invest-
ments in different countries. Finally, I extend the latent variable framework to
test a conditional version of the consumption-based asset pricing model. The
findings of this chapter can be summarized as follows:
Predictability is evidence for time varying risk premiums: Beta pricing models
allow to determine whether the predictability of excess stock market returns is
evidence for either market inefficiency or time varying risk premiums on a small
number of risk factors. Latent variable models are much more parsimonious, but
impose meaningful cross-sectional restrictions on a sample of international stock
market returns.
• Models with two and three latent variables describe global stock market re-
turns reasonably well. I have argued that it is only the assumption of constant
conditional betas that makes the search for the mere number of factors a
meaningful question. The magnitudes of the pricing errors show no major
Testing a conditional version of the consumption-based model 207
advantage from going beyond two risk factors. The time variation in ex-
pected excess returns can be explained by time variation in economic risk
premiums on at least two global risk factors. Loosely speaking, this result
could be interpreted as empirical support for the theoretical formulations of
international capital asset pricing models by Solnik (1974) and AdlerlDumas
(1983).
Consumption growth rates do not qualify as a proxy for the latent factors: Mer-
ton's (1973) intertemporal capital asset pricing model (ICAPM) suggests that
rational investors hold a portfolio that is optimally hedged against changes in the
stochastic investment opportunity set. Breeden (1979) and Stulz (1981) simplify
the pricing equation and posit that assets are priced with respect to their covari-
ance with aggregate consumption. Thereore, a single factor beta model results.
A natural approach is to extend the latent variable framework and examine
whether various measures of aggregate G7 consumption satisfy as proxies for
the unobserved risk factors.
• Testing a conditional linear factor model, different measures for world con-
sumption rates fail to proxy for the underlying state variables. The estimates
of constant conditional consumption betas are as expected, however, GMM
tests for the model's overall fit are rejected. In a nutshell, a conditional ver-
sion of the consumption-based asset pricing model cannot explain the level
of the historical risk premiums on international stock markets, which is a
manifestation of the equity premium puzzle. The empirical evidence is not
sensitive to a particular specification of the model. In addition, using mim-
icking portfolios for G7 consumption instead of reported consumption does
not help. In other words, consumption-mimicking portfolios are conditionally
mean-variance deficient.
• More fundamentally, the results shed negative light on the underlying eco-
nomic intuition of the interplay between consumption smoothing and risk
correction. The empirical findings in this chapter do not give credible support
to the notion in FamalFrench (1989) that the consumption-smoothing be-
havior of rational investors accounts for the time variation in economic risk
premiums. Negative evidence has already been reported in earlier literature.
208 Chapter 5
There is evidence that international stock markets are integrated: Capital mar-
kets are said to be integrated if financial assets with the same risk characteristics
have identical expected returns, irrespective of the market to which they belong.
The problem inherent in all empirical test of integration is the joint nature of the
underlying assumption. Rejecting the null hypothesis of integration could possi-
bly have two reasons: either markets are segmented and/or the underlying pric-
ing model does not correctly describe the risk-return relationship. The cross-
sectional restriction imposed by the latent variable model allows testing for inte-
gration on the basis of (almost perfect) correlation between expected returns on
global stock markets. The encouraging results for the latent variable models with
two and three unobservable factors provide evidence in favor of global stock
market integration. Two latent state variables suffice to obtain near perfect cor-
relation of expected excess returns on global stock markets, with the constant
conditional beta factors as the coefficients of proportionality. The ultimate test
for integration from a risk-sharing perspective is to use global consumption
rates. Unfortunately, this model is rejected.
It could be argued that one reason for the failure of the consumption-based
model is the poor quality of the data. More fundamentally, another possibility is
that the model imposes too many restrictions on consumption rates at the same
time, a problem closely related to the equity premium puzzle. In chapter 6 vola-
tility bounds for stochastic discount factors will be introduced, a concept origi-
nally proposed by Hansen/Jagannathan (1991). Their framework allows to reex-
amine the results of this chapter in stochastic discount factor language. Finally,
in chapter 7 I introduce a model that does no longer attempt to explain the level
of historical equity premiums, but rather explores the lessons we can still learn
from consumption-based valuation models. We cannot simply abandon these
appealing models in general and still believe that portfolio-based models do
hold. Those latter models are not derived by the assumptions of explicit frictions
that de-link consumption from asset returns.
209
Chapter 6
valuation model. More general, knowing a lower bound for the volatility of sto-
chastic discount factors that are consistent with a given set of asset returns offers
a completely new way of judging other commonly known asset pricing models.
The first of several obvious questions that arise in this context is whether it is
harder to simultaneously price international assets rather than domestic assets
alone. Assuming that the law of one price holds, a discount factor that perfectly
prices all assets always exists in a complete market, irrespective of the specific
set of assets. The question is, however, whether the qualitative properties of a
'local' purely Swiss discount factor and a 'global' discount factor are signifi-
cantly different, as judged by the minimum volatility that is necessary to price a
given set of assets. BekaertlHodrick (1992) find that international diversification
imposes stronger pricing restrictions, but they do not report any measures of
statistical significance. From this point of view, the test I propose in the empiri-
cal part of this chapter can be interpreted as an unparametric test of market inte-
gration, i.e. a test that does not assume a specific asset pricing model. 1 Another
major issue in international finance is the well known argument that investing in
foreign assets reduces the volatility of domestic portfolios. This observation is
attributed to low correlations between equity returns in different markets? How-
ever, very few studies have tested whether the increase in the risk-return spec-
trum is significant or simply lost in sampling error. The relevant question is
whether or not investors can mimic foreign stock returns with domestically
traded securities alone. In this chapter I also test whether a very simple currency
hedging strategy helps to further increase the risk-return spectrum available to a
global investor. This is an important practical question asset managers are con-
fronted with. Finally, I include both developed and emerging stock markets to
perform asset pricing related versions of spanning tests. Recently, emerging
markets have attracted a lot of attention from both researchers and practitioners.
Speidel/Sappenfield (1992) and DivechaJDrachlStefek (1992) argue that there is
a 'diversification free lunch' available for investors in emerging markets. Dro-
betz (1999) shows that emerging markets seem very attractive on an ex-ante ba-
sis, but are largely responsible for negative ex-post performance of globally di-
versified portfolios. Hence, the inclusion of emerging equity markets enables me
to present a fuller picture of the true benefits of international diversification.
urement. 7 In this chapter I will make use of the HI regions to perform mean-
variance spanning tests in the spirit of De Santis (1995) and Bekaert/Urias
(1996). This section provides the theoretical framework. First, it is shown how
security returns can be used to derive an efficiency region for stochastic discount
factors that are consistent with asset pricing data, completely divorced from any
parametric specification. The discussion follows closely along HansenlIaganna-
than (1991). Second, their methodology can be generalized to incorporate con-
ditioning information, as forcefully demonstrated by Cochrane (1996). Third, I
show how the equity premium puzzle can be refOlIDulated in this new frame-
work. The attempt to solve the puzzle by simply searching for more volatile
candidate discount factors reveals another problematic issue, which Co-
chraneIHansen (1992) call the 'correlation puzzle'. Finally, I demonstrate that
my empirical test design described in a later section exploits the duality between
mean-standard deviation frontiers for portfolio returns and mean-standard de-
viation frontiers for investors' marginal rate of substitution.
(6.1)
of the same equation must hold for the expectations E(.lfI>,) and EO. Hence,
conditioning down gives: 9
(6.2)
The sample counterparts of the orthogonality conditions derived from (6.2) form
the basis of many tests using the Generalized Method of Moments (GMM) de-
veloped by Hansen (1982). However, it is only the specific parameterization of
m'+l that gives the equation its empirical content. In particular, if m,+l is a linear
function of the return on the market portfolio, the capital asset pricing model
(CAPM) falls out. Merton (1973), Breeden (1979) and Fama (1996) showed that
in the context of an intertemporal capital asset pricing model (ICAPM) m,+l can
be defined as a linear combination of hedging portfolio returns. Finally, many
empirical studies define m,+l as a linear function of a number of observable (but
rather ad-hoc) fundamental or macroeconomic factors. ChenIRolllRoss (1986)
and FersonIHarvey (1994) empirically pursued this approach in an unconditional
framework, to name only a few. Cochrane (1996), FersonIHarvey (1993), Fer-
sonlKorajczyk (1995), and Dertman (1997) are examples for studies that further
condition the observable factors on well-known information variables. By speci-
fying proper utility functions, the stochastic discount factor setup also subsumes
dard deviation of any valid stochastic discount factor, using the returns on a
given set of securities only. Their procedure can be regarded as a generalization
of the early work by Shiller (1979, 1981). Starting at the constant discount-rate
present value model, Shiller found that the pricing kernel implied by this simple
framework was not volatile enough to explain the differences in expected re-
turns. Generalizing the model by allowing for time varying discount rates helps,
but volatility tests still reject, as reported in Campbell/Shiller (1988a,b ).15
(6.3)
where Et+1 is the error term in the regression with E(Et+,Rt+d = o. The standard
least-square formula for the N x 1 vector of regression coefficients Jl gives:
(6.4)
Without direct data on the stochastic discount factor, however, the regression
coefficients cannot be estimated in the usual fashion. But one can exploit the
fact that mt+1 must be a valid discount factor to infer the sensitivities. Multiply-
ing both sides of (6.3) by R t+" taking expectations on both sides, and applying
that E(mt+,Rt+,) == 1 yields an expression for (J. Substituting the regression coef-
ficients back into (6.3) gives the fitted values m;+1 of the projection as:
(6.5)
Notice that m;+1 is a valid stochastic discount factor, because using m;+1 instead
of mHI perfectly satisfies the Euler-equation in (6.2). m;+1 can as well as m tH
(6.6)
This implies that var(m,+l) ~ var( m;+l)' which is the key intuition of HJ lower
bounds on the variance of m'+l. While the true m'+l is unknown, m;+l depends
only on the matrix of second moments of the N asset returns and, hence is ob-
servable by the econometrician. Contrary to the usual approach in the literature,
no assumption about a specific pricing model has been required to derive (6.6).
Substituting from expressions (6.3)-(6.5) and omitting time subscripts yields:
var(m) ~ var(m*)
=p'var(R)p
= (cov(m,R')var(Rr1)var(R)( var(Rr 1cov(m,R')), (6.7)
Finally, substituting from (6.8) into the right hand side of inequality (6.7) yields
the lower bound for the variance of a valid discount factor:
The covariance term in equation (6.8) is zero for the risk-free asset, and so
E(m) = Rf"l. As the hypothetical values of E(m) = Rf"l are varied over the real
line, equation (6.9) describes a parabola in [E(m), var(m)] space. For a given set
of asset returns which determine E(R) and var(R), inequality (6.9) describes a
lower bound for the variance of m'+l as a function of its mean. Taking the
square root of equation (6.9) determines the HJ bound in a more familiar mean-
standard deviation space. The necessary (but not sufficient) condition is that any
valid stochastic discount factor must have a mean and a standard deviation that
places it within the parabola. However, the reverse argument does not necessar-
ily hold, as shown in section 6.1.4.
to test whether the addition of small firm returns expands the mean-variance
frontier beyond the spectrum available using large firms' returns only.
Figure 6.1 demonstrates how the equity premium puzzle can be explored within
the HJ framework. The solid line traces the parabolic HJ region for valid sto-
chastic discount factors, using quarterly stock returns for the sample of G7
countries from chapter 5 and a GDP-weighted 90-days Eurocurrency interest
rate (all denominated in Swiss francs). It is important to note again that this re-
gion is derived by a simple projection argument and, hence does not rely on any
existing asset pricing model. The dotted line plots means and standard devia-
tions for the marginal rate of substitution generated by the canonical consump-
tion-based model with time separable utility. The time discount factor is set ar-
bitrarily to ~ = 0.97 , and the degree of relative risk aversion 'Y varies along the
bend of the curve. Quarterly growth rates of aggregate G7 consumption are used
in the computation. The result is striking. Excessive curvature coefficients are
required to generate the volatility of the discount factor as implied by the HJ eq-
uity premium region. 19 The points do not enter the feasible region until relative
19 In their classic paper, FriendIBlume (1975) argue that realistic coefficients of relative
risk aversion should not exceed 10.
Volatility bounds for stochastic discount factors 219
risk aversion "( reaches a value of well above 200(!). This is definitely implausi-
ble. The reason for the failure of the consumption-based model becomes evi-
dent. Consumption rates are simply not volatile enough to qualify as a candidate
stochastic discount factor.
Figure 6.1
Feasible region for stochastic discount factors implied by G7 data
.
5.00
4.50
4.00
,,(>200 y
3.50
3.00
E 2.50
---.
.
"6'
2.00
1.50
1.00
( ri
0.50 ......... .---:----
.... . .
0.00
0.5 0.6 0.7 0.8 0.9 1.0 1.1
E(m)
The figure shows a modem representation of the equity premium puzzle, originally proposed by Me-
hralPrescott (1985). The solid line traces the parabolic 81 region for valid stochastic discount factors,
using stock returns for a sample of G7 countries and a GDP-weighted 9O-days Eurocurrency interest
rate (all denominated in Swiss francs). The dotted curve displays means and standard deviations for
the stochastic discount factor, assuming power utility with a time discount factor ~=O.97 and varying
degrees of risk aversion y. Quarterly growth rates of aggregate G7 consumption are used.
(6.10)
21 For yet a completely different approach to assess alternative discount factors see the
'distance measure' proposed by HansenlJagannathan (1997). They show that the maxi-
mum possible pricing error, expressed as expected return error per unit of standard de-
viation, is proportional to the standard deviation of the difference between the true and
the false discount factor. Their approach accounts for the 'correlation puzzle' and
avoids to reward variability of candidate stochastic discount factors (see section 6.l.4).
22 See section 2.4.5.
Volatility bounds for stochastic discount factors 221
gressing a frontier m;+1 onto the space of payoffs (or equivalently, returns) is
one. In fact, most general equilibrium models assume only one common source
of shocks, so that consumption growth, returns, and the discount factor become
perfectly correlated conditionally.23 Empirically, however, candidate discount
factors implied by standard asset pricing models are not highly (unconditionally)
correlated with returns, that is, they would produce R-squares substantially be-
low one. For example, take the series of aggregate G7 consumption rates from
chapter 5. The correlation between consumption growth and the return on the
world stock index has only been 0.14 over the period from 1973.01 to 1998.03.
(6.11)
where R2 denotes the goodness of fit from regressing mt+1 on the vector as asset
returns R'+I as in (6.3). This relation reveals a clear tradeoff: if a candidate dis-
count factor is less than perfectly correlated with the return vector, and so
R 2 < 1, it must be more volatile than implied by the HJ bound. CochranelHansen
(1992) construct iso-R-square contours above the original HJ bounds, resulting
in a series of tighter bounds for the minimum volatility of mt+1 as the projection
R-square decreases?4 Again, the intuition is that correlation is the driving con-
cept of asset pricing. A valid discount factor must either be highly correlated
with asset returns, or have a higher standard deviation than indicated by the
minimum HJ bound for a given set of assets. But a discount factor that is com-
pletely uncorrelated with asset returns will fail any pricing test. In other words, a
candidate discount factor can have lots of volatility and perfectly fall into the HJ
bound, but still price miserably if it is uncorrelated with the returns on a given
set of assets. This might explain the results of Mankiw/Shapiro (1986). They
report that market betas drive out consumption betas in mUltiple regressions. In
a habit formation framework that is calibrated to hold by construction, Camp-
belVCochrane (1999b) show that the capital asset pricing model (CAPM) per-
fonns better than the standard consumption-based model because the market
return is more closely unconditionally correlated with the marginal rate of sub-
stitution than is consumption growth - even when a slightly more complex con-
sumption-based model holds by construction. 25 As a final remark, it should be
noted that HJ bounds prove extremely useful to evaluate standard consumption-
based models with a unique series for the discount factor (e.g., the analysis in
figure 6.1). However, volatility bounds are less useful for assessing factor mod-
els. It is easy to add just another risk factor and generate sufficient volatility for
the pricing kernel to fall into the HJ feasible region. However, these models are
rather ad-hoc and do not claim to have a model for the discount factor. 26
Time subscripts are not needed in this section, so they are again omitted for
convenience. Subtracting both expressions gives the Euler-equation in tenns of
excess returns, denoted by r:
~( m »R-1
v
E(r') -R-1ISRI
- r cr(r') - r . (6.17)
For each point on the efficient frontier there is a unique pair (R r ,ISRI), and a
corresponding point on the volatility bound which is (ISRI/R r , l/R r ). Hence,
both curves contain the same set of information. This is shown graphically in
figure 6.2, which is adapted from De Santis (1995).28 As the values of E(m) are
varied, the slope of the tangent to the efficient frontier becomes lower and the
HJ bound declines. At the mean return corresponding to the minimum variance
portfolio, the bound attains its minimum. 29 Continuing, the Sharpe ratio rises
again, and so does the bound.
To sum up, the slopes of rays from the origin to a particular point on the HJ
bound are equal to the Sharpe ratios of the specific risk-rate plus tangency port-
folio combinations on the efficient frontier. This analysis directly corresponds to
the results derived by Zimmermann (1999).30 He also demonstrates that the dy-
namics of the stochastic discount factor are sufficiently characterized by the
risk-free rate and the market price of risk. Later on in this chapter I will intro-
duce a test that measures by how much the addition of international assets raises
the HJ bound for a portfolio of merely Swiss assets. The discussion in this sec-
tion should have made clear that this is exactly the same thing as asking by how
much these additional assets expand the mean-standard deviation spectrum
available to a global investor.
29 If the risk-free rate reaches the variance of the minimum-variance portfolio (MVP), the
tangency line becomes the asymptote to the efficient frontier. This is the minimum
Sharpe ratio attainable. According to (6.16), the absolute value of the asymptote's slope
equals the slope of the ray from the origin to the minimum standard deviation point on
the HI bound.
30 See Zimmermann (1999), sections 14 and 15, p. 72f.
31 See section 2.10.
Volatility bounds for stochastic discount factors 225
Figure 6,2
Hansen-Jagannathan bound and efficient frontier
0.4.----.---------------------~-__.,
0.35
[E(m),a(m)]
0.3
0.25
~
E
0' 0.2
0.15
0.1
Ray from origin with slope
0.05 = Sharpe Ratio (SR)
E(m)=I/R,
o+------~------_.------~~~~~-~
0.995 0.9955 0.996 0.9965 0.997
E(m)
0.06
0.055
0.05
2' 0.045
ill
0.04 E(R,)
0.035
a(R')
0.03 +-----.------'-r-'--'----,-----.,------,-----i
o 0.05 0.10 0.15 0.20 0.25 0.30
a(R)
The figure shows the link between the HansenlJ agannathan (1991) region for valid stochastic discount
factors and the efficient frontier for portfolio returns. There is a one-to-one relationship between the
two concepts, and both curves contain the same information. As the values of E(m)=llRr are varied,
the slope of the tangency to the efficient frontier (Sharpe ratio) becomes lower and the H1 bound de-
clines. If the risk-free rate attains the return on the minimum variance portfolio, the bound attains its
minimum. Continuing, the absolute value of the Sharpe ratio rises again, and so does the bound.
226 Chapter 6
tradeoffs. But knowing that Sharpe ratios and HJ bounds are related concepts,
higher risk-adjusted returns for any given risk-free rate of interest imply tighter
restrictions on the volatility of valid stochastic discount factors. To see this, let
again Zt be an L-dimensional vector of instruments contained in the informa-
tion set Wl> so that Zt C Wt . The space of scaled returns Rt+1 ® Zt can in prin-
ciple be infinite dimensional. Then equation (6.1) implies:
(6.18)
To denote the exact timing of the inflow of relevant information, time subscripts
are needed again. Taking unconditional expectations and applying the law of
iterated expectation results in:
(6.19)
Cochrane (1996) suggests to add managed portfolio payoffs and to proceed with
unconditional moments as if conditional information did not exist. In fact, the
Euler-equation in (6.2) must hold for these payoffs as well, and one can compute
unconditional HJ volatility bounds for scaled returns in the usual manner. In the
empirical tests that follow below I adopt simple linear scaling. But recent re-
search shows that linearity - as assumed in (6.18) - is not important. In fact,
BekaertlLiu (1999) show that there is an optimal scaling function (in the sense
that it results in the tightest HJ bound), which is decreasing in the conditional
variance of stock returns, but not monotonic in the conditional mean. 34 Simi-
larly, FersonlSiegel (1999) show that the optimal strategy puts less weight in the
risky asset when conditional expected returns become extreme. 35 Without going
into the details, their findings are intuitive. Scaling will only improve the HJ
bound significantly if the weights applied contain infonnation about future re-
turns. Scaling by an independent random variable merely adds noise to a given
set of returns, but does not provide any additional pricing infonnation. 'Stacked'
bounds are computed using both scaled and unscaled returns, imposing addi-
tional restrictions on the projection in (6.3). Intuitively, stacked bounds can
never deteriorate compared to the original HJ volatility region. More important,
adding noise (i.e. conditioning infonnation without infonnational content) will
not result in a significant upward shift of the bound. In a nutshell, the space of
payoffs only increases if the instrument variables are somehow correlated with
future returns. In this case, stacking implies sharper volatility bounds. 36 This no-
tion also puts a practical limit on the number of instruments in Z, one needs to
scale by. In particular, the instrument variables must characterize the conditional
distribution of stock returns. Given the evidence in chapter 4, I use the world
dividend yield and the G7 tenn spread in the empirical analysis. More instru-
ments are only hard to incorporate due to the rapid proliferation of GMM-based
orthogonality conditions. 37 Finally, it must be noted that the scaling approach
cannot capture all implications of conditioning infonnation for the unconditional
version of a specific model. However, Cochrane (1999a) warns that this lack of
accuracy should be carefully weighted against the danger of possibly incorrect
parametric models for the conditional distribution of returns, as in the Han-
senlGallantffauchen (1990) approach.
Again, let RI+' be the Nxl vector of asset returns observed by the econometri-
cian. All the assets included in R,+, define a mean-variance efficient frontier in
the spirit of Markowitz (1959). Mean-variance spanning tests ask whether there
exists a subset of assets that span the entire mean-variance frontier by itself. For
230 Chapter 6
(6.22)
38 See Huberman/Kandel (1987), equations (3a) and (3b) and proposition 3, p, 877.
Volatility bounds for stochastic discount factors 231
(6.23)
(6.24)
RO.1+ 1 denotes the return on some additional benchmark asset, e.g., the risk-free
rate of return. As in chapter 5, II denotes an N x K matrix of betas, conveniently
partitioned as II = (111'112)' Finally, ((), again captures the information set avail-
able at time t. The econometrician cannot estimate both betas simultaneously
without further specifying the model. Similar to the analysis in section 5.1.2,
however, there exists an [(N -K)xK] matrix of constant coefficients C such
that: 40
(6.25b)
Expressions (6.25a) and (6.25b) reveal that the existence of K latent variables
imposes restrictions on the C matrix that are equivalent to those imposed by
mean-variance spanning. In particular, the residual orthogonality condition in
(6.25b) can be replaced with:
subject to the constraint that the c coefficients sum to one for each test asset i,
with i = K + 1, ... , N. This specification allows to test the null hypothesis of un-
conditional mean-variance spanning using the Generalized Method of Moments
(GMM). In addition, incorporating a vector of predetermined instrument vari-
ables Z, C ((), allows to verify the null hypothesis of conditional mean-variance
39 For the detailed derivation see FersonIFoerster/Keim (1993), equation (10), p. 136.
40 Of course, it is assumed that the fundamental beta pricing restriction in (5.1) holds.
232 Chapter 6
hypothesis of spanning was rejected, this would seem to suggest that home bias
in portfolio holdings can be explained by statistical uncertainty (at least from a
Swiss perspective).
Define XI and X2 as two payoff spaces spanned by R I.1+1 and R 2.1+1, respec-
=
tively. Let X XI + X2 be the payoff space obtained by combining R I.1+1 and
R 2.t+I' Obviously, one can construct HJ volatility regions for any given set of
payoffs separately, i.e. for XI' X2' and X. Denote the discount factors associ-
ated with each one of these projections as m;.I+I' m;.I+I' and m:+ I, respectively.
However, it is more interesting to explore whether the HJ volatility bound for
the payoff space of spanning assets XI is statistically distinguishable from the
lower bound associated with the combined payoff space of spanning and test
assets, denoted as X. In other words, the question is whether the qualitative
properties of a 'local' Swiss stochastic discount factor and a 'global' stochastic
discount factor are significantly different, as judged by the minimum volatility
that is necessary to properly price both sets of payoffs. The question is not
whether a common discount factor exists or not for the combined payoff space
X. As has been argued in section 2.4.5, if the law of one price holds, a unique
stochastic discount factor that prices all payoffs simultaneously always exists in
the payoff space Xc 9\s. But the stochastic properties of m;.t+1 and m;+I' as de-
scribed by their minimum volatility required to qualify as valid stochastic dis-
count factors, may be very different. Finding that it is 'harder' to price interna-
tional asset than Swiss asset alone (harder in the sense that an extremely volatile
stochastic discount factor is required in a global context) would raise doubt
about international stock market integration. However, any results can only be
understood as indicative, because a particular pricing model has not been im-
posed.
When a new set of test assets is added to the set spanning assets (that is, going
from XI to X), the key question is whether the HJ bound, which indicates the
lowest volatility a valid stochastic discount factors must exhibit, changes in a
statistically significant way. It was shown in equation (6.6) that the bound on
cr(mt+l) was obtained from a projection of m on a constant and the space of
payoffs. XI is a subset of X, XI eX, so the projection argument imposes addi-
tional restrictions on the pricing kernel as we go from XI (the set of spanning
234 Chapter 6
assets) to .K (the combined set of spanning and test assets). Therefore, the
bounds associated with .K will be at least as tight as those associated with .KI'
Again, the analogy to the textbook mean-variance world is straightforward.
Adding additional assets always shifts the efficient frontier to the left, thereby
increasing the risk-return tradeoff (or Sharpe ratio) an investor faces. Taking
into account that there is a one to one relationship between both concepts, the HJ
bound must become sharper. This relationship is shown in figure 6.3, which is
again adapted from De Santis (1995).42 The upper figure demonstrates the up-
ward shift of the HJ bound, the lower figure the corresponding change of the
efficient frontier, both in a mean-standard deviation space. A natural procedure
for an econometric test is to ask whether the stochastic discount factor m;+I'
which prices all assets in .K, can be modeled as a linear function of the test as-
sets in .KI alone. Specifically, consider a slightly different version of the projec-
tion in equation (6.3). Assume now that a regression of mHI is run onto a con-
stant c and the vector of demeaned (unexpected) returns, that is:
(6.27)
with R;+I =(R;.HI' R;.HI ) and II =(II; II;). The null hypothesis of mean-
variance spanning asserts that the N-dimensional vector R '+1 is priced by a lin-
ear combination of the K spanning assets included in R 1.H1 . This would imply
that all N - K coefficients of test assets in 112 are equal to zero. In other words,
the returns on the spanning assets R 1.,+1 suffice to mimic the risk-return spec-
trum associated with the returns on the set of test assets R 2 •, +I. Taking the un-
conditional expectation of (6.27) and using the fact that E(m,+I) = Rr.IHI gives:
(6.28)
Volatility bounds and efficient frontiers are related concepts, hence the test can
equally be interpreted in the traditional mean-standard deviation framework for
portfolio returns. In particular, for a given value of c the test can be used to in-
vestigate whether the risk-return spectrum associated with RI.I+ 1 is tangent to
that associated with R '+1 at the point with the highest Sharpe ratio.
Figure 6.3
Volatility bounds and the addition of new assets
0.4~-----r~----------------------------------------~r----'
0.35
[E(m),o(m}]
0.3
0.25
~
E
t5' 0.2
0.15
0.1
Ray from origin with slope
0.05 = Sharpe Ratio (SR)
E(m)=l/R,
O+-----------~------------~----------~~~~~--~
0.995 0.9955 0.996 0.9965 0.997
E(m)
0.06
0.055
0.05
[ G(R),E(R)]
20.045
iif
0.035
0.03 -I--------r------Lr-------r--------r-------r---------I
o 0.05 0.10 0.15 0.20 0.25 0.30
o(R)
The upper figure shows the H1 region for valid stochastic discount factors and illustrates the change of
the frontier due to the addition of new assets. The lower line traces out the Hansen-Jagannathan (1991)
bound for Swiss stocks and the 9O-days Eurocurrency deposit denominated in Swiss francs (payoff
space XI). The upper line describes the bound that results from an addition of stocks from France,
Germany, the United States, and the United Kingdom, all denominated in Swiss francs (payoff space
X). The lower figure demonstrates the equivalent analysis in the traditional mean-standard deviation
space for portfolio returns.
236 Chapter 6
In other words, the test explores whether a portfolio including only R 1.H1 has the
highest Sharpe ratio conditional on the choice of the risk-free rate. A statistically
significant shift of the HJ volatility bound equivalently implies a significant in-
crease in the Sharpe ratio, i.e. an expanded risk-return spectrum available to a
diversifying investor. The idea to test whether the two bounds are tangent at the
point E(m) = c captures exactly what HubermanlKandel (1987a) call intersec-
tion. 43 There is, however, a severe complication: the econometrician usually
does not observe the value of c. Assuming some arbitrary value for c and run-
ning the test is of no help. Being tangent at this arbitrary point does not imply
that the two frontiers intersect for other values of c as well; in particular, it is
possible that they do not intersect for the true value of c. Obviously, one could
test the null hypothesis for all possible values of c, but this would be very cum-
bersome. To get around the problem, De Santis (1995) and BekaertlUrias (1996)
invoke the two-fund separation theorem. 44 Loosely speaking, this well-known
theorem from portfolio theory suggests that any frontier portfolio can be ob-
tained as a linear combination of two other frontier portfolios. Hence, if the
frontiers are tangent in any two points, they must coincide at all points. Running
the test then for two arbitrary (but different) values of c implies testing whether
the bounds coincide at all points. This is what Huberman/Kandel (1987a) refer
to as mean-variance spanning. 45 Formally, defining c1 and c 2 as two arbitrary
values of E(m,+,), such that C 17; c2, and plugging (6.27) into the Euler-equation
in (6.2), the orthogonality conditions for a GMM-based test of unconditional
mean-variance spanning are as follows: 46
The test in (6.29) has two advantages over the original test proposed by Huber-
man/Kandel (1987a). First, it is well known that stock returns deviate from nor-
mality. In particular, they are fat-tailed empirically. This problem is particularly
severe for emerging markets' stock data, which will be used in the empirical
analysis below. The convenient feature of GMM-based tests is that they do not
require asset returns to be normally distributed. Second, and more important
from an economic point of view, the system in (6.29) directly incorporates asset
pricing implications via E(m,+,Rt+') = 1. Even better, incorporating conditioning
information by scaling returns implies testing a fully conditional model in this
special setup. Recall from section 2.10, a conditional model is a model of the
form m'+1 = b;f,+" satisfying E(m,+,Rt+') = 1. This is in contrast to unconditional
fixed-weight factor pricing models, i.e. models where the b coefficients do not
carry time subscripts. To test some implications of fully conditional models, Co-
chrane (1996) suggests scaling the b parameters with proper instrument vari-
ables. Again, a linear model is not restrictive, but easy to implement. With a sin-
gle factor 1.+1 and a single instrument Z, this can be written as follows:
(6.30)
(6.31)
Since the 'factors' are already mean-adjusted returns (on the test assets), scaling
factors is already implied by adding conditioning information, that is, by scaling
returns. Given the results in chapter 4, I use two instrument variables: the world
dividend yield and a G7 term spread. With these two instrument variables, there-
fore, a conditional two-factor model (with Swiss stock returns and the 90-days
Eurocurrency interest rate denominated in Swiss francs as the spanning assets)
implies an unconditional six-factor model. This is also the idea that underlies the
influential paper by Jagannathan/Wang (1996).
I use fresh data to perform my analysis. Since the current framework no longer
requires consumption rates, I collect monthly (instead of quarterly) time series
of stock price indices from 1973.06 to 1998.08. This clearly enhances the power
of GMM-based tests. Postponed in chapter 4, this section gives a thorough de-
scription of the data set.
ity and free float. Finally, there is some attempt to ensure that the index reflects
the industry characteristics of the overall market. Hence, about 60 percent of
each industry group are targeted for inclusion in each MSCI country index.47
The indices are all market capitalization weighted on a total return basis, in-
cluding the dividends paid. Cumby/Glen (1990) further find that 99 percent of
non-U.S. stocks covered in the MSCI-world index are readily purchasable by
non-nationals. Hence, investors should have been able to implement most of the
strategies tested below.
[Fe stock market indices: Monthly data on emerging markets' equity indices for
the period from 1976.01 to 1998.08 are from the International Finance Corpora-
tion (IFC). In the past, IFC's definition of an emerging stock market was aligned
only to an emerging economy criterion: if a country's GNP per capita did not
exceed the World Bank's threshold for being a high income country, i.e. if a
country was eligible to borrow from the World Bank, its stock market was said
to be emerging. However, for the future the IFC intends to formulate a new
definition that takes both economic and stock market criteria into account. IFC
indices are intended to represent the performance of the most active stocks in
their respective stock markets and to be the broadest possible indicator of market
movements. The target aggregate market capitalization of IFC index constituents
is 60 to 75 percent of the total capitalization of all exchange-listed shares. IFC
indices do not take foreign investment restrictions into account and do not at-
tempt to replicate the composite, regional, or industry balances in overall market
capitalization.48 The indices are again all market capitalization weighted on a
total return basis, including the dividends paid. The emerging markets I include
in my analysis are Argentina, Brazil, Chile, Greece, India, Korea, Malaysia,
Mexico, Pakistan, Taiwan, Thailand, Turkey, Venezuela, the Philippines, and
Portugal.
47 Further details on market coverage and construction principles are provided in "Morgan
Stanley Capital International: Methodology & Index Policy", New York.
48 Further details on market coverage and construction principles are provided in ''The IFC
Indexes", published annually by the International Finance Corporation, Washington,
D.C.
240 Chapter 6
In the developed markets (arithmetic) mean returns in Swiss francs range from
16.286 percent (in Sweden) to 7.714 percent (in Singapore). The range is sig-
nificantly larger across emerging markets: an average annual (arithmetic) return
of 54.241 percent in Argentina is contrasted to an average annual loss of -9.342
percent in Indonesia. 49 There is no market among the MSCI sample that boosts
an arithmetic average of over 20 percent. In contrast, 9 out of the 18 IFC mar-
kets' mean returns exceed 20 percent (Argentina, Brazil, Chile, Columbia,
Mexico, Philippines, Portugal, Taiwan, and Turkey). These higher returns do not
come at zero cost. Emerging market returns are characterized by very high vola-
tilities, which explains the huge differences between the arithmetic and the geo-
metric returns for most of the IFC markets. The most dramatic example is Ar-
gentina, where the arithmetic average is 54.241 percent, but the geometric aver-
49 It should be noted, however, that the Indonesian sample is the shortest of all emerging
markets, starting only in 1990.01. It is not used in the empirical analysis below.
Volatility bounds for stochastic discount factors 241
age is only 18.766 percent. 50 Annualized volatilities in the IFC sample range
from 82.05 percent in Argentina to 29.16 percent in India. Volatilities in the de-
veloped markets are between 17 percent and 39 percent. The latter maximum is
for Hong Kong, which seems to be more of an exception rather than a represen-
tative value. The reported autocorrelations in the last column of the table meas-
ure persistence, that is, the predictability of market returns on the basis of past
market returns. There are six markets in the IFC sample that exhibit autocorre-
lations greater than 20 percent. This suggests that returns in emerging markets
are easier to predict on the basis of past information. 51
For the sample of MSCI markets I also report hedged returns, assuming that a
Swiss investor totally covers his or her exposure to currency risk. The return in
Swiss francs on foreign asset i from a unitary hedge strategy, denoted R~I+I' is:
R I,HI
h -R F.,t-S"I+I (6.32)
- 1,1+1 + S '
I.t
where R,.1+ 1 is the (uncovered) Swiss franc return on the market index from
country i, F..l denotes the time t forward Swiss franc price for foreign currency i,
and Si., is the spot price of currency i at time t. (F.., - S,,'+1 )lS", is the normalized
return on a short position in the forward contract (all returns are in log-terms).
Unfortunately, direct one-month forward prices for the full sample period are
only available for Canada, France, Germany, Japan, Netherlands, the United
Kingdom, and the United States from Data Resources Incorporated (DRI).52
Much shorter series are available for the remaining countries. Any missing for-
ward prices are reconstructed using covered interest rate parity:
(6.33)
50 Goetzman/lorion (1999b) argue that recent returns may not give a representative picture
of the expected performance of emerging markets. Emerging markets' return series suf-
fer from a 'survivorship bias', For example, Argentina was a very active stock market at
the beginning of the century, but ceased to exist afterwards to reemerge only in the
1970s. Their simulations show that recently emerged markets should have high ob-
served returns, and that the brevity of a market history is related to the bias in returns.
51 See Harvey (1995) and KirnlSingal (2000).
52 Forward prices are computed as averages of bid and ask prices.
Table 6.1
Summary statistics for stock returns I~
Period: 1973.06 - 1998.08 (monthly data)
Argentina 54.241 18.766 82.049 55.205 21.234 80.002 -0.01 0.021 -0.086
0'
1976.01 '"en
Brazil 1976.01 24.080 7.411 57.888 24.892 9.879 54.879 0.015 0.006 -0.011 0()
::r
Chile 1976.01 28.639 21.237 37.466 23.705 23.705 34.745 0.159 0.142 0.057 I>l
en
Columbia ~.
1985.01 24.469 10.822 29.436 28.119 23.870 27.692 0.294 0.127 -0.024
0-
Greece 1976.01 9.972 4.481 32.569 12.447 6.949 32.509 0.133 0.078 0.005 <n'
()
0
India 1976.01 13.113 8.780 29.158 15.052 11.248 27.212 0.152 0.011 -0.039
g=
Indonesia 1990.01 -9.342 -20.271 46.194 -10.201 -19.952 45.659 0.254 -0.067 -0.010
~
()
Korea 1976.01 11.892 4.936 36.803 13.561 7.404 34.646 0.031 0.062 0.069 0...
en
Malaysia 1985.01 1.646 -5.066 37.026 4.429 -1.018 33.274 0.174 0.016 0.024
Mexico 1976.01 23.448 11.417 50.681 24.055 13.885 46.541 0.225 -0.057 -0.029
Pakistan 1985.01 5.605 0.643 31.693 8.747 4.691 28.618 0.229 -0.049 -0.053
Phillipines 1985.01 23.717 16.127 38.528 26.929 20.176 36.024 0.329 0.058 0.090
Portugal 1986.02 30.321 22.826 37.382 32.386 25.262 36.119 0.227 0.018 0.047
Taiwan 1985.01 24.350 11.702 50.171 26.522 15.751 45.967 0.081 0.074 0.099
Thailand 1976.01 10.694 4.399 35.822 12.083 6.867 32.585 0.113 0.138 0.087
Turkey 1987.01 40.808 20.525 62.111 41.579 21.142 62.277 0.117 0.0121 -0.109
Venezuela 1985.01 17.608 5.797 49.617 21.245 9.845 48.658 0.007 0.156 -0.046
IFC-Composite 1985.01 10.083 5.902 29.080 12.526 9.951 22.662 0.181 0.115 0.035
The table provides descriptive statistics for global stock market returns. Total return indices on a monthly basis are provided by Morgan Stanley
Capital International (MSC) and the International Finance Corporation (IFC). I~
....,
244 Chapter 6
Recall from section 3.2, however, that there are good reasons for the existence
of currency risk premiums. Models of international portfolio choice such as Sol-
nik (1974a) and AdlerlDumas (1983) assert that investors hold foreign curren-
cies in their portfolios for both hedging purposes and for speculative reasons.
DrummenlZimmermann (1992) show that a considerable fraction of currency
risk is systematic. Empirical evidence for time variation in currency risk premi-
ums is provided in Dumas/Solnik (1995) and De Santis/Gerard (1998). At least
in theory, therefore, optimal hedge ratios should be below unity. However, em-
pirical evidence on the benefits of currency hedging is ambiguous. Glen/lorion
(1993) find that hedged strategies outperform unhedged ones, while Levy/Lim
(1994) find the opposite result. The analysis of my data set in table 6.1 reveals
that in only four cases (Canada, Hong Kong, Singapore, and Spain) hedging
leads to an increase in the average returns. In all cases, return volatility is suc-
cessfully reduced compared to the unhedged returns, which is consistent with
recent results by Solnik (1998).
53 Missing Eurocurrency rates for some countries during the early sample period were
substituted by locally available money market instruments.
54 See PeroldlSchulmann (1988).
Volatility bounds for stochastic discount factors 245
'Europe l' includes the large European equity markets, 'Europe 2' the smaller
European stock markets in the MSCI set, and 'Europe 3' the European stock
markets contained in the IFC database. The remaining MSCI markets are classi-
fied as 'North America' and 'Pacific Basin'. The grouping of IFC constituents is
similar. In addition to 'Europe 3', there are regional subsets denoted as 'Asia'
and 'Latin America'. A full set of data for the emerging equity markets is only
available starting at 1985.01. Investigating structural breaks is particularly inter-
esting for emerging markets. Taking data for all IFC countries from as late as
1985.01 onward, and dividing the series into two periods of equal length, leads
to a proliferation of orthogonality conditions. Hence, there is a danger of unreli-
able results. 55 To preserve the power of the GMM test, 'IFC1' includes only the
eight emerging stock markets with data available from as early as 1976.01, and
'IFC2' contains the two composite IFC indices for Asia and Latin America, both
available starting 1985.01. Admittedly, this approach is somewhat pragmatic,
but it helps to reduce the number of orthogonality conditions. To explore
whether any changes in the benefits of diversification have occurred over time,
both 'IFC1' and 'IFC2' are divided into two subperiods of approximately equal
length.
Remember, the system of equations in (6.29) tests the null hypothesis that the
two Swiss assets suffice to span the volatility bounds associated with different
international portfolio strategies. The statistical significance is measured via the
traditional qui-square test for the goodness-of-fit. However, this is only hard to
interpret economically. De Santis (1995) and BekaertlUrias (1996) propose a
simple method to assess the economic significance of this change. In particular,
they suggest to measure the distance between both HJ volatility bounds at the
value of E (mI+ 1 ) which corresponds to the minimum of the bound for the Swiss
spanning assets. A look back to equation (6.17) reveals that this measure is
equal to the change in the Sharpe ratio, divided by the risk-free rate of return
that is implied by the choice of E(mI+ 1 ) = Rf".l'+l' The mean discount factor at the
minimum point is very close to one and, therefore, the change in the volatility
bounds at this point can be approximately interpreted as an increase in expected
returns per unit of risk attainable through international diversification. In a nut-
shell, this measure allows an assessment of the expanded risk-return spectrum
that is available for global investors. The intuition of this measure is empirically
supported by the following observation. The minimum of the Swiss bound cor-
responds to a monthly risk-free rate of 0.39 percent, or 4.69 percent per year.
This value is close to the observed sample mean of the one-month Eurocurrency
interest rate denominated in Swiss francs, which is 0.36 percent on a monthly
basis, or 4.43 percent per year. Note, however, that this interpretation only ap-
plies in the unconditional framework. In the conditional case, spanning tests are
expressed both in terms of payoffs and returns instead of returns only. Although
the magnitude of the shift is reported for all conditional versions of the model,
these numbers should not be given too much importance.
To investigate the role of hedging, two versions of the model are estimated.
First, it is assumed that investors do not hedge and are fully exposed to currency
risk. Second, investors choose to fully hedge their currency exposure using cur-
rency forward contracts. Results of international portfolio theory indicate that
full hedging can be optimal only under four stringent conditions. 56 First, infla-
tion must be non-stochastic. Second, apart from inflation there are no other state
variables that influence prices and that are correlated with the exchange rate.
56 See Adler/Granito (1991), lorion (1994), Solnik (1998). and section 3.2.
Volatility bounds for stochastic discount factors 247
Third, the forward risk-premium must be zero at all times, i.e. purchasing power
parity holds. And fourth, domestic and foreign stock and bond prices must be
uncorrelated with the exchange rate. All of these conditions are violated empiri-
cally, at least in the short run. Intuitively, full hedging implies that investors ig-
nore the speculative character of currencies. They constitute an independent as-
set class with distinctive risk-return characteristics. In other words, full hedging
does not optimally exploit the correlation structure between stocks and forward
contracts. Therefore, if full hedging offers additional diversification benefits for
a Swiss investor, the results rather underestimate the true importance of currency
risk management.
more volatile markets. While large economically, the point estimates are not
statistically significant when using unhedged returns. As expected, the change in
the Sharpe ratio is inversely related to the p-values associated with the mean-
variance spanning tests. In general, the shifts are lost in sampling error. The
smallest p-value is 0.194 for 'Europe 2'. In a nutshell, the null hypothesis that
Swiss assets suffice to span the global HJ volatility bound cannot be rejected.
International stock returns are too noisy to detect statistically significant gains
from diversification.
Intuitively, rejecting the null hypothesis requires either higher average returns or
lower correlations. One way to reduce the volatilities is to hedge the currency
component of international returns (see table 6.1). In fact, the entries in the right
panel of table 6.2 indicate that hedging helps to further improve the performance
of an international portfolio strategy and to bring some of the p-values down to
statistically significant levels. First, full hedging increases the gains from global
diversification (,Global') by an additional (0.153 - 0.135)xO.0487 = 0.09 per-
cent. This is not obvious at all, because investors do not choose an optimal
amount of forward contracts in this case. Second, full hedging sufficiently re-
duces the noise in returns. The shifts in the bounds are now statistically signifi-
cant for the 'Global' and 'Europe plus North America' markets, at least at the 10
percent level. Again, the results are mainly driven by the smaller European
countries in 'Europe 2'.
the volatility of a common global stochastic discount factor become a lot tighter
once currency risk is taken into account.
Table 6.2
Unconditional investment strategies for MSCI stock markets
System of non-linear equations for GMM estimation:
Investment strategy
Results are shown in table 6.3. As expected, the point estimates for the condi-
tional HJ regions are higher than the corresponding unconditional ones. More
important, the p-values drop dramatically for all sets of test assets, except
'Europe 1 plus United States'. This is true for the hedged as well as the un-
hedged portfolio positions. The shifts become significant at the 1 percent level
for the sets 'Europe 2', 'Europe', and 'Pacific Basin'. The sharp drop of the p-
value for unhedged returns in 'Europe I plus United States' and its immediate
increase when hedged returns are used instead is only hard to interpret. How-
ever, the p-values do not indicate any statistical significance.
These results have three major implications. First, the dramatic improvement in
the HJ volatility bound for scaled returns provides another strong evidence for
the predictability of stock returns. Hence, the results can be interpreted as sup-
port for the simple return regressions in chapter 4 from a more elaborate angle.
Second, the results imply that tactical asset allocation with periodic adjustments
of portfolios based on new information helps to exploit the correlation structure
between international stock returns. The risk-return spectrum a global investor
faces expands both economically and statistically. In light of the recent findings
in BekaertlLiu (1999) and FersonlSiegel (1999) that optimal scaling functions
are nonlinear, the simple linear scaling rule might even underestimate the true
benefits of incorporating instrument variables. Finally, the decision of whether
Volatility bounds for stochastic discount factors 251
to hedge or not becomes an issue of minor importance for a Swiss investor once
conditioning information is taken into account.
Table 6.3
Conditional investment strategies for MSCI stock markets
System of non-linear equations for GMM estimation:
Investment strategy
The system of orthogonality conditions in (6.29) is tested by GMM for a fixed set of Swiss spanning
assets and varying sets of unhedged and hedged test assets. The spanning assets are the MSCI stock
market index for Switzerland and the 9O-days Eurocurrency interest rate, both denominated in Swiss
francs. The different sets of test assets contain international stock market indices and are defined as
follows: 'Europe I' (Germany, France, United Kingdom), 'Europe 2' (Austria, Belgium, Denmark,
Italy, Netherlands, Norway, Spain, Sweden), 'North America' (United States, Canada), and 'Pacific
Basin' (Australia, Japan, Hong Kong, Singapore). 'Europe' consists of Europe I and Europe 2. The
global information variables used to scale returns are the world dividend yield and a G7 term spread of
interest rates. ***1**1* denotes statistical significance at the I %15%110% level, df the degrees of free-
dom.
under this category. The attraction emerging markets have received can be ex-
plained by their distinct risk-return profiles, coupled with their low correlations
with other stock markets. DivechaIDrachlStefek (1992) and Speidel/Sappenfiel
(1992) even argue that a 'diversification free lunch' awaits investors in emerg-
ing stock markets. Empirically, it is true that IFC markets exhibited higher aver-
age returns than developed markets in the past. The average correlations be-
tween IFC and MSCI stock markets have also been lower than those between
the MCSI markets themselves. Clearly, this should make it easier to reject the
null hypothesis of mean-variance spanning.
Table 6.4 reports the results for mean-variance spanning tests using the returns
(in Swiss francs) on various subsets of IFC stock market indices. Panel A dis-
plays the results for the period from 1985.01 to 1998.08. For 'Asia' (India, Tai-
wan, Thailand, Malaysia, Korea, Pakistan, and the Philippines) the null hypothe-
sis of spanning can be rejected at the 5 percent level of significance. For a Swiss
investor portfolio diversification over these seven markets led to a monthly gain
of 0.12xO.0487 =0.58 percent. This is huge, recalling that global diversifica-
tion over all 18 stock markets in the MSCI database led to an only slightly
higher monthly gain of 0.64 percent. The results are different for 'Latin Amer-
ica'. This set contains some of the more established markets in the IFC database
(such as Argentina, Brazil, Chile, and Mexico). Although the monthly risk-
adjusted gain remains economically significant (0.40 percent monthly), with a p-
value of 0.173 the shift is only moderate from a statistical point of view. Unfor-
tunately, the results for 'Europe 3' are only hard to interpret. The shift of the
bound is neither economically nor statistically significant for this subset of IFC
markets.
The results could be driven by the most recent years, when correlations have in-
creased notably. To explore this possibility, subset 'IFC l' (Argentina, Brazil,
Chile, Greece, Mexico, Thailand, India, Korea) and 'IFC 2' (Asia Composite
and Latin America Composite) are each divided into two periods of equal
length. Estimation results are reported in panel B of tables 6.4. Indeed, for the
more advanced emerging markets in 'IFC l' the gains from diversification dur-
ing the early period from 1976.01 to 1986.12 were huge. A Swiss investor could
have gained 0.260xO.0487 = l.27 percent (!) on a monthly basis, and this shift
Volatility bounds for stochastic discount factors 253
was significant at the 1 percent level. It must have been this impressive increase
in risk-adjusted returns that established the myth of a 'diversification free-lunch'
in emerging stock markets. Unfortunately, these times are over. A look at the
results for the more recent period from 1987.01 to 1998.08 shows that the mag-
nitude of the shift in the bound has dropped significantly. The HI bound still be-
comes tighter, but monthly gains from diversification come back to a level com-
parable to that for the smaller European markets (a monthly 0.46 percent) and
are no longer significant at the 10 percent level. Results are similar for 'IFC 2',
the set containing the composite indices for Asia and Latin America. Over the
early period from 1985.01 to 1991.06 the gains of 0.54 percent per month are
significant at the 5 percent level. In the most recent period from 1991.07 to
1998.08, however, the magnitude of the shift drops to a monthly 0.41 percent
and is lost in sampling error. In a recent paper, Errunza/Hoan/Hung (999) fmd
similar results. The results are also consistent with Goetzman/lorion (1999b).
Their simulation evidence shows that high returns and low covariances are tem-
porary phenomena that can be attributed to the recent emergence of Asian and
Latin American stock markets. In particular, the history of emerging stock mar-
kets immediately after re-emergence provides an overly optimistic picture of
future investment performance.
Finally, the colunms on the right side of table 6.4 contain the results for the con-
ditional model. The point estimates rise to almost unbelievable levels, but again,
one should be careful with interpretations. Nevertheless, the results can be inter-
preted as evidence that IFC returns are better predictable than MSCI returns. 57
Admittedly, given that the conditional version of the model must be understood
as an actively managed portfolio strategy, the gains from diversification might
only be hard to realize. The strategy requires periodic adjustments of portfolios
based on newly available information, which might be difficult due to thin trad-
ing and a variety of direct and indirect barriers for foreign investors. Finally,
from an asset pricing perspective, the findings in table 6.4 indicate that emerging
stock markets have become more and more integrated with the world stock mar-
ket.
Table 6.4
Investment strategies for IFC stock markets
System of non-linear equations for GMM estimation:
Unconditional Conditional
Panel A:
24.267** 109.042***
Asia 14 0.120 42 0.450
(0.043) (0.000)
16.422* 101.765***
Latin America 12 0.082 36 0.231
(0.173) (0.000)
8.272 30.493**
Europe 3 6 0.012 18 0.127
(0.219) (0.033)
Panel B:
IFC 1 55.916*** 77.776***
16 0.260 48 0.342
( 1976.01-1986.12) (0.000) (0.004)
IFC I 23.040* 85.415***
16 0.094 46 0.734
(1987.01-1998.07) (0.113) (0.000)
IFC2 15.617** 30.773***
4 0.111 12 0.297
( 1985.01-1991.06) (0.029) (0.002)
IFC2 5.439 30.119***
4 0.084 12 0.345
(1991.07.1998.07) (0.245) (0.003)
The system of orthogonality conditions in (6.29) is tested by GMM for a fixed set of Swiss spanning
assets and varying sets of unhedged and hedged test assets. The spanning assets are the MSCI stock
market index for Switzerland and the 9O-days Eurocurrency interest rate denominated in Swiss francs.
The different sets of test assets contain international stock market indices and are defined as follows:
'Asia' (India, Taiwan, Thailand, Malaysia, Korea, Pakistan, Philippines), 'Latin America' (Argentina,
Brazil, Chile, Columbia, Mexico, Venezuela), and 'Europe 3' (Greece, Portugal, Turkey). In addition,
'IFC I' contains the indices for Argentina, Brazil, Chile, Greece, Mexico, Thailand, India, Korea (all
starting 1976.01), while 'IFC 2' consists of the composite indices for Asia and Latin America (both
available starting 1985.01). The global information variables used to scale returns are the world divi-
dend yield and a G7 term spread of interest rates. ***1**1* denotes statistical significance at the
I %/5%/10% level, dfthe degrees of freedom.
Volatility bounds for stochastic discount factors 255
The aims of this chapter have been manifold. First, given the negative evidence
for the consumption-based model in the previous chapter I try to analyze the
reasons for its failure. Second, I ask whether it is harder to price international
assets rather than domestic ones. This is an unparametric test of market integra-
tion. Third, I shed some light on the benefits of international diversification from
a Swiss perspective for both developed and emerging stock markets. Fourth, I
reexamine the predictability of stock returns on the basis of well-known instru-
ment variables. The results allow to assess the plausibility of the evidence pre-
sented in chapter 4. Finally, I explore the importance of currency hedging for the
risk-return menu faced by a globally oriented investor. All these issues can be
analyzed in a common framework, the HansenJJagannathan (1991) volatility
bounds for admissible stochastic discount factors. The results in this chapter can
be summarized as follows:
Mean-variance regions for stochastic discount factors: The main idea of Han-
senJJagannathan (1991) is to move from the familiar mean-standard deviation
diagram for portfolio returns to a mean-standard deviation representation for
admissible stochastic discount factors. In particular, given any set of asset re-
turns, one can calculate a feasible region in mean-standard deviation space for
valid stochastic discount factors to fall into.
• For any given mean stochastic discount factor, a higher Sharpe ratio delivers
a tighter bound on the volatility of valid stochastic discount factors. In other
words, a higher Sharpe ratio must be accompanied by a more volatile sto-
chastic discount factor. Unfortunately, consumption growth has been very
smooth in the past. This requires an unreasonably high coefficient of relative
256 Chapter 6
risk aversion for the canonical consumption-based model with time separable
risk aversion to generate enough volatility in the pricing kernel. In a nutshell,
the HJ framework allows a reinterpretation of the MehralPrescott (1985) eq-
uity premium puzzle in a more modern stochastic discount factor language.
International assets are harder to price than domestic ones: International finan-
cial stock market data is used to establish which assets are the most challenging
for standard asset pricing models and to measure the benefits of global diversifi-
cation. The HJ framework is extremely appropriate for this purpose, because it
incorporates asset pricing implications and assumes as little structure as possi-
ble.
• Spanning tests in the spirit of De Santis (1995) and BekaertlUrias (1996) are
performed to examine whether a Swiss investor can mimic foreign returns by
holding only domestically traded assets. In particular, benchmark bounds are
constructed using the return on the Swiss stock market index and the 90-days
Eurocurrency interest rate for Swiss francs as the spanning assets. New
bounds are then computed by adding different combinations of foreign stock
indices as test assets. The significance of the resulting shift is evaluated both
statistically (via the GMM qui-square test statistic) and economically (by cal-
culating the approximate increase in the Sharpe ratio).
quired for valid stochastic discount factors when going from purely Swiss
portfolio holdings to a portfolio that is diversified across varying sets of
global stocks.
• The estimated volatility bounds can be translated into expected gains from
diversification at the level of risk of the Swiss stock market. To give a notion
of the actual magnitude of the improvement, I find that the annual expected
gain was roughly 6 percent per year for a portfolio of all European and North
American stock markets in the MSCI dataset over the sample period. Full
hedging led to an additional I percent increase.
• Observing that the shifts of the volatility bound are often lost in sampling
error (at least for the MSCI markets) provides additional evidence for global
stock market integration. This insignificance can also be interpreted as a pos-
sible explanation for the home-bias puzzle. Most interesting, emerging stock
markets become increasingly integrated from a Swiss investor's point of
view. This interpretation is to be taken with due care, because my test design
does not rely on any specific form of an asset pricing model. Yet, the times-
if they ever existed - when 'diversification free lunches' were readily avail-
able in these markets seem to be over.
• Using a linear rule to scale returns with the world dividend yield and a G7
term spread, the significant shifts (both economically and statistically) in the
258 Chapter 6
HJ volatility bounds can be interpreted as yet another evidence for the pre-
dictability of stock returns. The results confinn earlier findings in chapter 4.
• The dominant strategy for a Swiss investor is to actively manage his or her
portfolio on the basis of newly available infonnation. Although this strategy
involves dynamic adjustment of the portfolio across different geographical
regions and, hence is harder to implement on a day-to-day basis, it allows the
investor to further increase the Sharpe ratio. In other words, using condi-
tioning infonnation makes it easier to actually capture a good part of the po-
tential benefits from international diversification.
259
Chapter 7
Recursive utility functions (Epstein/Zin (1989) and Wei I (1990)) and habit formation
(Constantinides (1990), Abel (1990), and Campbell/Cochrane (1999a)) are examples in
this direction.
2 See Cochrane (1996).
Mean reversion and rational asset pricing 261
t -1 will help to predict the change in consumption in t since part of that change
is the lagged response to the previous change in permanent income.
where 1;,'+1 denotes the continuously compounded rate of return on stock i over
the period from t to t + 1, and ~C'+I is the logarithmic growth rate of consump-
tion over the same period of time, i.e. ~C'+I = In (CI+ IjC, ). The intercept term is
defined as ~,==logP+<V2)(cr~ +icr~ +2)'O',c).5 According to equation (7.1),
changes in the anticipated growth rates of consumption should be reflected in
time variation of expected real stock returns. Divergent behavior of the two
variables' expectations indicates a failure of the model's restrictions. In general
the error term '11",+1 will be correlated with realized consumption. It is well
known that in the presence of endogenous regressors ordinary least square is not
an appropriate estimation method. However, '11,,1+1 is uncorrelated with any vari-
ables in the information set at time t. Consequently, the econometrician has to
identify proper instrument variables. Any lagged variables that are correlated
with asset returns can be used as instruments in instrumental variable (IV) re-
gressions. 6 For example, Campbell (1999) uses lagged asset returns, lagged con-
sumption growth rates, and the log price dividend ratio as instruments. His re-
sults are not particularly encouraging for the consumption-based model.
The framework set out in (7.1) can be extended in two respects. First, the results
of simple return regressions in chapter 4 demonstrate that global instrument
variables contain information about expected stock returns. Therefore, the same
set of global instrument variables should also contain prediction power for fu-
4 For example, see Hansen/Singleton (1983), Hall (1988), CampbelllMankiw (1990), and
Campbell (1999).
5 See section 2.5 for the exact notation. See also RestoylWeil (1998).
6 See chapter 16 in Greene (2000) for a description of instrumental variable estimation.
264 Chapter 7
ture growth rates of consumption.7 Second, the relevant states of the economy
for determining conditional moments of asset returns are not only characterized
by mean consumption growth, but also the volatility of consumption rates. This
can easily be seen by examining expression (2.46):
where r.,'+1 now denotes the excess return on asset i over the period from t to
t+ 1. 8 Recall from chapter 4 that the global instrument variables give strong sig-
nals about mean stock returns. In contrast, they do not give much information
about the standard deviation of returns, as indicated by low regression R-squares
for most stock markets. Kandel/Stambaugh (1990) and Cochrane (1997) provide
evidence that the standard deviation of returns slightly increases when declining
price dividend ratios indicate increasing expected stock returns. The increase in
standard deviation is much less than the increase in mean return, however, im-
plying that the Sharpe ratio varies over time. Specifically, it increases when
prices are low. Equation (7.2) reveals that this is possible if (i) the conditional
correlation of consumption growth and returns varies a great deal, and/or (ii)
there were times of high and low volatility of consumption growth, i.e. variation
in O'c, •. The first possibility is unlikely.9 Variation in conditional consumption
variances, on the other hand, has largely been neglected in the previous litera-
ture. To fill this gap, the main model of this chapter assumes constant correla-
tion between consumption growth and stock returns, but incorporates heterosce-
dasticity in consumption rates.
In a first ad-hoc approach conditional means and standard deviations of the loga-
rithmic rates of G7 consumption are estimated as linear functions of four global
instrument variables: the world dividend yield (iWLDY), the G7 term spread
(iTSG7), the TED spread (iTED), and the G7 short-term interest rate (iSRG7).IO
(7.3)
This specification does not contradict Hall's (1978) argument for a random walk
of marginal utility. Recall, he finds empirical rejection for his model using
lagged stock prices. Given the amount of correlation between stock returns and
lagged global instrument variables, evidence for predictable components in con-
sumption growth rates should come as no surprise. Of course, the linear relation
in (7.3) is only an approximation to what would almost surely be a nonlinear
relation between the expected consumption growth and the predetermined finan-
cial variables.
(7.4)
Rejecting the null hypothesis that all sensitivity coefficients in (7.4) are jointly
equal to zero is interpreted as evidence for heteroscedasticity in consumption
growth. The system of equations in (7.3) and (7.4) can be estimated jointly using
the Generalized Method of Moments (GMM). However, since the same predic-
tive variables are used in both equations, this is computationally equivalent to a
two-stage procedure. Ordinary least square (OLS) estimates of (7.3) are obtained
in the first stage, and the estimated residuals Tt'+l from this regression are then
used in the second stage estimation of (7.4). Series for the conditional mean and
the conditional standard deviation of consumption growth are obtained as the
fitted values using the parameter estimates at the beginning of each quarter.
Table 7.1
Conditional means and standard deviations of G7 consumption growth
System of equations to be tested using ordinary least square:
Estimated coefficients
The set of quarterly consumption data from chapter 5 is used to estimate the
model. Results are shown in table 7.1. The standard errors used to compute t-
values for the significance of the sensitivity coefficients are based on the esti-
mator of the covariance matrix proposed by NeweylWest (1987), taking care of
first-order autocorrelation in residuals to avoid the aggregation bias. 12 The null
hypothesis that the conditional mean and the conditional standard deviation of
consumption growth are independent from the global set of instruments can be
tested using standard Wald-test methodology. The chi-square test statistics, to-
gether with the associated p-values (in italics), are reported in the last column of
table 7.1. Many sensitivity coefficients are significant. All intercepts are signifi-
cant as well, indicating that the model does not capture all time variation. The
null hypothesis that the sensitivity coefficients on the instrument variables are
jointly equal to zero can be rejected for all mean consumption rate specifications
at the I percent level of significance. The regression results are less pronounced
for the volatility specifications, but there is some evidence for heteroscedasticity
in consumption growth (except for services and durable goods). Finally, the ad-
justed R-squares are remarkably high, especially for the conditional mean speci-
fications.
Figure 7.1 plots the time evolution of conditional means and conditional stan-
dard deviations of quarterly G7 consumption rates, using nondurable goods plus
services. The predictions are calculated as the fitted values of the regressions in
(7.3) and (7.4), using the least square coefficient estimates. Vertical bars indi-
cate the turning points of the business cycle as determined by the National Bu-
reau of Economic Research (NBER). Fortunately, the business cycle patterns of
consumption growth are as one would expect intuitively. Troughs in the condi-
tional mean rates tend to occur during recessions. At the same time, economic
slowdowns are characterized by peaks in the conditional standard deviation of
consumption growth. 13 In other words, the conditional mean of consumption
growth varies procyclical, while the conditional volatility exhibits countercycli-
cal movements. Overall, there is strong evidence that both conditional moments
vary over time. What is even more important is that these movements are related
to the states of real activity. Although the R-squares are relatively low, a modest
degree of predictability in the endowment process is sufficient to produce mean
reversion in stock returns.
Figure 7.1
Predicted means and standard deviations of G7 consumption rates
Period: 1973.01 - 1998.03 (quarterly data)
0.00
0.00
0.00
-0.00 "'-h-rtn-nnTTTlTTTlTTTTTTTTnTrmrrTTTirmn-mmnTTTlTTTlTTTlTTTl-rn'rnTrTTlTrTTTrTTTn-m:TTT1-rm-l
74 76 78 80 82 84 86 88 90 92 94 96 98
The figure exhibits the time evolution of the conditional mean and the conditional standard deviation
of aggregate G7 consumption rates, using nondurable goods plus services. Conditional moments are
modeled as being linearly dependent on a set of prespecified global instrument variables (the world
dividend yield, a G7 term spread, the TED spread, and a G7 short-term interest rate). The forecasts are
computed as the fitted values from (7.3) and (7.4), using the least square coefficient estimates. The
vertical bars denote the business cycle peaks and troughs as defined by the National Bureau of Eco-
nomic research (NBER).
The general hypothesis of this chapter is that the conditional mean and the con-
ditional volatility of consumption growth characterize the states of the economy
relevant for determining conditional moments of asset returns. In a first attempt
Mean reversion and rational asset pricing 269
to examine this proposition the following linear model for excess stock returns
in country i is considered:
(7.5)
where !le., denotes the estimated mean consumption growth rate in quarter t+ 1,
calculated as the fitted value of equation (7.3), and cr c.' is the estimated standard
deviation of quarter t+ 1 consumption growth, calculated as the fitted value of
equation (7.4).14 The set of quarterly stock returns on the MSCI indices from
chapter 4 is used again. The sample period is again from 1973.01 to 1998.03. All
returns are in Swiss francs and the I-month Eurocurrency rate for Swiss francs
is used to compute continuously compounded excess returns, denoted as ['.'+1 for
each country i. Note that estimating (7.5) corresponds to the second stage of an
instrumental variable (IV) regression. Unfortunately, ordinary least square pro-
duces frequent negative expected excess returns. Although such a situation is
possible from a theoretical standpoint, it is not a very plausible one. IS I follow
the approach by De Santis/Gerard (1998) and rewrite the model as follows:
(7.6)
This brute-force approach guarantees that the market risk premium takes on only
positive values. The model in (7.6) is estimated using non-linear least square. A
Wald-test rejects the null hypothesis that quarterly returns on the world stock
index do not depend on the fitted moments of quarterly consumption at the 10
percent level of significance. The results are similar for the larger stock markets
in the sample, but the null hypothesis of zero coefficients on the conditional
mean and the conditional volatility of consumption growth cannot be rejected
for several of the smaller stock markets (e.g., Austria with a p-value of 0.93).
Figure 7.2 shows the time evolution of expected excess returns on selected stock
markets and the world market index, calculated as the fitted values from (7.6).
Again, vertical bars denote the official business cycle dates as determined by the
National Bureau of Economic Research (NBER). The picture is striking: ex-
pected excess stock returns depend on the conditional moments of consumption
growth and exhibit a clear business cycle association. In particular, peaks in the
market risk premium occur closely before business cycle troughs, which is con-
sistent with the consumption smoothing story proposed throughout this book.
The average world market risk premium is 6.5 percent per annum. Given the 6
percent equity premium in long term U.S. data, this seems to be a plausible
value. 16 While this very ad-hoc model does an excellent job in replicating the
level of the global market risk premium, it clearly cannot solve the equity pre-
mium puzzle from a more elaborate point of view.
Figure 7.2
Market risk premiums on global stock markets
Period: 1973.01- 1998.03 (quarterly data)
r--.-----.-r--.--.-------r-r-------,
0.25 ....
020
0.15
~
'I
'I
"
"
" :1
,,
0.10
,,
0.05
74 76 78 80 82 84 86 88 90 92 94 96 98
The figure shows the time evolution of the stock market risk premiums for Gennany, Switzerland, the
United States, and the world market index according to (7.6). Expected excess returns are modeled as
linear functions of the fitted values of the conditional mean and the conditional volatility of G7 con-
sumption rates (nondurable goods plus services). The vertical bars denote the official business cycle
turning points as detennined by the National Bureau of Economic Research (NBER).
The important point is that the results provide convincing evidence for the pre-
sumed interplay between consumption smoothing and risk correction. Recall,
the argument runs as follows. A level of current income below trend is consid-
ered as transitory due to trend reversion. The life cycle-permanent income hy-
pothesis implies a significantly lower level of aggregate savings because inves-
tors attempt to transfer future purchasing power into the present by selling off
some of their financial assets. However, all stocks must be held in equilibrium,
implying that expected returns must be high to compensate for the risk of a more
volatile consumption pattern, that is, recession risk.
Given the evidence for the ad-hoc model in (7.3) and (7.4), the remainder of this
chapter presents a clean economic model for the time variation of expected stock
returns. The argument remains basically the same and rests on two ingredients:
(i) trend reversion in the endowment process and (ii) the life cycle-permanent
income hypothesis of ModiglianifBrumberg (1954) and Friedman (1957). Pre-
dictability of asset returns is a natural consequence of consumption smoothing
and risk correction. Given the empirical results for the conditional version of a
consumption-based model in chapter 5, the model presented here is less ambi-
tious. No attempt is undertaken to explain or even replicate the level of the his-
torical equity premium. Nevertheless, if there is some truth in the consumption
smoothing story, the standard model should alleast be able to capture the nega-
tive autocorrelations in long-horizon stock returns. Specifically, assuming that
there is a structural relationship between consumption and stock returns, mean
reversion and return predictability should be consistent with a variant of the
classical Lucas (1978) model of an endowment economy.
(7.7)
272 Chapter 7
If markets are complete, marginal rates of substitution are equated across all in-
vestors and the stochastic discount factor mt+1 is uniquely specified. While dif-
ferent models of capital market equilibrium rely on specific structures of the dis-
count rate, the Euler-equation in (7.7) establishes a necessary condition for mar-
kets to be intertemporally efficient. In particular, in an efficient market the sto-
chastic process defined by the product of the total return on an asset and the
pricing kernel must be unpredictable using any information available at time t.
However, this does not require that the returns themselves are unpredictable. In
particular, early work by LeRoy (1973) and Lucas (1978) shows that rational
expectation equilibrium prices need not follow a random walk, and not even a
martingale sequence. Depending on the covariances between the stochastic dis-
count factor and scaled stock returns, Kirby (1998) derives values for the inter-
cept, the slope coefficients, and the regression R-squares in instrumental vari-
able prediction models consistent with rational asset pricing. 17 In general, the
only circumstance in which the Euler-equation rules out predictability is when
the stochastic discount factor mt+1 is equal to a constant. In this case risk-neutral
pricing prevails. While risk neutrality is sufficient to rule out predictability, risk
aversion does not necessarily imply predictability. For example, the lognormal
model of Hansen/Singleton (1983) implies that if agents have constant relative
risk aversion and consumption growth is unpredictable, asset returns will be un-
predictable as well.
The essence of predictability is that there is an immediate link between the sto-
chastic properties of the pricing kernel and the underlying capital market equi-
librium. If both the stochastic discount factor and the risk-free rate of return are
i.i.d., the Euler-equation implies that prices follow a random walk in the spirit of
Samuelson (1965). This is also Fama's (1970) dictum of informationally effi-
cient markets. If, however, there is cyclical behavior in the stochastic process of
the discount factor, the old version of market efficiency no longer captures the
equilibrium price process. Instead of a random walk process, negative serial cor-
relation of stock returns can be expected at long horizons. However, the eco-
nomic rationale for mean reversion is not unique. LakonishoklSchleitberNishny
(1994) argue that high (low) firm fundamentals relative to stock prices announce
high (low) returns because stock prices are irrationally low (high) compared to
their fundamental value. Specifically, investors extrapolate past growth rates too
far into the future. If the expected growth rate does not materialize, growth
(value) stocks with low (high) dividend price ratios command low (high) re-
turns. On the other hand, FamalFrench (1988a,b) suggest that if rational equilib-
rium expected returns are correlated but mean reverting, and if shocks to ex-
pected returns are independent of shocks to rational forecasts of dividends, a
shock to expected returns has no long term effect on expected prices. However,
it must be compensated by an opposite movement in the current price. Mean re-
version in prices is then implied by the mean reversion in expected returns.
Clearly, only an equilibrium model for the time evolution of expected returns
allows to distinguish between these two theories.
( c -X )=(C,-x,t - l
Tl
(7.8)
u " 1-11'
where X, denotes the level of habit that slowly adjusts to consumption. The
symbol 11 is used for the power, because curvature and risk aversion are no
longer equal. Time-nonseparability makes marginal utility volatile even when
consumption itself is smooth, because consumers derive utility from consump-
tion relative to its recent history and not from the absolute level of consumption.
Risk aversion, denoted as "(, depends on how far consumption is above habit:
-Cu C
"(, = __,_c_c =11C -'X, (7.9)
UC t I
18 Another important issue is whether the habit level is modeled as a function of an agent's
own consumption (e.g., Constantinides (1990) and BoldrinlChristianolFisher (1995» or
aggregate consumption (e.g., Abel's (1990, 1994) 'catching up with the Joneses'). Fi-
nally, the speed with which habit reacts to individual or aggregate consumption must be
specified. DunnlSingleton (1986) and FersonlConstantinides (1991) model habit as de-
pendent on one lag of consumption only, while BoldrinlChristianolFisher (1995) let the
habit level react only gradually to changes in consumption.
Mean reversion and rational asset pricing 275
E, ( C..t+I) _ y 0 p (7.10)
cr - t c,t le,t'
I.'
where [,.t+1 again denotes the excess return on stock i over the risk-free rate of
interest. To get a constant risk-free rate of interest (as observed in the data),
Campbell/Cochrane (1999a) use a back-of-the-envelope argument to determine
how fast habit adjusts to contemporaneous consumption. Specifically, they cali-
brate the model such that the intertemporal substitution effect and the precau-
tionary savings effect offset each other, implying a constant riskless rate of in-
terest. 19 Given all these assumptions, they are able to replicate a slowly time
varying, countercyclical market risk premium. As consumption rises in a boom
and falls toward habit in a recession, the Sharpe ratio in (7.10) moves counter-
cyclical (that is, high in recessions and low in booms). Note that time variation
in consumption volatility 0c.t is not required to explain mean reversion.
To sum up, habit formation implies that the investor slowly adjusts to a new
level of consumption, creating mean reversion in marginal utility. Consumption
itself follows an i.i.d. lognormal process, i.e. habit models do not require that
there are predictable components in consumption. Unfortunately, the model by
Campbell/Cochrane (1999a) cannot duplicate a high equity premium with low
risk aversion. It only disentangles risk aversion and intertemporal substitution so
that high risk aversion can be consistent with low and constant interest rates and
smooth consumption. The model also explains the predictability in stock returns.
(7.ll)
(7.12)
In this type of economy, the risky asset is merely a claim on the aggregate con-
sumption good. Therefore, the asset's payoff Y'+I can be replaced with con-
sumption CHI' hence:
(7.13)
20 See expression (2.2) for the specification of power utility and section 2.2.3.
21 This assumption is clearly restrictive. Abel (1994) introduces an equilibrium model
where output (consumption) is the sum of labor income and capital income (dividends).
DanthinelDonaidson (1999) also build a model where consumption is no longer con-
strained to equal dividends.
22 For u.S. evidence see table I in ChecchettilLamlMark (1990), p. 402.
Mean reversion and rational asset pricing 277
(7.14)
The realized gross one-period return of the risky asset can then be written as a
function of the price/consumption ratio:
(7.15)
Furthermore, the risk-free rate of return is given from equation (7.12) as:
1
Rc., =p-' (7.16)
f.'
The model in (7.15) and (7.16) has been rejected empirically. It is unable to si-
multaneously explain the level of the historical equity premium and the low risk-
free rate. 23 Strictly speaking, in this simple economy a risk-free rate does not
even exist. Incorporating the risk-free asset is also hard to deal with empirically.
In a framework similar to the present one Bonomo/Garcia (1996) demonstrate
that the risk-free rate implied by the simple endowment economy is excessively
volatile, predicting an almost constant risk premium. They conclude that the
main failure of the model comes from the characteristics of excess returns. 24 The
goal in this chapter is not to report yet another failure of the consumption-based
model, but to demonstrate that rational asset pricing is consistent with consump-
tion smoothing and mean reversion in expected returns. Hence, the risk-free rate
will be neglected in the analysis that follows. Instead, an empirical specification
of the consumption process is estimated, and the resulting stochastic properties
of endogenous asset prices are compared with the empirical ones.
where s, is the state of the economy at time t, p and q are the probabilities of
remaining in state 1 and 0, respectively. These probabilities are constant through
time. The state of the economy is defined as a sequence of Markov random vari-
ables {s,} that take on values of I or 0 with transition probabilities p and q. Re-
25 For example, Filardo (1994) extends the model to time varying transition probabilities,
26 Bonomo/Garcia (1996) allow for three different states in the consumption process, but
they are harder to interpret economically.
Mean reversion and rational asset pricing 279
calling that ilct+1 =In (C,+J C, ), the endowment process follows a random walk
in logarithms with state-dependent drift Ils and state-dependent volatility s . cr
The intuitive thing about this model is that the two states can be interpreted as
being directly linked to the business cycle. State I can be defined as an eco-
nomic boom, while state 0 denotes an economic recession. The main intuition is
that the statistical properties of consumption growth are assumed to differ across
these two states of the economy, that is, over the business cycle. The specifica-
tion in (7.17) is slightly more general than that by CecchettilLamlMark (1990)
in that it also allows for state-dependent variances. 27
(7.18)
where N U denotes the density function for a normal distribution. By the prop-
erties of the lognormal distribution the following expression holds:
To solve for the price/consumption ratio plug (7.19) into (7.14). For PI' the con-
sumption/price ratio in the upstate, it is then possible to write (omitting the time
subscripts):
(7.20)
The analogue expression holds for Po' the price/consumption ratio in the down-
state. Collecting both price/consumption ratios in a vector, denoted as P, equa-
tion (7.14) can further be rewritten as:
27 In a different context Whitelaw (1997) experiments with time varying transition prob-
abilities, but the parameters of this part turn out to be statistically insignificant.
280 Chapter 7
and
1t' =[ pexp ( (1-Y)111 +~(1- yf crn (1- p )exp( (1-Y)l1o + ~(1- yf cr~)].
(1-q)exp( (1- Y)111 + ~ (1- y)2 crn q exp( (1- y) 110 + ~ (1- y)2 cr~ )
(7.22)
with I and 1 denoting the identity matrix and a vector of ones, respectively.
Plugging (7.22) into (7.15), it is easy to compute gross (one-period) returns on
the risky asset. It is now possible to compare the regime switching model's im-
plications for returns with actual stock market data, i.e. to verify whether mean
reversion is rational or not.
28 This, of course, requires that the £' s are normally distributed. For detailed information
about the estimation algorithm the interested reader is referred to chapter 22 in Hamil-
ton (1994).
29 See section 3.1.1 for a description of consumption data.
Mean reversion and rational asset pricing 281
data, i.e. the entries in the table reveal marked differences in state dependent
moments across the sample of G7 countries.
30 Even though both estimates for the French growth rates in consumption are significant,
the two entries in table 7.1 show that the regime switching model cannot clearly distin-
guish between the two hypothetical states. Another uncomfortable observation is the
excessively high, but insignificant, variance in Japan in a bad state.
282 Chapter 7
duration than phases of economic downturn. More important, this result is con-
sistent with the procedure in chapter 4, where business cycle fluctuations in
macroeconomic time series were defined as cycles of six to thirty-two quarters
in length. In other words, the mean reverting components in consumption rates
relate to the business cycle. Hence, equilibrium stock returns should also be as-
sociated with the business cycle.
Table 7.2
Maximum likelihood estimates of the Markov regime switching model
Two-state regime switching model to be estimated:
p q 111 Ilo
(J2
I
(J2
0
The table shows the estimates of the transition probabilities and the state dependent moments of the
regime switching model in (7.17) for logarithmic growth rates of quarterly national consumption (non-
durable good plus services). The numbers are in percent per quarter, standard errors are reported un-
derneath the coefficient estimates (in italics). 'I' denotes an upstate. '0' a downstate.
Mean reversion and rational asset pricing 283
Even if the true values of (p, q, 111,110, of, (6) were known, the exact regime Sl
in place at every date in the sample would be unknown. However, given the es-
timates of the parameters in (7.17) it is possible to make an inference about
which regime was more likely to have been responsible for producing the date t
observation of consumption. Hamilton (1994) shows that the inference about the
value of Sl depends on all observations available up to date t. 31 Figure 7.3 pres-
ents the inferences about the unobserved regime. The line in each box shows the
filter probabilities of being in the state I (upstate) in each quarter, given the in-
formation available up to that point. For the United States the filter probabilities
closely track the business cycle as it is dated by the National Bureau of Eco-
nomic Research (NBER) and indicated by the vertical bars.
Figure 7.3
Filter probabilities for consumption growth
Period: 1973.01-1998.03 (quarterly data)
1.0 --r---::::::-"T--,r-::--::---rr....,...-r-::--:-~-----.....,.......-----,
0.8
0.6
0.4
0.2
Figure 7.3
Continued
Period: 1973.01 - 1998.03 (quarterly data)
1.0,-_ _ _ _ _ _ _ _ _ _ _---, 10,-_ _ _ _ _ _ _ _ _ _ _ _ _---,
08 08
06 0.6
04 04
0.2 02
70 75 80 85 90 95 70 75 80 85 90 95
Canada France
10~------------------------_, 1.0...,.-::>'1.."-7'<""--,...,-----....,,...----------,
08
0.6
04 0.4
02 02
70 75 80 85 90 95 70 75 80 85
Germany - Italy
1.0 r-'
"tJrv""';--;.,...........~.......--.,.---o·"...\
08 V
0.6
04
02
70 75 80 85 90 95 70 75 80 85 90 95
The figure displays the inferences about the unobserved regime s, in (7.17). The lines exhibit the time
evolution of the filter probabilities for being in state I (upstate or boom) at each point of time during
the sample period, given the information available up to that point. The associated probabilities for
being in state 2 (downstate or recession) are given by one minus the state I probabilities.
Mean reversion and rational asset pricing 285
Recession probabilities (the probabilities for state 0) are high during the 1973n4
recession, during the slumps in real activity in the early 1980s, and the short
downturn in 1990/91. At all other times during the sample period state 1 prob-
abilities (interpreted as an expansion of real activity) are clearly dominant. The
filter probabilities are very similar in Canada, Italy, and the United Kingdom,
and they are highly volatile in Germany. The Japanese case is particularly inter-
esting. During the entire sample period, except the oil shock in 1974, the upstate
probabilities are close to one. Given the dynamic history of the Japanese econ-
omy, this result is perfectly understandable. Unfortunately, the French results
are only hard to interpret. The estimates suggest that France suffered a recession
for most of the sample period. Taking closer look at table 7.2 reveals the prob-
lem. The growth rate of consumption is almost identical in the two states (0.557
percent versus 0.525 percent per quarter), and the downstate is more persistent
than the upstate (p = 0.936 versus q = 0.981). This has two obvious effects.
First, once a bad state is reached, it is extremely unlikely to leave that state. 32
Second, the switching model is unable to clearly discriminate between the two
states. Even in a 'bad' state consumption growth is high, and hardly lower that
in a 'good' state. Hence, although the model replicates the historical growth in
French consumption, it does so by misinterpreting the deeper meaning of an
economic recession.
To sum up, a two-state regime switching model characterizes the stochastic pro-
cess of logarithmic consumption rates in the G7 countries reasonably well. Both
states are estimated with statistical precision. In particular, consumption volatil-
ity is higher in recessions than in booms of real activity. This observation is im-
portant because standard results of state preference theory allow to interpret the
volatility of the discount factor as the market price of risk.33 Most important, the
stochastic properties of the pricing kernel directly determine the underlying re-
turn process. Having identified the business cycle components in consumption
growth, return simulations on the basis of (7.22) and (7.15) should produce evi-
dence against the random-walk model. Recalling the fundamental expression in
(7.2), the market price of risk is higher in recession states and lower in booming
32 A recession lasts on average 52 quarters, which is almost half of the sample period.
33 For example, see Zimmermann (1999).
286 Chapter 7
states. Hence, the long-horizon returns on the risky asset are expected to exhibit
mean reversion. LolMacKinlay (1988) emphasize that mean reversion may be
interpreted as a rejection of some economic model of efficient price formation,
but there may be other (and maybe more) plausible models that are consistent
with the empirical findings. Therefore, the next step is to examine whether mean
reversion is indeed an empirical phenomenon in historical return data, and if so,
whether the Lucas-type model asset pricing is able to reproduce these patterns.
Serial correlation in historical returns: Variance ratios have been used exten-
sively to examine mean reversion in return series. 34 The idea is to exploit the
fact that the variance of the increments of a random walk is linear in the sam-
pling interval. Assume that stock prices are in fact generated by a random walk
(possibly with drift). In this case the variance of yearly sampled returns must be
four times as large as the variance of quarterly sampled returns. More specifi-
cally, let r, be the continuously compounded one period rate of return, and rt
the k period return. Given that stock prices follow a random walk, it holds that:
(7.23)
Poterba/Summers (1988) define the variance ratio for returns at the k-th horizon
as follows:
var( rt)
VR(k)= k var ()" (7.24)
r,
(7.25)
where PJ is the j-th autocorrelation of quarterly returns. When returns are seri-
ally uncorrelated, the variance ratio is equal to one in large samples for all k.
Hence, the random-walk model is taken as the null hypothesis, and deviations of
variance ratios from the theoretical value are usually interpreted as market inef-
ficiencies 35 • In contrast, the story proposed here is that mean reversion in stock
prices is perfectly consistent with rational asset pricing. Given that consumption
growth reveals cyclical components and investors try to smooth consumption,
stock prices exhibit negative serial correlation. Hence, one would expect the
variance ratios to fall below one for a high enough value of k.
Table 7.3 shows the variance ratios estimated from historical stock returns up to
a sampling period of 10 quarters. Again, continuously compounded stock returns
from Morgan Stanley Capital International (MSCI) on a quarterly basis for the
period from 1973.01 to 1998.03 are used. The results are uniform across the G7
stock markets. Specifically, the variance ratios indicate positive autocorrelation
for k up to six quarters, but eventually fall below one in all countries, except in
Italy and Japan. To interpret this table, recall from (7.25) that variance ratios can
be expressed in terms of return's autocorrelation. For example, ratios with k = 2
are approximately equal to 1 plus the first-order autocorrelation coefficient esti-
mator of quarterly returns. Hence, the numbers in table 7.4 are a confirmation of
previous research for the U.S. applied to international stock market data. For
example, LolMacKinlay (1988) find positive autocorrelation for weekly holding
period returns. In contrast, FamalFrench (1988a) report negative serial autocor-
35 Poterba/Summers (1988) use Monte Carlo simulations to obtain standard errors for their
variance ratios under the null hypothesis of a random walk. In contrast, LolMacKinlay
(1988) develop simple, asymptotically standard normal test statistics.
288 Chapter 7
relation for long-term holding period returns. The fact that the variance ratios for
Italy and Japan are still above one in the final column of the table (for k=10)
might be due to the restriction on the length of the sampling interval. Further
increasing k might bring the ratios down below one for these countries as well.
Recent evidence for mean reversion on global stock markets has also been pro-
vided recently by BalverslWulGilliand (2000).
Table 7.3
var ( [ k)
VR{k)=--'- where r,k is the k-period continously compounded stock market return
k var{r,)
Canada 1.139 1.126 1.085 1.035 0.987 0.932 0.886 0.851 0.748
France 1.071 1.066 1.096 1.110 1.107 1.082 1.013 0.944 0.842
Germany 1.020 0.938 0.938 0.944 0.908 0.879 0.832 0.771 0.687
Italy 1.043 1.195 1.277 1.376 1.423 1.433 1.417 1.447 1.278
Japan 1.134 1.248 1.328 1.346 1.342 1.351 1.343 1.326 1.162
United Kingdom 1.064 1.018 1.003 0.921 0.868 0.843 0.849 0.884 0.839
United States 1.135 1.022 0.946 0.911 0.883 0.865 0.857 0.839 0.776
The entries in the table show the values of variance ratios up to k= I 0 quarters using quarterly stock re-
turns from the G7 countries. Roughly, a value above one indicates positive serial correlation of returns
(momentum), a value below one implies negative serial correlation (mean reversion).
The results are shown in graphical form in figure 7.4. Each box displays the
point estimates of historical variance ratios, the median, and both the 66 percent
and 90 percent confidence intervals of the Monte Carlo simulations. The hori-
zontal axis graphs the number of quarters k used to calculate returns, and the
vertical axis indicates the variance ratios. Overall, the results in figure 7.4 sup-
port the consumption smoothing story proposed throughout this book. Far from
being anomalous within a rational pricing framework, Monte Carlo inferences
suggest that mean reversion is consistent with the standard Lucas-type model in
all G7 countries. Given a coefficient of relative risk aversion of 'Y = 3, the medi-
ans of the sample distributions of the variance ratios are well below one at every
horizon. ThIs is not surprising, recalling that the model is calibrated to capture
the mean reverting components in the consumption process. Most important, the
variance ratios calculated from historical data eventually fall within the 66 per-
36 This assumes a very segmented view of the world. For an analysis of the return gener-
ating process in response to a world business cycle see DumasIHarvey/Ruiz (1999).
290 Chapter 7
cent confidence interval generated through Monte Carlo simulation. The fit of
the model is almost perfect for Canada, Germany, the United Kingdom, and the
United States. The country medians of the simulated distributions closely track
the historical variance ratios for these countries. The empirical variance ratios
eventually fall within the 66 percent confidence interval for France and Japan as
well, and the 90 percent confidence interval for Italy. The intuition is clear:
mean reversion does not produce a challenge for an equilibrium model of the
Lucas-type. Negative autocorrelation of stock returns over longer holding peri-
ods is perfectly consistent with simple rational equilibrium models of asset
pricing. This is also in line with Campbell's (1996) results. Recall his pricing
condition in expression (2.64). Mean reversion implies a positive covariance
between stock returns and the current return on wealth (VIm> 0), but a negative
covariance between stock returns and revisions in expected future returns on
wealth (Vlh > 0). Indeed, using a variance decomposition argument this is ex-
actly what Campbell (1996) finds. 37 All this leads to the conclusion that the
value for the intertemporal rate of substitution should be in the range suggested
here to explain the empirically observed mean reversion. However, as the corre-
sponding value of the relative risk aversion is not able to duplicate historical ex-
cess returns, the results impose another puzzle to equilibrium asset pricing: the
mean reversion puzzle. A level of risk aversion consistent with the equity pre-
mium would result - in addition to excessive volatility in the risk-free rate - in a
too high speed of mean reversion.
Apart from mean reversion, however, a second effect becomes apparent when
looking at the graphs for France and Italy. Even though the empirical variance
ratios do fall within the simulated confidence regions for large enough time ho-
rizons, they lie well above the 90 percent confidence interval for shorter holding
periods. This phenomenon has become known as 'momentum' in the literature. 38
Unfortunately, the simple equilibrium model is not able to entirely capture the
stock market overreaction in these two G7 countries over shorter holding peri-
ods. The debate about the mechanisms behind the momentum effect is still go-
ing on. In fact, momentum is a challenge for rational asset pricing stories. For
Figure 7.4
Variance ratios: equilibrium returns generated by concave utility
Period: 1973.01- 1998.03 (quarterly data)
Canada France
1.6...-_ _ _ _ _ _ _ _ _ _ _----. 1.4...-_ _ _ _ _ _ _ _ _ _ _ _-,
---- -------
-----
1.4
------
1.2
1.2
_--"",==--:5;:;:
~~:::.:~===:..: _____ _
~-="'~--------------------------
1.0 \
0.8
\~~~----- -----------
---
0.6
" ........
0.2 +---,--r---,-----r--r-~__,-_r___l 0.4 +---.--...--.--,---,-.......,----,-----,---1
2345678910 1 2 4 5 6 7 8 9 10
Germany Italy
1.6...-_ _ _ _ _ _ _ _ _ _ _--.
1.6.---------------.
------------- ---------
1.4 1.4
0--::':-------------------------------
:: ~~~;;;;=~-~~-------
1.0
---- 1.0
0.8 ~:::.::-.::---------
0.6 ................ --
--------------- ---- --------
0.4 +---,--r---,-----r--r-~__,-_r___l 0.4 +--,.---.,.-..---,----,--..----,----,---1
2345678910 2 4 5 6 7 8 9 10
39 Carhart (1997) reports that momentum stocks move together, as do value stocks and
small stocks. He finds that a 'momentum factor' works to 'explain' momentum portfo-
lio returns. However, it is doubtful that adding a new factor for every anomaly could be
the appropriate solution. See Danielffitman (1997) and LiewNassalou (1999).
292 Chapter 7
Figure 7.4
Continued
Period: 1973.01 - 1998.03 (quarterly data)
--
1.6...-_ _ _ _ _ _ _ _ _ _ _-, 1.6...-_ _ _ _ _ _ _ _ _ _ _---,
1.4
---- ---- ------ 1.4
---- -- --
\.2
1.0
/--
,..::: ------------------------------- \.2 -- -----
/,:.------ -------------
\~~~:~:=~----------
1.0
0.8
0.6
0.4
---- ----.
-----
0.8
0.6
------------
------
0.2 +--.--...,..--.........--.--..--""T""---.-...---I
12345678910 2 4 5 6 7 8 9 10
United States
1.6 -..-_ _ _ _ _ _ _ _ _ _ _ _--,
1.4
1.2
---- ------ -- ----
1.0
0.8
0.6
0.4
0.2 +--.--...,..--.........--.--..----r----.-...---I
2 3 4 5 6 7 8 9 10
Using the maximum likelihood estimates of the Markov regime switching model in (7.17), equilibrium
stock returns are simulated according to (7.15), with r=3 and 13=0.98. Variance ratios are calculated on
the basis of simulated returns. Each box shows the point estimates of historical variance ratios, the
median, and both the 66 percent and 90 percent confidence intervals of Monte Carlo simulations. The
horizontal axis denotes the number of quarters k used to calculate returns, variance ratios are plotted
on the vertical axis.
The analysis clearly shows that predictability of long horizon returns is not at all
inconsistent with the concept of market efficiency. Accordingly, stock prices
need not follow a random walk in an efficient market. Changes in the equilib-
rium return on stocks can be predicted to the extent that there is predictability in
the endowment process. To maximize utility, investors attempt to smooth con-
sumption by adjusting the required rate of return for financial assets. For exam-
ple, assume that investors know that the economy is in a good state (Sl =1) .
Given the maximum likelihood estimates of the system in (7.17), being in a
good state today implies that the economy is very likely to remain in a high-
growth state in the future. The intertemporal models by Campbell (1993, 1996)
and RestoylWeil (1998) show that this has two effects that work in opposite di-
rections. 41 First, there is an intertemporal relative price effect in which the
higher expected future endowment implies a lower relative price of future
goods. This induces agents to increase savings and increase their demand for
assets. Increased asset demand works to raise current asset prices. Second, an
intertemporal substitution effect, arising from investors' attempt to smooth con-
sumption, works in the opposite direction. When the expected future endowment
is high, investors attempt to increase current consumption by selling off part of
their holdings. In equilibrium, however, the agent has to hold his pile of assets,
which can only happen - accepting a price-based explanation - if asset prices
are low and expected returns are high. Experimenting with different degrees of
risk aversion, it turns out that log utility (y =1) is indeed the borderline case in
which the intertemporal price effect and the substitution effect cancel out. When
the utility function is more concave than implied by log utility, the consumption
smoothing effect dominates. This is perfectly in line with Campbell's (1996)
cross-sectional restriction in (2.64). By the same logic, investors anticipating
lower real activity in the future will attempt to transfer wealth to this anticipated
period of scarcity by buying additional assets. Hence, today's stock prices must
be high, implying low expected rates of return.
Figure 7.5
Stock returns over the business cycle
Period: 1970.01- 1998.08 (monthly data)
25%
20%
15%
10%
~
5%
I.
11- '1111 I - '1
0%
-5%
1 I
-10%
-15%
-20%
-25%
.
c: :g
oS
.~
E '"
~
1l >-
ii 0
c:
j ii -Ol >-
oS
c:
aCo. "'il"
"0
ii
E
0 !l 1;
~ 13
[;i E ~
"- E :.: ~ 1iii ~
5 en ~
§ gp en9
"0
~
'"
U
~ " c: z Vl
;;;: ]
0 0
:c "ii
z
~
c
en ] ::>
'c
::>
The graph shows average returns on developed stock markets, classified into boom and recession
phases. All numbers are in percent per year. Business cycle dates are taken from the National Bureau
of Economic Research (NBER). Monthly stock market data are from Morgan Stanley Capital Interna-
tional (MSCI).
Overall, the results in this chapter confirm the general notion presented through-
out this book: the predictable time variation of expected returns is closely related
to real activity. Starting with the proposition that consumption is the best meas-
ure of bad times, evidence has been provided for cyclical components in con-
Mean reversion and rational asset pricing 295
sumption growth. Given the linkage between consumption growth and stock re-
turns via the Euler-equation in (7.7), stock returns should also exhibit a predict-
able component that is related to the business cycle. This might indicate a pre-
mium for holding macroeconomic risks associated with the business cycle, i.e.
for holding assets that do poorly in bad states of the world. The relevant infor-
mation is reflected in prices: high prices lead to low expected returns and low
prices lead to high expected returns. Clearly, the model presented here can nei-
ther replicate the high equity premium nor any other quantity. However, it gives
rise to the notion that the historical equity premium is in fact related to inves-
tors' aversion to hold stocks because of their exposure to some relevant state
variable or risk factor, e.g., 'recession risk' as Campbell/Cochrane (1999a) call
it. In other words, the results indicate a premium for holding macroeconomic
risk associated with the business cycle and for holding stocks that do poorly in
recessions. To further illustrate this intuition, figure 7.5 shows the average re-
turns on developed stock markets, classified into recession and boom phases.
The business cycle dates are taken from the National Bureau of Economic Re-
search (NBER). Monthly stock returns over the period from 1970.01 to 1998.08
are taken from Morgan Stanley Capital International (MSCI). Average returns
are positive during booms, and negative during recessions (all numbers are in
percent per year). This seems to be a global phenomenon, and it is in accordance
with the general notion: for expected returns to be high in recessions, prices
must fall. Fortunately, the empirical results are also consistent with the eco-
nomic story presented to motivate the ad-hoc instrumental regression approach
in chapter 4.
capital asset pricing model (CAPM). Every investor holds the market portfolio;
the only decision is how much of it to hold. The 'two-fund' theorem also implies
that the investor's time horizon does not matter. This has a good reason. Tradi-
tional finance argued that prices are close to a random walk, making prices in-
dependent over time. Therefore, stocks are not safer in the long run, and the
risk-free/risky asset allocation is independent of the time horizon. The ratio of
means and variances of stock returns is constant, regardless of the time hori-
zon. 42 However, these propositions of the traditional view have come under at-
tack only recently. First, the simulation results in this chapter provide strong
evidence for a 'recession factor' that is relevant for pricing a cross-section of
assets. Cochrane (1999c) combines these observations within Fama's (1996)
framework of a multi-factor efficient frontier. 43 Investors do not only want port-
folios with higher average returns and lower volatility, but also require portfo-
lios not to go down in recessions. Specifically, if a portfolio keeps paying off in
a recession, investors might be willing to accept a little lower expected return
and a somewhat higher standard deviation. In other words, the classical mean-
variance tradeoff must be extended. The market return is no longer efficient in a
multifactor setup, and the average investor - even if he or she is immune against
recession risk - should no longer hold the market portfolio. Intuitively, the new
tangency portfolio is tilted towards recession sensitive stocks. The 'two-fund'
theorem becomes a 'three-fund' theorem: the optimal set of portfolios consists
of (i) the risk-free rate, (ii) the market-portfolio, and (iii) one additional multi-
factor efficient portfolio. Of course, this analysis is directly linked to Merton's
(1973) intertemporal capital asset pricing model (ICAPM), where the investor
holds hedge portfolios to be immune against changes in the investment opportu-
nity set.44
Second, if stock returns are mean reverting, stocks seem less risky in the long
run. Mean and variance no longer scale the same way with the time horizon.
Negative serial correlation implies that the variance of long horizon returns is
lower than the variance of one-period returns times the holding period. Variance
ratios below one make stocks an attractive long horizon investment. Indeed, re-
cent research by Barberis (2000) shows that predictability leads to a greater allo-
cation in stocks, even if parameter uncertainty in predicting returns is taken into
account. Brandt (1999) reverses the traditional approach of empirical asset
pricing tests. He uses the Generalized Method of Moments (GMM) to estimate
portfolio rules that best satisfy the intertemporal Euler-equation, given the sto-
chastic properties of historical price data. In other words, while the standard ap-
proach is to fix the consumption or wealth process and estimate the preference
parameters, Brandt fixes the preference parameters and assumes instead that the
composition of the optimal portfolio is unknown. He extends this technique to
multiperiod problems and problems in which the allocation decision depends on
conditioning information, i.e. market-timing. Brandt reports a mild horizon ef-
fect, and a strong market timing effect. Specifically, (i) longer term investors
hold more stocks and (ii) the optimal fraction of wealth invested in stocks is
very sensitive to changes in the instrument variables. CampbellNiceira (1999)
calculate an actual solution to the optimal market-timing question when invest-
ment opportunities are time varying. While in most other models investors are
generally assumed to have power utility defined over wealth at a single terminal
date, they model investors with a desire for lifetime consumption rather than
portfolio returns at a fixed horizon. Their results show that the optimal invest-
ment strategy takes strong advantage of market timing. Unfortunately, their 'op-
timal' strategy would not have captured the strong bull market in recent years. 45
This reveals that market timing requires the willingness to suffer losses from a
portfolio that does not track the index, explaining the reluctance of portfolio
managers to follow this advice. In the end, higher returns require higher risk. In
particular, market timing requires buying at the bottom when everyone else is in
panic (i.e. in bad states), and selling at the top when the crowd wants to hold
stocks. While the market-timer outperforms in the long run, he or she often un-
derperforms a benchmark over shorter periods of time.
An important caveat is that all these models are partial equilibrium in nature.
Specifically, they solve the microeconomic problem of a given investor facing
exogenous asset returns, but it is not shown how these asset returns could be
consistent with general equilibrium. There is no answer to this question yet, but
it is important to remember that the average investor must hold the market.
Therefore, multiple factors and predictability do not alter the portfolio advice for
the average investor. The regime switching model can provide the intuition.
Modern asset pricing theory interprets capital market equilibrium in terms of
portfolio optimality conditions. 46 In particular, in equilibrium the marginal util-
ity gain from an additional asset bought at time t to be sold for consumption
purposes at t + 1 is exactly offset by the utility loss caused by consuming less in
the current period. Any attempt to exploit predictability will increase variation in
consumption, thereby decreasing expected utility. More specifically, Balvers!
CosimanolMcDonald (1990) argue that an investor would have to give up con-
sumption right in times when he or she likes to consume much. In a bottom of a
bust an investor would have to time high expected returns and put all funds into
the stock market. This, however, is not consistent with utility-maximizing be-
This leads to a final interesting question: once investors know about predictabil-
ity of stock returns, will it vanish in the future? If the business cycle story pro-
posed in this book is true, there is no reason that predictability in returns disap-
pears due to intertemporal ' adjustments'. Predictability is due to changes in real
returns over time caused by the persistence of real macroeconomic shocks, not
by some irrational market inefficiency or any kind of behavioral phenomenon.
Some investors may adjust intertemporal consumption plans based on expected
real returns, but there is no reason that the magnitude of adjustment could cause
predictable changes in returns to fully disappear. However, those that 'exploit'
predictability should be different from everyone else. A market-timer must be
recession-immune to be a provider of 'insurance' .47 Recession averse investors
can reduce their exposure to stocks that are particularly sensitive to declines in
real activity only by paying a premium to those investors who deliberately take
on recession risk. In the end, the rationale of any risk-based explanation is that a
strategy that works in the market must have a real economic function.
This chapter starts with an ad-hoc model of the relationship between consump-
tion and asset returns. Global instrument variables with prediction power for
stock returns are used to identify mean-reverting components in consumption
growth. Given the conditional means and volatilities of consumption rates, time
varying expected excess returns can be estimated in a simple regression setup. In
the second part of this chapter a thorough economic model is specified and
simulated using the maximum likelihood estimates of a two-state regime
switching model for consumption growth. The intuition is simple: the stochastic
properties of the consumption process are assumed to have direct implications
for the behavior of stock returns. Specifically, mean reverting components in
consumption imply mean reversion in stock returns, and this is fully consistent
with rational asset pricing. Investors prefer smooth consumption patterns and
adjust the return they expect from investing in stocks accordingly. The main re-
sults of this chapter are as follows:
Conditional moments of consumption growth are time varying: There are good
reasons to believe that consumption provides the best measure of bad times. In-
vestors consume less when their income prospects are low or if they think future
returns will be bad. When consumption is low investors would like to hold port-
folios that do not perform badly, and would even be willing to pay a premium to
Mean reversion and rational asset pricing 301
ensure that wish. Hence, the general hypothesis is that the states of the economy
that are relevant for determining equilibrium asset returns can be described by
the conditional mean and the conditional volatility of consumption growth.
• The null hypothesis that the conditional mean and the conditional standard
deviation of consumption growth do not depend on a set of global instrument
variables can be rejected. As expected, the conditional mean growth is lowest
during recessions and highest during booms of real activity. In contrast,
peaks in the conditional standard deviation occur during recessions. In other
words, the conditional volatility of consumption growth exhibits countercy-
clical movements.
• There is empirical evidence that the stock market risk premium depends on
the conditional moments of logarithmic consumption growth. Peaks in ex-
pected excess returns occur closely before business cycle troughs. In contrast,
expected excess returns are lowest at the end of a boom, when prices are
high. This is perfectly consistent with the consumption smoothing story. The
average world market risk premium is estimated 6.5 percent per annum.
Mean reversion is compatible with rational asset pricing: There are two ingre-
dients for an economic model with time variation in expected stock returns: (i)
trend reversion in the stochastic discount factor and (ii) the life cycle-permanent
income hypothesis. A variant of the classical Lucas-type endowment economy
captures both consumption smoothing and risk-correction. The key intuition of
the model is that there is an immediate link between the properties of the sto-
chastic discount factor and the underlying capital market equilibrium.
• The identified states are very persistent. When the economy is in a boom, the
results show that the probability that it remains in the good state during the
next quarter is well above 0.7 for all G7 countries. The chances to remain in
a low state in the next quarter once the economy has come into a recession
are somewhat lower, but they also reveal considerable persistence. This is
reasonable, because it implies that booms on average are longer in duration
than recessions.
• Looking at historical stock returns, there is clear evidence for mean reversion
at holding periods greater than six quarters. At the same time, there is evi-
dence for positive autocorrelation at shorter horizons. In more technical
words, variance ratios start above one, but eventually fall below one in most
G7 countries.
• As the corresponding value of the relative risk aversion is not able to repli-
cate historical excess returns, yet another puzzle arises: the mean reversion
puzzle. A level of risk aversion consistent with the equity premium results in
Mean reversion and rational asset pricing 303
a too high speed of mean reversion. In addition, the model cannot explain the
short-term momentum in stock returns. For time horizons shorter than six
months the empirical variance ratios lie well above the 90 percent confidence
interval generated by Monte Carlo simulations.
Predictability does not necessarily imply that market timing is optimal: Predict-
ability raises the question whether investors can gain excess profits. Mean rever-
sion makes stocks more attractive investments in the long run, and market tim-
ing would imply increased allocation to stocks when expected returns are high.
• The average investor must hold the market. The portfolio advice for the aver-
age investor is not altered by multiple factors and return predictability. The
latter is consistent with the general notion of market efficiency, implying that
any attempt to exploit return predictability increases variation in consump-
tion, thereby decreasing expected utility. The market timer must give up con-
sumption right in times when he or she likes to consume much, i.e. in a re-
cession. Any attractiveness of higher expected stock returns is exactly offset
by the risk of a less smooth consumption pattern. The selling pressure in-
duced by an attempt to run down savings in a recession is exactly balanced
by a countercyclical increase in expected returns.
305
Chapter 8
This study has reported many figures and tables. Hop'!fully, the big picture has
not been lost in too much formalism and technicalities. The goal of this book has
been to explore the interplay between time varying expected returns, consump-
tion, and the business cycle on global stock markets. The main idea that has
come up again and again is that expected returns vary with the business cycle. In
a nutshell, if the time variation in expected returns is rational, driven by shocks
to taste or technology, then the variation in expected returns should be related to
variation in consumption, investment, and savings. In this sense, predictability
of stock returns is perfectly consistent with the concept of market efficiency, and
stock prices need not follow a random walk. This notion is empirically tested
using different methodologies. In particular, the study starts with simple linear
regressions, extracts business cycle components from macroeconomic time se-
ries, tests both beta pricing and stochastic discount factor models in their uncon-
ditional as well as conditional versions and, finally, simulates a slightly adapted
version of the Lucas (1978) endowment economy. The examination includes
quarterly and monthly data of both developed and emerging stock markets
starting in the early 1970s (if available). Returns are all denominated in Swiss
francs.
• Predictability does not imply that stock markets are inefficient: Investors
dislike variation in their consumption patterns. To smooth consumption over
time and across states of nature, rational investors buy stocks. Prices and ex-
pected returns must adjust accordingly. This risk-based explanation implies
that investors do not require a high equity premium because they dislike
stocks per se, but because they do poorly in recessions. To illustrate the no-
tion of consumption-based asset pricing, assume that the economy is ex-
pected to do well in the future. Investors attempt to increase current con-
sumption by selling off part of their holdings. In equilibrium, however, the
agent has to hold his pile of assets, which can only happen - accepting a risk-
based explanation - if asset prices are low and expected returns are high.
Alas, expected returns are higher in recessions than in booms. Even though
current models cannot explain the level of the historical equity premium, this
notion is comfortable from an economic point of view. Predictability is
caused by changes in real returns over time caused by the persistence of real
macroeconomic shocks, not merely by some irrational market inefficiency or
any kind of behavioral phenomenon. Recession averse investors can reduce
their exposure to stocks that are particularly sensitive to declines in real ac-
tivity, but to do so they have to pay an insurance premium to the investors
deliberately taking on the recession risk. Empirically, these results are sup-
ported by the following observations. First, extracting the business cycle
components of instrument and stock price series reveals that the correlations
On the contributions of this study 307
on these medium-tenn frequencies are high, especially for the dividend yield
and the tenn spread. Second, meaningful cross-sectional restrictions are im-
posed on the time series behavior of expected returns in multibeta models.
Time variation of risk premiums on a small number of risk factors (in combi-
nation with constant conditional betas) can explain a good fraction of the
predictable variation in expected returns. Unfortunately, direct evidence for a
conditional version of the international consumption capital asset pricing
model is less encouraging. Finally, a fully rational equilibrium asset pricing
model in the spirit of Lucas' endowment economy with two states defined
over the conditional moments of consumption rates is simulated using his-
torical data from the G7 countries. This framework fully incorporates the in-
terplay between consumption smoothing and risk correction, allowing to ex-
plain mean reversion over longer horizons, but not momentum over shorter
horizons.
• Global stock markets are becoming increasingly integrated: With the glob-
alization of stock markets, it is widely believed that capital markets are be-
coming more integrated. The fact that returns on global stock markets be-
come more correlated, however, does not imply efficient pricing. In particu-
lar, capital markets are said to be integrated when assets in different countries
display the same risk-adjusted expected return. The question is not only a
theoretical one, but has important practical implications. Most important, the
cost of capital can be expected to decrease once a market becomes integrated
into the world market. This, by itself, promotes economic growth, which is
particularly relevant in the emerging countries. Another testable implication
is that integration directly bears on the desirability of international diversifi-
cation. If national stock markets are segmented, international portfolios
should exhibit superior risk-adjusted perfonnance because some of the sys-
tematic risk can be diversified away by investing internationally without
paying a price in tenns of lower returns. The empirical evidence in this study
is twofold. First, latent variable models with time varying risk premiums on a
small number of common factors (in combination with constant conditional
betas) assume perfect integration, i.e. perfect correlation of expected returns
across global stock markets. The ultimate test for integration from a risk-
308 Chapter 8
sharing perspective is the use of global consumption rates as the only state
variable. Unfortunately, this test fails, which is a manifestation of the equity
premium puzzle. Second, spanning tests in the HansenlJagannathan (1991)
framework show that the minimum volatility of the stochastic discount factor
required to price a set of international assets increases compared to a set of
Swiss assets only. However, the shifts in bounds (or equivalently, the shifts
in the efficient frontiers for portfolio returns) are often economically insig-
nificant and lost in sampling error. This is especially the case for the emerg-
ing markets, which have become more integrated only recently. The times of
'diversification-free lunches' are over. However, hedging the currency com-
ponent of returns makes it easier to realize the potential gains from interna-
tional diversification. More important, conditioning returns on instrument
variables and managing a portfolio tactically allows to achieve a significantly
improved risk-adjusted performance, both from an economic and a statistical
point of view.
A lot of research is still necessary. Most important, the equity premium puzzle is
still an unsolved question. Further analysis is also necessary to explore the equi-
librium implications of predictability for optimal portfolio choice. Despite the
promise of such research, it is probably unrealistic to hope for a fully rational,
risk-based explanation of all the empirical patterns that have been discovered in
stock returns. All existing models to explain the equity premium still require
extremely high risk aversion. The general lesson, however, is that to justify the
equity premium, people must be averse to holding stocks because of their expo-
sure to some other risk factor in addition to wealth risk, such as recessions or
changes in the investment opportunity set. To believe in the equity premium,
one has to believe in these types of stories. Alternatively, of course, the long-run
equity premium is much smaller than the average post-war excess return. The
standard model was right after all, historically high stock returns were largely
due to a transitory phenomenon, not to say luck. Unfortunately, this would mean
that one has to reduce the estimates for the future equity premium.
309
Bibliography
Abel, A., 1990, Asset Prices under Habit Formation and Catching Up with the
lonses, American Economic Review 80, pp. 38-42
Abel, A., 1994, Exact Solutions for Expected Rates of Return Under Markov
Regime Switching: Implications for the Equity Premium Puzzle, Journal of
Money, Credit and Banking 26, pp. 345-361
Abel, A., and O. Blanchard, 1986, The Present Value of Profits and the Cyclical
Variability of Investment, Econometrica 54, pp. 249-273
Adler, M., and B. Dumas, 1983, International Portfolio choice and corporation
finance: A synthesis, Journal of Finance 38, pp. 925-984
Adler, M., and M. Granito, 1991, Should International Portfolios be Perma-
nently Hedged?, in: I. Khoury (ed.), Recent Developments in International
Banking and Finance, Amsterdam: Elsevier, pp. 277-313
Adler, M., and B. Prasad, 1992, On universal currency hedges, Journal of Fi-
nancial and Quantitative Analysis 27, pp. 19-38
Admati, R., and P. Pfleiderer, 1985, Interpreting the factor risk premia in the
arbitrage pricing theory, Journal of Economic Theory 35, pp. 191-195
Altug, S., and P. Labadie, 1994, Dynamic Choice and Asset Markets, San Di-
ego: Academic Press
Anderson, R., and I.P. Danthine, 1981, Cross Hedging, Journal of Political
Economy 89, pp. 1183-1197
Antoniou, A., I. Garrett, and R. Priestley, 1998a, Calculating the Equity Cost of
Capital Using the APT: The impact of the ERM, Journal of International Money
and Finance 17, pp. 949-966
Antoniou, A., I. Garrett, and R. Priestley, 1998b, Macroeconomic Variables as
Common Pervasive Risk Factors and the Empirical Content of the Arbitrage
Pricing Theory, Journal of Empirical Finance 5, pp. 221-240
Arrow, K., 1964, The Role of Securities in the Optimal Allocation of Risk
Bearing, Review of Economic Studies 31, pp. 91-96
Asness, C., 1991, Negative expected returns and the business cycle, working
paper, Graduate School of Business, University of Chicago
310 Bibliography
Backus, D., P. Kehoe, and F. Kydland, 1992, International real business cycles,
Journal of Political Economy 100, pp. 745-775
Bakshi, G., and A. Naka, 1997, An Empirical Investigation of Asset Pricing
Models Using Japanese Stock Market Data, Journal of International Money and
Finance 16, pp. 81-112
Ball, R., 1978, Anomalies in relationships between securities' yields and yield
surrogates, Journal of Financial Economics 6, pp. 103-126
Balvers, R., T. Cosimano, and B. McDonald, 1990, Predicting Stock Returns in
an Efficient Market, Journal of Finance 45, pp. 1109-1128
Balvers, R., Y. Wu, and E. Gilliand, 2000, Mean Reversion Across National
Stock Markets and Parametric Contrarian Investment Strategies, Journal of Fi-
nance 55, pp. 745-772
Bansal, R., A. Hsieh, and S. Viswanathan, 1993, A new approach to interna-
tional arbitrage pricing, Journal of Finance 48, pp. 1719-1747
Barberis, N., 2000, Investing for the Long Run when Returns are Predictable,
forthcoming in: Journal of Finance
Barberis, N., A. Schleifer, and R. Vishny, 1998, A Model oflnvestor Sentiment,
Journal of Financial Economics 49, pp. 307-344
Barro, R., 1990, The stock market and investment, Review of Financial Studies
3, pp. 115-131
Baxter, M., 1994, Real exchange rates and real interest differentials: Have we
missed the business cycle component, Journal of Monetary Economics 33, pp.
5-37
Baxter, M., and U. Jerrnann, 1995, The international diversification puzzle is
worse than you think, working paper, Rochester Series for Economic Research
Baxter, M., and R. King, 1998, Measuring Business Cycles: Approximate Band
Pass Filters for Economic Time Series, working paper, University of Virginia
Beckers, S., G. Connor, and R. Curds, 1996, National Versus Global Influences
on Equity Returns, Financial Analysts Journal 52, pp. 31-39
Beckers, S., R. Grinold, A. Russ, and D. Stefek, 1992, The relative importance
of common factors across the European equity markets, Journal of Banking and
Finance 16, pp. 75-95
Bekaert, G., 1995, Market integration and investment barriers in emerging eq-
uity markets, World Bank Economic Review 9, pp. 75-107
Bekaert, G., and C. Harvey, 1995, Time-varying world market integration,
Journal of Finance 50, pp. 403-444
311
Bekaert, G., and C. Harvey, 1997, Emerging Equity Market Volatility, Journal
of Financial Economics 43, pp. 29-78
Bekaert, G., and C. Harvey, 2000, Foreign Speculators and Emerging Equity
Markets, Journal of Finance 55, pp. 565-613
Bekaert, G., and R. Hodrick, 1992, Characterizing Predictable Components in
Excess Returns on Equity and Foreign Exchange Markets, Journal of Finance
47, pp. 467-509
Bekaert, G., and J. Liu, 1999, Conditioning Information and Variance Bounds
on Pricing Kernels, NBER working paper 6880
Bekaert, G., R. Lumsdaine, and C. Harvey, 1998, Dating the Integration of
World Capital Markets, NBER working paper 6724
Bekaert, G., and M. Urias, 1996, Diversification, Integration, and Emerging
Market Closed End Funds, Journal of Finance 51, pp. 835-869
Bekaert, G., and M. Urias, 1997, Is There a Free Lunch in Emerging Market In-
vesting?, working paper, Stanford University
Black, F., 1972, Capital market equilibrium with restricted borrowing, Journal
of Business 45, pp. 444-454
Black, F., 1990a, Mean Reversion and Consumption Smoothing, Review of Fi-
nancial Studies 1, pp. 107-114
Black, F., 1990b, Equilibrium exchange rate hedging, Journal of Finance 45, pp.
899-907
Bodhurta, N., and N. Mark, 1991, Testing the CAPM with time-varying risks
and returns, Journal of Finance 46, pp. 1485-1506
Boldrin, M., L.J. Christiano, and J.M.D. Fisher, 1995, Asset pricing lessons for
modeling business cycles, NBER working paper 5262
Bollerslev, T., R. Engle, and J. Wooldridge, 1988, A Capital Asset Pricing
Model with Time Varying Covariances, Journal of Political Economy 96, pp.
116-131
Bonomo, M., and R. Garcia, 1996, Consumption And Equilibrium Asset Pric-
ing: An Empirical Assessment, Journal of Empirical Finance 3, pp. 240-266
Bossaerts, P., and R. Green, 1989, A general equilibrium model of changing risk
premia: Theory and tests, Review of Financial Studies 2, pp. 467-493
Bossaerts, P., and P. Hillion, 1999, Implementing Statistical Criteria to Select
Return Forecasting Models: What Do We Learn, Review of Financial Studies
12, pp. 405-428
Boudhoukh, J., M. Richardson, and T. Smith, 1993, Is the ex ante risk premium
always positive?, Journal of Financial Economics 34, pp. 387-408
312 Bibliography
Errunza, V., and E. Losq, 1985, International asset pricing under mild segmen-
tation: Theory and test, Journal of Finance 40, pp. 104-124
Estrella, A., and G. Hardouvelis, 1991, The Term Structure as a Predictor of
Real Economic Activity, Journal of Finance 46, pp. 555-576
Evans, D., 1994, Expected returns, time varying risk, and risk premia, Journal of
Finance 49, pp. 655-679
Fama, A., 1970a, Efficient Capital Markets: A Review of Theory and Empirical
Work, Journal of Finance 25, pp. 383-417
Fama, A., 1970b, Multiperiod Consumption-Investment Decisions, American
Economic Review 60, pp. 163-l74
Fama, E., 1981, Stock returns, real activity, inflation, and money, American
Economic Review 71, pp. 545-565
Fama, E., 1990, Stock Returns, Expected Returns, and Real Activity, Journal of
Finance 45, pp. 1089-1108
Fama, E., 1991, Efficient Capital Markets: II, Journal of Finance 46, pp. 1575-
1618
Fama, E., 1996, Multifactor Portfolio Efficiency and Multifactor Asset Pricing,
Journal of Financial and Quantitative Analysis 31, pp. 441-465
Fama, E., and K. French, 1988a, Permanent and Temporary Components of
Stock Prices, Journal of Political Economy 96, pp. 246-273
Fama, E., and K. French, 1988b, Dividend Yields and Expected Stock Returns,
Journal of Financial Economics 22, pp. 3-27
Fama, E., and French, K., 1989, Business conditions and expected returns on
stocks and bonds, Journal of Financial Economics 25, pp. 23-49
Fama, E., and K. French, 1992, The Cross-Section of Expected Stock Returns,
Journal of Finance 47, pp. 427-465
Fama, E., and K. French, 1993, Common factors in the returns on stocks and
bonds, Journal of Financial Economics 33, pp. 3-56
Fama, E., and K. French, 1996, Multifactor Explanations of Asset Pricing
Anomalies, Journal of Finance 51, pp. 55-84
Fama, E., and 1. MacBeth, 1973, Risk, Return, and Equilibrium: Empirical
Tests, Journal of Political Economy 71, pp. 607-636
Fama, E., and W. Schwert, 1977, Asset returns and inflation, Journal of Finan-
cial Economics 5, pp. 115-146
319
Ferson, W., S. Sarkissian, and T. Simin, 1999a, The alpha factor asset pricing
model: A parable, Journal of Financial Markets 2, pp. 49-68
Ferson, W., S. Sarkissian, and T. Simin, 1999b, Spurious regressions in Finan-
cial Economics? working paper, University of Washington
Ferson, W., and R Schadt, 1996, Measuring Fund Strategy and Performance in
Changing Economic Conditions, Journal of Finance 51, pp. 425-462
Ferson, W., and A. Siegel, 1999, Stochastic Discount Factor Bounds with Con-
ditioning Information, working paper, University of Washington
Filardo, A.J., 1994, Business Cycle Phases and Their Transitional Dynamics,
Journal of Business and Economic Statistics 12, pp. 299-308
Fischer, S., and R. Merton, 1984, Ecroeconomics and Finance: The Role of the
Stock Market, with S., in: Brunner, K., and A.H. Meltzer (eds.), Essays on Mac-
roeconomic Implications of Financial and Labor Markets and Political Proc-
esses, Vol. 21, Amsterdam: North-Holland
Fisher, I., 1930, The Theory of Interest, New York: Macmillian
French, K., W. Schwert, and R Stambaugh, 1987, Expected stock returns and
volatility, Journal of Financial Economics 14, pp. 217-236
Friedman, M., 1957, A theory of the consumption function, Princeton: Princeton
University Press
Friend, I., and M. Blume, 1975, The Demand for Risky Assets, American Eco-
nomic Review 65, pp. 900-022
r
Foster, F.D., T. Smith, and RE. Whaley, 1997, Assessin Goodness-of-Fit of
Asset Pricing Models: The Distribution of the Maximal R , Journal of Finance
52, pp. 591-607
Froot, K., and K. Rogoff, 1995, Perspectives on PPP and long-run real exchange
rates, in: Grossman, G., and K. Rogoff, Handbook of international economics,
Vol. 3, Amsterdam: North Holland
Gallant, R, 1997, An Introduction to Econometric Theory, Princeton: Princeton
University Press
Gallant, R, L.P. Hansen, and G. Tauchen, 1990, Using Conditional Moments of
Asset Payoffs to Infer the Volatility of Intertemporal Marginal Rates of Substi-
tution, Journal of Econometrics 45, pp. 141-179
Geske, R, and R Roll, 1983, The monetary and fiscal linkage between stock
returns and inflation, Journal of Finance 38, pp. 1-33
Ghysels, E., 1998, On Stable Factor Structures in the Pricing of Risk: Do Time
Varying Betas Help or Hurt?, Journal of Finance 53, pp. 549-498
321
Gibbons, M., and W. Ferson, 1985, Testing asset pricing models with changing
expectations and an unobservable market portfolio, Journal of Financial Eco-
nomics 14, pp. 217-236
Giovannini, A., and P. Weil, 1989, Risk Aversion and Intertemporal Substitution
in the Capital Asset Pricing Model, NBER working paper 2824
Glen, 1., and P. Jorion, 1993, Currency Hedging for International Portfolios,
Journal of Finance 48, pp. 1865-1886
Glosten, L., R. Jagannathan, and D. Runkle, 1993, On the Relation Between the
Expected Value and the Volatility of the Nominal Excess Stock Return on
Stocks, Journal of Finance 48, pp. 1779-1801
Goetzman, W., and P. Jorion, 1999a, Global Stock Markets in the Twentieth
Century, Journal of Finance 54, pp. 953-980
Goetzman, W., and P. Jorion, 1999b, Re-emerging markets, Journal of Finan-
cial and Quantitative Analysis 34, pp. 1-32
Gordon, M., and E. Shapiro, 1956, Capital Equipment Analysis: The Required
Rate of Profit, Management Science 3, pp. 102-110
Grauer, F., R. Litzenberger, and R. Stehle, 1976, Sharing rules and equilibrium
in an international capital market under uncertainty, Journal of Financial Eco-
nomics 3, pp. 233-256
Greene, W., 2000, Econometric Analysis, Upper Saddle River: Prentice Hall, NJ
Griffin, H., and A. Karolyi, 1998, Another look at the role of the industrial
structure of markets for international diversification strategies, Journal of Fi-
nancial Economics 50, pp. 351-373
Grinold, R., A. Rudd, and D. Stefek, Global Factors: Fact or Fiction?, Journal of
Portfolio Management 16, pp. 79-88
Grossman, S., A. Melino, and R. Shiller, 1987, Estimation of the continuous-
time consumption based asset pricing model, Journal of Business Economics
and Statistics 5, pp. 315-327
Grossman, S., and R. Shiller, 1981, The Determinants of the Variability of Stock
Market Prices, American Economic Review 71, pp. 222-227
Grossman, S., and R. Shiller, 1982, Consumption Corre1atedness and Risk
Measurement in Economics with Non-Traded Assets and Heterogeneous Infor-
mation, Journal of Financial Economics 10, pp. 195-210
Grossman, S., and J. Stiglitz, 1980, On the Impossibility of Informationally Ef-
ficient Markets, American Economic Review 70, pp. 393-408
Grubel, H.G., and K. Fadner, 1971, The Interdependence ofInternational Equity
Markets, Journal of Finance 26, pp. 89-94
322 Bibliography
Gultekin, M., B. Gultekin, and A. Penati, 1989, Capital Controls and Interna-
tional Capital Market Segmentation: The Evidence from the Japanese and
American Stock Markets, Journal of Finance 44, pp. 849-869
Hall, R., 1978, Stochastic implications of the life cycle-permanent income hy-
pothesis: theory and evidence, Journal of Political Economy 96, pp. 339-357
Hall, R., 1988, Intertemporal Substitution in Consumption, Journal of Political
Economy 96, pp. 221-273
Hall, R., 1989, Consumption, in: Barro, R. (ed.), Modern Business Cycle The-
ory, Cambridge: Harvard University Press, pp. 153-177
Hamilton, J., 1989, A New Approach to the Economic Analysis of Nonstation-
ary Time Series and the Business Cycle, Econometrica 57, pp. 357-384
Hamilton, J., 1994, Time series analysis, Princeton: Princeton University Press
Hansen, L.P., 1982, Large Sample Properties of generalized method of moments
estimators, Econometrica 50, pp. 1029-1054
Hansen, L.P., J. Heaton, and E. Luttmer, 1995, Econometric evaluation of asset
pricing models, Review of Financial Studies 8, pp. 237-274
Hansen, L.P., and R. Hodrick, 1983, Risk averse speculation in the forward ex-
change market: an econometric analysis of linear models, in: Frenkel, J. (ed.),
Exchange Rates in International Macroeconomics, Chicago: University of Chi-
cago Press
Hansen, L.P., and R. Jagannathan, 1991, Implications of Security Market Data
for Models of Dynamic Economics, Journal of Political Economy 99, pp. 225-
262
Hansen, L.P., and R. Jagannathan, 1997, Assessing specification errors in sto-
chastic discount factor models, Journal of Finance 52, pp. 557-590
Hansen, L.P., and S. Richard, 1987, The role of conditioning information in de-
ducing testable restrictions implied by dynamic asset pricing models, Econo-
metrica 50, pp. 1269-1286
Hansen, L.P., and K. Singleton, 1982, Generalized Instrumental Variables Esti-
mation of Nonlinear Rational Expectation Models, Econometrica 50, pp. 1269-
1288
Hansen, L.P., and K. Singleton, 1983, Stochastic Consumption, Risk Aversion
and the Temporal Behavior of Asset Returns, Journal of Political Economy 93,
pp.249-268
Hardouvelis, G., D. Kim, and T. Wizman, 1996, Asset pricing models with and
without consumption data: An empirical evaluation, Journal of Empirical Fi-
nance 3, pp. 267-301
323
Hargis, K., and W. Maloney, 1996, Emerging Equity Markets: Are they for
Real?, working paper, University of Illinois at Urbana-Champaign
Harrison, M., and D. Kreps, 1979, Martingales and Arbitrage in Multiperiod Se-
curities Markets, Journal of Economic Theory 20, pp. 381-408
Harrison, M., and S. Pliska, 1981, Martingales and Stochastic Integrals in the
Theory of Continuous Trading, Stochastic Processes and Their Applications 11,
pp.215-260
Harvey, C.R., 1988, The real term structure and consumption growth, Journal of
Financial Economics 22, pp. 305-314
Harvey, c.R., 1991a, The World Price of Covariance Risk, Journal of Finance
46, pp. 111-157
Harvey, C.R., 1991b, The term structure and world economic growth, Journal of
Fixed Income, pp. 7-19
Harvey, C.R., 1991c, The specification of conditional expectations, working pa-
per, Duke University
Harvey, C.R., 1994, Comments on Dumas (1994b), in: Frankel, J. (ed.), The In-
ternationalization of Securities Markets, Chicago: University of Chicago Press
Harvey, C.R., 1995, Predictable Risk and Returns in Emerging Markets, Review
of Financial Studies, 8, pp. 773-816
Harvey, C.R., 1995, Global risk exposure to a trade-weighted foreign currency
index, working paper, Duke University
Harvey, C.R., 1997, The International Cost of Capital and Risk Calculator,
working paper, Duke University
Harvey, c.R., B. Dumas, and P. Ruiz, 1999, Are Common Swings in Interna-
tional Stock Returns Justified by Subsequent Changes in National Outputs,
working paper, Duke University and HEC
Harvey, C.R., and C. Kirby, 1996, Instrumental Variables Estimation of Condi-
tional Beta Pricing Models, in: Maddala, G. and C. Rao (eds.), Handbook of
Statistics 14, Amsterdam: Elsevier, pp. 35-60
Harvey, c.R., B. Solnik, and G. Zhou, 1995, What determines expected interna-
tional asset returns?, working paper, Duke university
Hawawini, G., and D. Keim, 1995, On the Predictability of Common Stock Re-
turns: World-Wide Evidence, in: R. Jarrow, et.al., Handbooks in OR and MS
(Vol. 9, 'Finance'), Amsterdam: Elsevier, pp. 497-544
Hazuka, T., 1984, Consumption betas and backwardation in commodity mar-
kets, Journal of Finance 39, pp. 647-766
324 Bibliography
Heaton, J., and D. Lucas, 1996, Evaluating the Effects of Incomplete Markets on
Risk Sharing and Asset Pricing, Journal of Political Economy 104, pp. 668-712
Henry, B., 2000, Stock Market Liberalization, Economic Reform, and Emerging
Equity Prices, Journal of Finance 55, pp. 529-564
Heston, S., and G. Rouwenhorst, G., 1994, Does industrial structure explain the
benefits of international diversification?, Journal of Financial Economics 36,
pp.3-27
Heston, S., and G. Rouwenhorst, G., 1995, Industry and country effects in inter-
national stock returns, Journal of Portfolio Management 22, pp. 53-58
Heston, S., G. Rouwenhorst, and R Wessels, 1995, The structure of interna-
tional stock returns and the integration of capital markets, Journal of Empirical
Finance 2, pp. 173-197
Heston, S., G. Rouwenhorst, and R. Wessels, 1999, The role of size and beta in
the cross-section of European stock returns, European Financial Management 4,
pp.4-28
Hodrick, RJ., and E.C. Prescott, 1997, Postwar u.S. Business Cycles: An Em-
pirical Investigation, Journal of Money, Credit, and Banking 29, pp. 1-16
Hodrick R.J., D. Ng, and P. Sengmueller, 1999, An International Dynamic Asset
Pricing Model, NBER working paper 7157
Huang, c., and R Litzenberger, 1988, Foundations for Financial Economics,
New York: North Holland
Huberman, G., and S. Kandel, 1987a, Mean-Variance Spanning, Journal of Fi-
nance 42, pp. 873-888
Huberman, G., and S. Kandel, 1987b, Mimicking portfolios and exact arbitrage,
Journal of Finance 42, pp. 1-10
Ikeda, S., 1991, Arbitrage asset pricing under exchange risk, Journal of Finance
46,pp.447-466
Ilmanen, A., 1995, Time-Varying Expected Returns in International Bond Mar-
kets, The Journal of Finance 50, pp. 481-506
Ingersoll, J.E., 1984, Some results of the theory of arbitrage pricing, Journal of
Finance 39, pp. 1021-1039
Ingersoll, J.E., 1987, Theory of Financial Decision Markets, Savage: Rowman &
Littlefield
Jagadeesh, N., 1990, Evidence of Predictable Behavior of security Returns,
Journal of Finance 45, pp. 881-898
Jagannathan, R, and Z. Wang, 1996, The conditional CAPM and the cross-
section of expected returns, Journal of Finance 51, pp. 3-53
325
Jorion, P., 1992, Term Premiums and the Integration of the Eurocurrency Mar-
kets, Journal of International Money and Finance, pp. 17-39.
Jorion, P., 1994, MeanlVariance Analysis of Currency Overlays, Financial
Analysts Journal, pp. 48-56
Jorion, P., and J. Khoury, 1996, Financial Risk Management, Cambridge: Basil
Blackwell
Jorion, P., and E. Schwartz, 1986, Integration vs. Segmentation in the Canadian
Stock Market, Journal of Finance 41, pp. 603-617
Kandel, S., and R.F. Stambaugh, 1990, Expectations and volatility of consump-
tion and asset returns, Review of Financial Studies 3, pp. 207-232
Kandel, S., and RF. Stambaugh, 1991, Asset Returns and Intertemporal Prefer-
ences, Journal of Monetary Economics 27, pp. 39-71
Kandel, S., and RF. Stambaugh, 1995, Portfolio Inefficiency and the Cross-
Section of Expected Returns, Journal of Finance 50, pp. 157-184
Kaplanis, E., 1988, Stability and Forecasting of the Comovement Measures of
International Stock Market Returns, Journal of International Money and Fi-
nance 7, pp. 63-76
Kasa, K., 1997, Consumption-based versus production-based models of interna-
tional equity markets, Journal of International Money and Finance 16, pp. 653-
680
Kaul, G., 1987, Stock returns and inflation: The role of the monetary sector,
Journal of Financial Economics 17, pp. 253-276
Keirn, S., and RF. Stambaugh, 1986, Predicting returns in the stock and bond
markets, Journal of Financial Economics 17, pp. 358-390
Keirn, S., and RF. Stambaugh, 1987, On correlations and interferences about
mean-variance efficiency, Journal of Financial Economics 18, pp. 61-90
Kim, E., and V. Singal, 2000, Stock Market Openings: Experience of Emerging
Economies, The Journal of Business 73, pp. 25-66
Kimball, M., 1990, Precautionary Savings in the Small and in the Large,
Econometrica 58, pp. 53-73
King, M., and S. Wadhwani, 1990, Transmission of Volatility between Stock
Markets, Review of Financial Studies 3, pp. 5-33
Kirby, C., 1998, The Restrictions on Predictability Implied by Rational Asset
Pricing Models, Review of Financial Studies 11, pp. 343-382
Koch, P., and T. Koch, 1991, Evolution in dynamic linkages across daily na-
tional stock indexes, Journal of International Money and Finance 10, pp. 231-
251
326 Bibliography
Lo, AW., and A.C. MacKinlay, 1992, Maximizing Preditctability in the Stock
and Bond Markets, working pape, Sloan School of Management, Massachusetts
Institute of Technology
Lo, AW., and AC. MacKinlay, 1999, A Non-Random Walk Down Wall Street,
Princeton: Princeton University Press
Lombard, T., J. Roulet, and B. Solnik, 1999, Pricing of Domestic Versus Multi-
national Companies, Financial Analysts Journal 55, pp. 35-49
Long, J., 1974, Stock prices, inflation, and the term structure of interest rates,
Journal of Financial Economics 2, pp. 313-170
Longin, F., and B. Solnik, 1995, Is the correlation in international equity returns
constant: 1960-1990?, Journal of International Money and Finance 12, pp. 3-26
Lucas, RE. Jr., 1978, Asset prices in an exchange economy, Econometrica 46,
pp. 1429-1445
Lucas, R.E. Jr., 1987, Models of business cycles, Oxford: Basil Blackwell
Mankiw, N.G., and M.D. Shapiro, 1986, Risk and Return: consumption versus
market beta, Review of Economics and Statistics 68, pp. 452-459
Mankiw, N.G., and S.P. Zeldes, 1991, The consumption of stockholders and
nonstockholders, Journal of Financial Economics 29, pp. 97-112
Markowitz, H., 1959, Portfolio Selection, New York: John Wiley & Sons
McCulloch, H., 1975, The Monte Carlo Cycle in Business Activity, Journal of
Political Economy 91, pp. 39-69
McCurdy, T., and I. Morgan, 1992, Evidence of risk premiums in foreign cur-
rency markets, Review of Financial Studies 5, pp. 65-84
McQueen, G., and V.V. Roley, 1993, Stock Prices, News, and Business Condi-
tions, Review of Financial Studies 6, pp. 683-707
Mehra, R., and E. Prescott, 1980, Recursive competitive equilibrium: The case
of homogeneous households, Econometrica 48, pp. 1365-1379
Mehra, R., and E. Prescott, 1985, The Equity Premium: A Puzzle, Journal of
Monetary Economics 15, pp. 145-161
Mehra, R., and E. Prescott, 1988, The Equity Risk Premium: A Solution?, Jour-
nal of Monetary Economics 22, pp. 133-136
Merton, R, 1971, Optimum Consumption and Portfolio Rules in a Continuous-
Time Model, Journal of Economic Theory 3, pp. 373-413
Merton, R, 1973, An intertemporal capital asset pricing model, Econometrica
42,pp.867-887
328 Bibliography
Merton, R.c., 1980, On Estimating the Expected Return on the Market, Journal
of Financial Economics 8, pp. 323-361
Merton, R.C., 1987, On the Current State of the Stock Market Rationality Hy-
pothesis, in: Dornbusch, R., S. Fischer and J. Bossons (eds.), Macroeconomics
and Finance: Essays in Honor of Franco Modigliani, Cambridge: MIT Press
Modigliani, F., and Brumberg, 1954, Utility analysis and the consumption func-
tion: an interpretation of cross-section data, in: K. Kurihara (ed.), Post Keynes-
ian economics, New Brunswick: Rutgers University Press, pp. 388-436
Mossin, J., 1966, Equilibrium in a Capital Asset Pricing Model, Econometrica
35, pp. 768-783
Nelson, c., and C. Plosser, 1982, Trends and Random Walks in Macroeconomic
Time Series: Some Evidence and Implications, Journal of Monetary Economics
10, pp. 139-162
Newey, W., and K. West, 1987, A Simple Positive Semi-Definite, Heterosce-
dasticity and Autocorrelation Consistent Covariance Matrix, Econometrica 55,
pp.703-708
Ng, L., 1991, Test of the CAPM with Time-Varying Covariances: A Multivari-
ate GARCH Approach, Journal of Finance 46, pp. 1507-1521
Obstfeld, M., 1989, How integrated are world capital markets? Some new tests,
in: Calvo, G., R. Findlay, P. Kouri, and J. de Macedo (eds.), Debt stabilization
and development: Essays in memory of Carlos Diaz-Alejandro, Oxford: Basil
Blackwell
Obstfeld, M., 1994, Risk-taking, global diversification and growth, American
Economic Review 85, pp. 1310-1329
Obstfeld, M., and K. Rogoff, 1997, Foundations of International Macroeco-
nomics, Cambridge: MIT Press
Odier, P., and B. Solnik, 1993, Lessons for International Asset Allocation, Fi-
nancial Analysts Journal, pp. 63-77
Oertmann, P., 1997, Global Risk Premia on International Investments, Gabler
Verlag
Oertmann, P., and H. Zimmermann, 1997, Global Economic Risk Profile,
Schweizer Bank, Heft 4/1998, pp. 24-27
Oertmann, P., and H. Zimmermann, 1998, Global economic conditions and risk
premia on international investments, working paper, University of St. Gallen
Ostdiek, B., 1998, The world ex ante risk premium: an empirical investigation,
Journal of International Money and Finance 17, pp. 967-999
329
Perold, A., and E. Schulman, 1988, The Free Lunch in Currency Hedging; Im-
plications for Investment Policy and Performance Standards, Financial Analysts
Journal, pp. 45-50
Poterba, J., and L. Summers, 1988, Mean Reversion in Stock Returns: Evidence
and Implications, Journal of Financial Economics 22, pp. 27-60
Reisman, H., 1992, Reference variables, factor structure, and the approximate
mutltibeta representation, Journal of Finance 47, pp. 1303-1314
Restoy, F., and P. Wei1, 1998, Approximate Equilibrium Asset Pricing, NBER
working paper 6611
Rietz, T.A., 1988, The equity risk premium: A solution, Journal of Monetary
Economics 22, pp. 117-131
Roll, R., 1977, A Critique of the Asset Pricing Theory's Tests: Part I, Journal of
Financial Economics 4, pp. 129-176
Roll, R., 1992, Industrial Structure and The Comparative Behavior of Interna-
tional Stock Market Indices, Journal of Finance 47, pp. 3-41
Roll, R., and S. Ross, 1980, An Empirical Investigation of the Arbitrage Pricing
Theory, Journal of Finance 35, pp. 1073-1103
Roll, R., and S. Ross, 1994, On the Cross-Sectional Relation between Expected
Returns and Betas, Journal of Finance 49, pp. 101-122
Ross. S., 1976, The Arbitrage Theory of Capital Asset Pricing, Journal of Eco-
nomic Theory 13, pp. 341-360
Ross, S., 1978, A Simple Approach to the Valuation of Risky Securities, Jour-
nal of Business 51, pp. 453-460
Rouwenhorst, G., 1995, Asset pricing implications of equilibrium business cycle
models, in: Cooley, T. (ed.), Frontiers of Business Cycle Research, Princeton:
Princeton University Press
Rozeff, M., 1984, Dividend yields and equity risk premiums, Journal of Portfo-
lio Management 11, pp. 68-75
Rubinstein, M., 1974, An aggregation theorem for securities markets, Journal of
Financial Economics 1, pp. 225-244
Rudolf, M., and H. Zimmermann, (1998), An Algorithm for International Port-
folio Selection and Optimal Currency Hedging, in: Mulvey, J., and W. Ziemba
(eds.), Worldwide Asset and Liability Modeling, Cambridge: Cambridge Univer-
sity Press, pp. 315-240
Salyer, K., 1998, Crash states and the equity premium: Solving the puzzle raises
another, Journal of Economic Dynamics and Control 22, pp. 955-965
330 Bibliography
Samuelson, P., 1965, Proof that Properly Anticipated Prices Auctuate Ran-
domly, Industrial Management Review 6, pp. 41-49
Samuelson, P., 1969, Lifetime Portfolio Selection by Dynamic Stochastic Pro-
gramming, Review of Economics and Statistics 51, pp. 239-246
Schwert, W., 1990, Stock return and Real Activity: A Century of Evidence,
Journal of Finance 45, pp. 1237-1258
Selden, L., 1978, A New Representation of Preferences Over 'Certain X Uncer-
tain' Consumption Pairs: The 'Ordinal Certainty Equivalent' Hypothesis,
Econometrica 46, pp. 1045-1060
Sercu, P., 1980, A generalization of the international asset pricing model, Revue
de I 'Association Francaise de Finance, pp. 91-l35
Shanken, J., 1985, A multi-beta CAPM or equilibrium APT?: A reply, Journal
of Finance 40, pp. 1189-1196
Shanken, 1., 1987, Multivariate proxies and asset pricing relations: Living with
the Roll critique, Journal of Financial Economics 18, pp. 91-110
Shanken, J., 1990, Intertemporal asset pricing: An empirical investigation, Jour-
nal of Econometrics 45, pp. 99-120
Shanken, J., 1992, On the Estimation of Beta-Pricing Models, Review of Finan-
cial Studies 5, pp. 1-34
Sharpe, W., 1964, Capital asset prices: A theory of market equilibrium under
conditions of risk, Journal of Finance 19, pp. 425-442
Shiller, R., 1979, The volatility of long-term interest rates and expectations
models of the term structure, Journal of Political Economy 87, pp. 1190-1219
Shiller, R., 1981, Do Stock Prices Move Too much to be Justified by Subse-
quent Changes in Dividends?, American Economic Review 71, pp. 421-436
Shleifer, A., and R. Vishny, 1986, The stock market as a cause of investment,
working paper, Massachusetts Institute of Technology
Snow, K., 1991, Diagnosing Asset Pricing Models using the Distribution of As-
set Returns, Journal of Finance 46, pp. 955-983
Solnik, B., 1974a, Why not diversify Internationally rather than Domestically?,
Financial Analysts Journal, pp. 48-54
Solnik, B., 1974b, An Equilibrium Model of the International Capital Markets,
Journal of Economic Theory 8, pp. 500-524
Solnik, B., 1974c, The International Pricing of Risk: An Empirical Investigation
of the World Capital Structure, Journal of Finance 29, pp. 365-378
331
Solnik, B., 1983a, International arbitrage pricing theory, Journal of Finance 38,
pp.449-457
Solnik, B., 1983b, The relationship between stock returns and inflationary ex-
pectations: the international evidence, Journal of Finance 38, pp. 35-48
Solnik, B., 1993a, Currency Hedging and Siegel's Paradox: On Black's Univer-
sal Hedging Rule, Review of International Economics, pp. 180-187
Solnik, B., 1993b, The performance of international asset allocation strategies
using conditioning information, Journal of Empirical Finance 1, pp. 33-47
Solnik, B., 1998, Global Asset Management: To hedge or not to hedge - a ques-
tion that cannot be ignored, Journal of Portfolio Management 25, pp. 43-51
Solnik, B., 1999, International Investments, Reading: Addison Wesley, MA
Speidell, L.S., and R. Sappenfield, 1992, Global Diversification in a Shrinking
World, Journal of Portfolio Management 19, pp. 57-67
Stambaugh, R., 1982, On the Exclusion of Assets from Tests of the Two Pa-
rameter Model, Journal of Financial Economics 10, pp. 235-268
Stehle, R., 1977, An Empirical Test of the Alternative Hypothesis of National
and International Pricing of Risky Assets, Journal of Finance 32, pp. 493-502
Stock, J., and M. Watson, 1989, New Indexes of Coincident and Leading Eco-
nomic Indicators, NBER Macroeconomics Annual, pp. 351-394
Stulz, R., 1981, A model of international asset pricing, Journal of Finance 36,
pp.923-934
Stulz, R., 1984, Pricing capital assets in an international setting: An introduc-
tion, Journal of International Business Studies, pp. 55-72
Stulz, R., 1995, International Portfolio Choice and Asset Pricing; An Integrative
Survey, in: in: R. Jarrow, et.al., Handbooks in OR and MS (Vol. 9, 'Finance'),
Amsterdam: Elsevier, pp. 201-223
Stulz, R., 1999, Globalization of Equity Markets and the Cost of Capital, NBER
working paper 7021
Tauchen, G., and R. Hussey, 1991, Quadrature-based methods for obtaining ap-
proximate solutions to non-linear asset pricing models, Econometrica 59, pp.
371-396
Telmer, c.1., 1993, Asset pricing puzzles and incomplete markets, Journal of
Finance 48, pp. 1803-1832
Tesar, L., 1995, Evaluating the gains from international risk sharing, Carnegie-
Rochester Conference Series on Public Policy 42, pp. 95-143
332 Bibliography
Tesar, L., and I. Werner, 1995, Home bias and high turnover, Journal of Inter-
national Money and Finance 14, pp. 46-492
Vassalou, M., 2000, Exchange Rate and Foreign Inflation Risk Premiums in
Global Equity Returns, forthcoming in: Journal of International Money and Fi-
nance
Von Wincoop, E., 1994, Welfare gains from international risk sharing, Journal
of Monetary Economics 34, pp. 175-200
Weil, P., 1989, The equity premium puzzle and the risk-free rate puzzle, Journal
of Monetary Economics 24, pp. 401-421
Wei!, P., 1992, Equilibrium Asset Prices with Undiversifiable Labor Income
Risk, Journal of Economic Dynamics and Control 16, pp. 769-790
Wheatley, S., 1989, A Critique of Latent Variable Tests Of Asset Pricing Mod-
els, Journal of Financial Economics 23, pp. 325-338
Wheatley, S., 1988a, Some test on international equity integration, Journal of
Financial Economics 21, pp. 177-212
Wheatley, S., 1988b, Some tests of the consumption-based asset pricing model,
Journal of Monetary Economics 22, pp. 193-215
Whitelaw, R., 1997, Stock Market Risk and Return: An Equilibrium Approach,
working paper, Stern School of Business, New York University
Working, H., 1960, Note on the Correlation of First Differences of Averages in
a Random Chain, Econometrica 28, pp. 916-918
Zhou, G., 1994, Analytical GMM Tests: Asset Pricing with Time-Varying Risk
Premiums, Review of Financial Studies 7, pp. 687-710
Zimmermann, H., 1996, Finanzanalyse und Kapitalmarkttheorie am Beispiel
schweizerischer Wirtschaftssektoren, Finanzmarkt und Portfoliomanagement
10, pp. 148-171
Zimmermann, H., 1998, State-Preference Theorie und Asset Pricing: Eine Ein-
fiihrung, Heidelberg: Physica-Verlag
Zimmermann, H., 1999, State-Preference Primer, manuscript, University of St.
Gallen