ANNUAL

REVIEWS Further Asset Allocation

Click here for quick links to
Annual Reviews content online,
including: Jessica A. Wachter
t
Other articles in this volume Department of Finance, The Wharton School, University of Pennsylvania,
t Top cited articles Philadelphia, Pennsylvania 19104; email: jwachter@wharton.upenn.edu.
t Top downloaded articles
t
0VS
DPNQSFIFOTJWF
TFBSDI

Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org
by University of Pennsylvania on 11/26/10. For personal use only.

Annu. Rev. Financ. Econ. 2010. 2:175–206 Key Words

First published online as a Review in Advance on portfolio choice, predictive regression, recursive utility
September 22, 2010

The Annual Review of Financial Economics is Abstract

online at financial.annualreviews.org
This review article describes recent literature on asset allocation,
This article’s doi: covering both static and dynamic models. The article focuses on the
10.1146/annurev-financial-073009-104026
bond-stock decision and on the implications of return predictability.
Copyright © 2010 by Annual Reviews. In the static setting, investors are assumed to be Bayesian, and the
All rights reserved
role of various prior beliefs and specifications of the likelihood are
1941-1367/10/1205-0175$20.00 explored. In the dynamic setting, recursive utility is assumed, and
attention is paid to obtaining analytical results when possible.
Results under both full- and limited-information assumptions are
discussed.

175
 1. INTRODUCTION
The study of portfolio allocation has played a central role in financial economics, from its
very beginnings as a discipline. This field of study has attracted (and continues to attract)
the attention that it does because it is both highly practical and amenable to the application
of sophisticated mathematics.
This study reviews the recent academic literature on asset allocation. Two important
simplifications are employed: First, the field has drawn a distinction between the study
of allocation to broad asset classes and allocation to individual assets within a class.
This article focuses on the former. In fact, the empirical applications in this article
assume an even more specific case, namely an investor who chooses between a broad
stock portfolio and a riskless asset. Second, the surveyed models assume, for the most
part, no financial frictions. That is, I assume that the investor does not face unhedgeable
labor income risk or barriers to trading in the assets, such as leverage or short-sale
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

constraints. This is not to deny the importance of other asset classes or of financial
frictions. Recent surveys on portfolio choice (encompassing portfolios of many assets)
by University of Pennsylvania on 11/26/10. For personal use only.

include Cochrane (1999), Brandt (2009), and Avramov & Zhou (2010). Campbell
(2006) and Curcuru et al. (2009) survey work on asset allocation under realistic frictions
faced by households.
I focus on two broad classes of models: static models (in which the investor looks one
period ahead) and dynamic models (in which the investor looks multiple periods ahead and
takes his future behavior into account when making decisions). For static models, the
solution where investors have full information about asset returns has been known for
some time (Markowitz 1952), so the focus is on incorporating uncertainty about the return
process. In contrast, much has been learned in recent years about dynamic models, even in
the full-information case. A barrier to considering dynamic models is often their complex-
ity: For this reason, I devote space to analytical results. These results, besides being inter-
esting in their own right, can serve as a starting point for understanding the behavior of
models that can be solved numerically only.
Finally, in both the static and dynamic sections, I consider in detail the model in which
excess returns on stocks over short-term Treasury bills are in part predictable. A substantial
empirical literature devotes itself to the question of whether returns are predictable; the
asset allocation consequences of such predictability are striking and well-known in at least
a qualitative sense since Graham & Dodd (1934).
Ultimately, the goal of academic work on asset allocation is the conversion of the time
series of observable returns and other variables of interest into a single number: Given the
preferences and horizon of the investor, what fraction of her wealth should she put in
stock? The aim is to answer this question in a “scientific” way, namely by clearly specifying
the assumptions underlying the method and developing a consistent theory based on these
assumptions. The very specificity of the assumptions and the resulting advice can seem
dangerous, imputing more certainty to the models than the researcher can possibly possess.
Yet, only by being so highly specific, does the theory turn into something that can be clearly
debated and ultimately refuted in favor of an equally specific and hopefully better theory.
This development implies the use of mathematics to model the investment decision.
Throughout this article, the reader is encouraged to remember that the subject of the
modeling is an individual or household making a decision with significant consequences
for lifelong financial security.

176 Wachter
 2. STATIC MODELS

2.1. The Basic Decision Problem

In this section, I consider the problem of an investor maximizing wealth as of time
T by allocating wealth between a risky and a riskless asset. The portfolio decision takes
^ < T. Let Rtþ1 denote the simple return on the risky asset between times
place at a time T
t and t þ 1 and Rf,tþ1 is the simple return on the riskless asset between t and t þ 1. Let Wt
denote the investor’s wealth at time t. The investor solves
" #
WT1"g
max E T^ , ð1Þ
z 1"g

where
0 1
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

Y
T Y
T
WT ¼ WT^@z Rs þ (1 " z) Rf ,s A: ð2Þ
^þ1
s¼T ^þ1
s¼ T
by University of Pennsylvania on 11/26/10. For personal use only.

The parameter g is assumed to be positive, and g ¼ 1 should be interpreted as logarith-

^ and
mic utility. Note that the investor described above decides on the allocation z at time T
then does not trade. This is a buy-and-hold investor. The implicit weight on the risky
security can, and almost certainly will, change over time; however, the problem written as
above assumes that the investor does not rebalance back to the original weights. For now,
it is assumed that z can take on any value: Short sales and borrowing at the risk-free rate
are allowed. For the purposes of solving the model, I assume Rfs, for s between T^ þ 1 and T,
is known to the investor at time T ^. Following much of the literature, I assume that the
investor’s utility takes a power form, implying that relative risk aversion is constant and
that asset allocation does not depend on wealth. The scale invariance of power utility has
broad empirical support in that interest rates have remained stationary despite the fact that
wealth has grown.
Define the continuously compounded excess return to be

yt ¼ log Rt " log Rft

and assume yt follows the process

ytþ1 ¼ a þ bxt þ utþ1 , ð3Þ

and

xtþ1 ¼ y þ rxt þ vtþ1 , ð4Þ

where
! "
utþ1
j yt , . . . , y1 , xt , . . . , x0 & N(0, S), ð5Þ
vtþ1

and
" #
s2u suv
S¼ : ð6Þ
suv s2v

www.annualreviews.org ' Asset Allocation 177

 That is, ytþ1 has a predictable component xt that follows a first-order autoregressive
process. The errors are assumed to be serially uncorrelated, homoskedastic, and jointly
normally distributed. A substantial and long-standing empirical literature documents pre-
dictability in excess returns, in the sense that running regression (Equation 3) for observ-
able xt generates statistically significant coefficients: For example, see Fama & Schwert
(1977), Keim & Stambaugh (1986), Campbell & Shiller (1988), Fama & French (1989),
Cochrane (1992), Goetzmann & Jorion (1993), Hodrick (1992), Kothari & Shanken
(1997), Lettau & Ludvigson (2001), Lewellen (2004), Ang & Bekaert (2007), Boudoukh
et al. (2007). In what follows, I focus on the case where xt is the dividend yield because the
dividend yield and future expected excess returns are linked through a present value
identity (Campbell & Shiller 1988). Theory suggests, therefore, that if returns are pre-
dictable, the dividend yield should capture at least some of that predictability. The general
setting that I have laid out here largely follows that of Barberis (2000) but with some
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

important differences.
I assume the investor does not know the parameters of the system above. Rather, he is a
Bayesian, meaning that he has prior beliefs on the parameters, and after viewing the data,
by University of Pennsylvania on 11/26/10. For personal use only.

makes inferences using the laws of probability (Berger 1985). Let b^ be the ordinary least
squares (OLS) estimate of b in the regression (Equation 3). Bayesian analysis turns the
standard frequentist analysis on its head: Instead of asking for the distribution of the test
statistic b^ (which depends on the data) as a function of the true parameter b, Bayesian
analysis asks for the distribution of the true parameter b as a function of the data (which
^
often comes down to a function of sufficient statistics, such as b).
For notational convenience, stack the coefficients from Equations 3 and 4 into a vector:

b ¼ ½a, b, y, r)> :

The investor starts out with prior beliefs p(b, S). Let L(Djb, S) denote the likelihood
function, where D is the data available up until and including time T ^. It follows from
Bayes’ rule that the posterior distribution is given by
L(D j b, S) p(b, S)
p(b, S j D) ¼ ,
p(D)
where p(D) is an unconditional likelihood of the data in the sense that
Z
p(D) ¼ L(D j b, S) p(b, S) dbdS,
b,S

namely that p(D) integrates out b and S. It follows that

p(b, S j D) / L(D j b, S) p(b, S), ð7Þ

where / denotes “proportional to” because p(D) does not depend on b or S. The likeli-
^ years of data, is equal to
hood function, given T
^Y
T "1
L(D j S, b) ¼ ptþ1jt (ytþ1 , xtþ1 j xt , S, b)p0 (x0 j b, S):, ð8Þ
t¼0

where ptþ1jt (ytþ1 , xtþ1 j xt , S, b) is given by a bivariate normal density function as described
in Equations 3–6 and p0 (x0 j b, S) gives the initial condition of the time series. Given the
posterior, the predictive density for returns from time T ^ to T is defined as
178 Wachter
 Z
p(yT^þ1 , . . . yT j D) ¼ p(yT^þ1 , . . . , yT j b, S, xT^) p(b, S j D) db dS: ð9Þ

The predictive distribution (Equation 9) summarizes the agent’s beliefs about the return
distribution after viewing the data. The expectation in Equation 1 is taken with respect to
this distribution.
How might predictability influence an investor’s optimal allocation? Kandel & Stambaugh
(1996) find that optimal allocation for the single-period case can be approximated by

1 E½yT ) þ 12 Var(yT )
z* , ð10Þ
g Var(yT )

where the mean and the variance are taken under the investor’s subjective distribution of
returns (which is Equation 9 in the Bayesian case). Holding the variance constant, an upward
shift in the mean increases the allocation. This is not surprising given that the investor prefers
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

more, not less, wealth. This approximation is valid only for short horizons and small shocks.
However, it is useful as a first step to understanding the portfolio allocation.
by University of Pennsylvania on 11/26/10. For personal use only.

2.2. The Conditional Bayesian Model

The initial condition p0 (x0 j b, S) in Equation 8 is problematic. Kandel & Stambaugh
(1996) show that if this term were to disappear, the system (Equations 3–6) would take
the form of a classical multivariate regression model (Zellner 1971, ch. 8).1 Assuming the
standard noninformative prior for regression,

p(b, S) /j S j"3=2 , ð11Þ

the posterior distribution for all the parameters could then be obtained in closed form.
Indeed, conditional on S, b would be normally distributed around its OLS estimate b. ^
However, p0 (x0 j b, S) is there, and something must be done about it. One approach is to
let it stay and specify what it should be. I refer to the resulting set of assumptions and
results as the “exact Bayesian model” (described Section 2.3). Another approach is to
assume x0 conveys no prior information. Thus,
p(b, S j x0 ) ¼ p(b, S): ð12Þ

Because the prior is now conditional on x0, the likelihood can condition on x0 as well.
The posterior is, of course, conditional on x0 because it is conditional on all the data. That
is, the assumption in Equation 12 allows Equation 7 to be replaced by
p(b, S j D) / Lc (D j b, S, x0 ) p(b, S j x0 ),

where Lc is the likelihood conditional on x0:

Y
T "1
Lc (D j S, b, x0 ) ¼ ptþ1jt (ytþ1 , xtþ1 j xt , S, b): ð13Þ
t¼0

1
However, it still would not be a classical regression model. An assumption of classical regression is that the
dependent variable is either nonstochastic or independent of the disturbance term ut at all leads and lags (Zellner
1971, ch. 3). As emphasized in Stambaugh (1999), the independence assumption fails in predictive regressions.
Under the assumptions of classical regression, Gelman et al. (1996) show that the likelihood function for the
regressor is irrelevant to the agent’s decision problem (and so, therefore, is the initial condition).

www.annualreviews.org ' Asset Allocation 179

 I refer to this as the “conditional Bayesian model.” I comment further on the assumption in
Equation 12 below.
This conditional Bayesian model is used by Barberis (2000) to study asset allocation
for buy-and-hold investors. For comparison, Barberis also solves the model in the full-
information case, namely when the investor knows the parameters in Equations 3–6.
Figure 1 shows the resulting optimal allocation for a risk aversion of 5 at various horizons
and values of the dividend yield: the full-information case (Figure 1a) and the results from
the conditional Bayesian model (Figure 1b).
A striking feature of Figure 1b is the degree to which portfolio weights respond to
changes in the dividend yield. That is, the investor aggressively engages in market timing,
with the dividend yield as the signal of how much to allocate to equities. For an investor
with a one-year horizon, the optimal allocation is approximately 80% when the dividend
yield is at its mean. When the dividend yield is at one or two standard deviations above
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

the long-run mean, the investor has all her wealth in stocks. When the dividend yield is at
one standard deviation below the long-run mean, the optimal allocation falls to 20%. The
optimal allocation is bounded above and below because the power utility investor would
by University of Pennsylvania on 11/26/10. For personal use only.

never risk wealth below zero. Because the distribution for returns is unbounded from
above, this investor would never hold a negative position in stock. The investor would also

a b
1 1
Allocation to stocks

0.5 0.5

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Horizon (years) Horizon (years)

Figure 1
Static allocation as a function of horizon assuming return predictability: (a) when there is no parameter uncertainty and
(b) incorporating parameter uncertainty. The solid line corresponds to the optimal (buy-and-hold) allocation when the dividend
yield is at its sample mean (3.75%). The dash-dotted lines correspond to the allocations when the dividend yield is one standard
deviation above or below its mean (2.91% and 4.59%, respectively). The dotted lines correspond to the allocations when the
dividend yield is two standard deviations above or below its mean (2.06% and 5.43%, respectively). The agent has power utility
over terminal wealth with relative risk aversion equal to 5. Some lines may lie on top of each other. The allocations weakly
increase as a function of the dividend yield except at very long horizons in panel b. The model is estimated over monthly data
from 1952 to 1995. This figure is adapted from Barberis (2000, figure 3).

180 Wachter
 never hold a levered position in stock, for this too implies a positive probability of negative
wealth, because returns could be as low as "100%. Note that these endogenous bounds
on the optimal portfolio illustrate errors in the approximation given in Equation 10, which
contains no such bounds.
What about the case where the investor incorporates estimation risk into her decision
making? One may think that estimation risk would make a substantial difference because,
as Barberis (2000) reports, the evidence for predictability at a monthly horizon is of only
borderline significance in the relevant sample. However, the results incorporating estima-
tion risk are quite similar to those that do not at short horizons. Indeed, differences start to
become noticeable only at buy-and-hold horizons of five years or more. (Although such
long buy-and-hold horizons may characterize the behavior of some investors, from the
normative perspective of this article, such infrequent trading seems extreme.) Thus, for the
statistical model for stock returns above, parameter uncertainty resulting from the regres-
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

sion is small compared with the measured uncertainty of holding stocks.2

Figure 1 also reveals that the optimal allocation is increasing in the horizon in the case
of full information as well as for all but the longest horizons in the parameter uncertainty
by University of Pennsylvania on 11/26/10. For personal use only.

case [see Stambaugh (1999) for an explanation of the reversed relation between holdings
and the dividend yield at the longest horizons]. Because innovations to the dividend yield
are negatively correlated with innovations to returns, stocks when measured at long hori-
zons are less risky than stocks measured at short horizons [mean reversion in stock returns
is pointed out in earlier work of Poterba & Summers (1988)]. The implications of mean-
reversion for long-horizon investors are also the subject of Siegel (1994).

2.3. The Exact Bayesian Model

The results above implicitly assume Equation 12, namely that prior beliefs are independent
of the initial observation x0. This assumption enables the use of the conditional likelihood,
which combined with the prior (Equation 11) leads to closed-form expressions for the
posterior distributions of the parameters. Although Equation 12 is convenient, how realis-
tic is it? Under Equation 12, the agent believes that the value of x0 conveys no information
about the parameters of the process for x or y. For instance, the initial value of the dividend
yield would tell you nothing about, say, the average dividend yield. There is nothing
mathematically incorrect about specifying such prior beliefs; the agent can, in principal,
believe anything so long as it does not entail a logical inconsistency or require a peek ahead
at the data. However, such beliefs do not seem reasonable. For instance, the logic of these
beliefs would allow the agent to exclude an arbitrary amount of the data from consider-
ation, just by making the prior parameters independent of these data.
The question of whether to allow the initial condition appears to be of a technical
nature, but it has unexpectedly deep implications for Bayesian estimation and for the
portfolio-allocation decision. Stambaugh (1999) describes these implications. Stambaugh
also shows that OLS estimation of the coefficient b implies results are upward biased.

2
Note that the effect of the dividend yield attenuates at longer horizons both when parameter uncertainty is taken
into account and when it is not. This occurs because the dividend yield is mean-reverting and because the investor
cannot rebalance as the dividend yield reverts to its mean. In the limit, an investor with an infinite horizon (who cares
about wealth at the end of the horizon) would care only about the unconditional distribution of returns and not
about the current value of the dividend yield. Because the dividend yield is so persistent, the effect attenuates very
slowly as a function of the horizon.

www.annualreviews.org ' Asset Allocation 181

 As mentioned above, the posterior mean of b implied by the conditional Bayesian
model is also given by the OLS estimate b^ and thus is also biased, despite the fact that
Bayesian estimation explicitly takes the finite-sample properties of the regression into
account. The bias in the posterior mean of b may be an indication that all is not right
with this model.
Without Equation 12, the conditional likelihood (Equation 13) is no longer correct and
the so-called exact likelihood (Equation 8) must be used. Immediately, a decision must be
made about the distribution of the initial observation x0. Stambaugh (1999) assumes that
x0 is drawn from the stationary distribution of Equation 4. If r is between "1 and 1, this
stationary distribution exists and is given by
# $
y s2v
x0 & N , ð14Þ
1 " r 1 " r2
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

(Hamilton 1994, p. 53). The relevant likelihood function is therefore Equation 8, where p0
is the normal density given by Equation 14.
by University of Pennsylvania on 11/26/10. For personal use only.

The use of the unconditional likelihood requires that r be between "1 and 1.
Stambaugh (1999) therefore modifies the assumption in Equation 11 as follows:

p(b, S) /j S j"3=2 , r 2 ("1, 1): ð15Þ

Stambaugh also considers the alternative prior specification:

p(b, S) / (1 " r2 )"1 j S j "5=2 , r 2 ("1, 1): ð16Þ

What is the rationale for Equation 16 or, for that matter, for Equations 15 or 11? The
prior of Equation 11 is standard in regression models. Its appeal is best understood by the
fact that it embodies three conditions: (a) b and S should be independent in the prior;(b)
for the elements of b, ignorance is best represented by a uniform distribution (which, in the
limit, becomes a constant as in Equation 11); and (c)

p(S) /j S j"3=2 , ð17Þ

which generalizes the assumption that, for a single system, the log of the standard deviation
should have a flat distribution on "1 and 1. Jeffreys (1961, p. 48) proposes these rules
for cases when there is no theoretical guidance on the values of the parameters. An
additional appeal of Equation 11 (discussed above) is that, when combined with the
likelihood (Equation 13), explicit expressions for the posterior distributions of the param-
eters can be obtained.
This discussion would seem to favor the prior given in Equation 15 (because theory now
requires a stationary process) in combination with the exact likelihood. However, applying
these rules does not constitute the only approach. Jeffreys (1961) proposes an alternative
means of defining ignorance: Inference should be invariant to one-to-one changes in the
parameter space. This criterion is appealing in the case of the predictive model (Equations
3–6) in which the particular parametrization appears arbitrary. The exact form of Jeffreys
prior depends on the sample size T and is derived by Uhlig (1994). Stambaugh (1999)
derives an approximate Jeffreys prior that becomes exact as the sample size approaches
infinity. This approximate Jeffreys prior is equal to Equation 16. Relative to the flat prior
for r (Equation 15), more weight is placed on values of r close to "1 and 1.

182 Wachter
 Table 1 Posterior means of b and r under various combinations of the likelihood and the
prior*

Posterior means

Specification b r
"3=2
Conditional likelihood; p(b, S) /j S j , r 2 ("1, 1) 0.437 0.9800
"3=2
Conditional likelihood; p(b, S) /j S j , r 2 ("1, 1) 0.441 0.9798
"3=2
Exact likelihood; p(b, S) /j S j , r 2 ("1, 1) 0.375 0.9828

Exact likelihood; p(b, S) / (1 " r2 )"1 s2v j S j"5=2 , r 2 ("1, 1) 0.276 0.9872

*Results are from Stambaugh (1999, figure 1). The predictor variable is the dividend-price ratio. Data are monthly
from 1952 to 1996. The conditional likelihood refers to Equation 13; the exact likelihood refers to Equation 8 with
initial condition given by Equation 14.
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

Table 1 shows the implications of these specification choices for the posterior mean of
by University of Pennsylvania on 11/26/10. For personal use only.

the regressive coefficient b and the autocorrelation r.3 For the conditional likelihood
and prior (Equation 11), the posterior mean of beta equals the OLS regression coef-
ficient (which is biased upward). When values of r are restricted to be between "1 and 1,
the posterior mean of b is slightly higher. By contrast, when the exact likelihood is used, the
posterior mean of b is lower and the difference is substantial, regardless of whether the
uniform prior or the Jeffreys prior is used.
To understand these differences in posterior means, consider the following approximate
relation (Stambaugh 1999):
! "
suv
E½b j D) * b^ þ E 2 j D ðE½r j D) " r ^Þ: ð18Þ
sv

Because suv is negative, positive differences between the posterior mean of r and r ^ trans-
late into negative differences between b and b. ^ Equation 18 is the Bayesian version of the
observation that the upward bias in b^ originates from the downward bias in r ^. OLS
estimates the persistence to be lower than what it is in population: This bias arises from
the need to estimate both the sample mean and the regression coefficient at the same time;
the observations revert more quickly to the sample estimate of the mean than the true mean
(Andrews 1993). Because of the negative correlation, OLS also estimates the predictive
coefficient to be too high. Intuition for this result is as follows: If r is above the OLS
estimate r^, then r
^ is “too low,” i.e., in the sample, shocks to the predictor variable tend to
be followed more often by shocks of a different sign than would be expected by chance.
Because shocks to the predictor and the return variables are negatively correlated, this
implies that shocks to the predictor variable tend to be followed by shocks to returns of
the same sign. This implies that b^ will be “too high.”
Compared with the uniform prior over "1 to 1, the prior that restricts r to be
between –1 and 1 lowers (slightly) the posterior mean of r because it rules out draws of r
that are greater than one. For this reason it raises (slightly) the posterior mean of b even

3
The specifications involving the exact likelihood or the Jeffreys prior do not admit closed-form solutions for the
posterior distribution. Nonetheless, the posterior can be constructed using the Metropolis-Hastings algorithm (see
Chib & Greenberg 1995, section 5). See Johannes & Polson (2006) for further discussion of sampling methods for
solving Bayesian portfolio choice problems.

www.annualreviews.org ' Asset Allocation 183

 above the OLS value. This result is analogous to the fact that imposing stationarity in a
frequentist framework implies additional evidence in favor of predictability (Lewellen
2004, Campbell & Yogo 2006, Campbell 2008, Cochrane 2008). Introducing the exact
likelihood leads to an estimate of r that is higher that r ^. This result (which is sample
dependent) arises from two sources of evidence on r: the evidence from the covariance
between xt and xtþ1 and the evidence from the difference between x0 and the sample mean.
If x0 is relatively far from the sample mean, the posterior of r shifts toward higher values.
This implies that r ^ is lower than r and, therefore, that b^ is higher than b. Introducing the
Jeffreys prior in combination with the exact likelihood further shifts r back toward 1; this
raises r relative to r ^ The net bias reduction resulting from these
^ and lowers b relative to b.
modifications is smaller than standard frequentist-based estimates of the bias. Whether this
is good, bad, or merely neutral depends on one’s perspective (Sims & Uhlig 1991).
Table 2 shows the implications for expected returns and asset allocation by reporting
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

these values at various levels of the dividend yield. Comparing the first and last rows of
each panel shows that Bayesian estimation with the conditional likelihood and prior
(Equation 11) have implications that are virtually identical to ignoring parameter uncer-
by University of Pennsylvania on 11/26/10. For personal use only.

tainty and using the OLS estimates. For the exact likelihood and prior (Equation 15), both
the expected returns and allocations are less variable, as one would expect given the lower
posterior mean of b. Surprisingly, not only are the expected returns less variable, but they
are also substantially lower for both values of the dividend yield, leading to lower alloca-
tions as well. In fact, the average excess stock return is different in the various cases
(Wachter & Warusawitharana 2009b). As explained in that paper, differences in estimates
of average excess stock returns arise from differences in estimates of the mean of the

Table 2 Expected returns and optimal allocations under various combinations of the likelihood and prior (monthly
horizon)*

Current dividend yield

Specification 3% 4% 5%

Panel A: Expected excess returns (in percent)

"3=2
Conditional likelihood; p(b, S) /j S j , r 2 ("1, 1) 2.0 7.3 12.5
"3=2
Exact likelihood; p(b, S) /j S j , r 2 ("1, 1) 1.0 5.5 10.0

Exact likelihood; p(b, S) / (1 " r2 )"1 s2v j S j"5=2 , r 2 ("1, 1) 1.9 5.2 8.5

Conditional MLEs as true parameters 2.0 7.3 12.5

Panel B: Stock allocation (in percent)

Conditional likelihood; p(b, S) /j S j"3=2 , r 2 ("1, 1) 22 61 97

"3=2
Exact likelihood; p(b, S) /j S j , r 2 ("1, 1) 15 46 79

Exact likelihood; p(b, S) / (1 " r2 )"1 s2v j S j"5=2 , r 2 ("1, 1) 21 45 68

Conditional MLEs as true parameters 22 60 98

*Results are from Stambaugh (1999, tables 3, 4). The predictor variable is the dividend-price ratio. Data are monthly from 1952 to 1996. The
conditional likelihood refers to Equation 13; the exact likelihood refers to Equation 8 with initial condition given by Equation 14. The table
assumes that the investor has a horizon of one month and has constant relative risk aversion equal to 7.

184 Wachter
 predictor variable. Over this sample, the conditional maximum likelihood estimate of the
dividend yield is below the exact maximum likelihood estimate. Therefore, shocks to the
predictor variable over the sample period must have been negative on average; it follows
that shocks to excess returns must have been positive on average. Accordingly, the poste-
rior mean of returns is below the sample mean.

2.4. Informative Priors

Introducing a Jeffreys prior and the exact likelihood has the effect of making portfolio
choice less sensitive to the dividend yield, as compared with the conditional Bayesian
model. However, the agent still engages in market timing to a large degree. As Wachter &
Warusawitharana (2009a) show, the priors described above may assign an unrealistically
high probability to high R2 statistics in the regression equation. These authors also argue
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

that economic theory points toward low levels of the R2, should predictability exist at all.
Let
by University of Pennsylvania on 11/26/10. For personal use only.

s2v
s2x ¼ , ð19Þ
1 " r2

and note that Equation 19 is the unconditional variance of xt. The population R2 for the
regression given in Equation 3 is defined to be the ratio of the variance of the predictable
component of the return to the total variance. It follows from Equation 19 that the R2 is
equal to

b2 s2x
R2 ¼ : ð20Þ
b2 s2x þ s2u

Wachter & Warusawitharana (2009a) consider a class of priors that translate into
distributions on the population R2. Specifically, they define a “normalized” b:

! ¼ s"1
u sx b:

They also assume that the prior distribution for ! equals

! & N(0, s2! ): ð21Þ

The population R2 can be rewritten in terms of !:

!2
R2 ¼ : ð22Þ
!2 þ1

Equation 22 provides a mapping between a prior distribution on ! and a prior distribution

on the population R2. The prior distribution for ! implies a conditional prior for b.
Namely,

b j a, y, r, S & N(0, s2! s"2 2

x su ): ð23Þ

Because sx is implicitly a function of r and sv, the prior on b is also a function of these
parameters. The approximate Jeffreys prior for the remaining parameters is given by
5
p(a, y, r, S) / sx su j S j"2 : ð24Þ

www.annualreviews.org ' Asset Allocation 185

 Equations 23–24 form a class of prior distributions indexed by !. For s! ¼ 0, the prior
dogmatically specifies that there can be no predictability: b is identically equal to zero. For
s! ¼ 1, the prior is uninformative and is in fact equal to the approximate Jeffreys prior in
Stambaugh (1999). Because of the relation between ! and the R2, a prior on ! translates
directly into a prior on the R2. An appeal of this approach is its scale-invariance: It is hard
to imagine putting an economically meaningful prior on b without knowing something
about the variance of the predictor variable x.
Figure 2 illustrates the implications of different values of s! for the prior distribution on
the R2. On the one hand, the prior with s! ¼ 0 implies a dogmatic view that there can be no
predictability, which is why the R2 is a point mass at zero. On the other hand, with s! ¼ 100
(which well approximates the Jeffreys prior), the R2 is nearly flat over the single-digit
range, dipping down in a region close to 1. Figure 2 shows that uninformative beliefs imply
not only that high values of the population R2 are possible, but also that they are extremely
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

likely. The prior assigns a probability of nearly 100% to the R2 exceeding any given value,
except for values that are an infinitesimal distance from one.
The literature has considered other specifications for informative priors. Kandel &
by University of Pennsylvania on 11/26/10. For personal use only.

Stambaugh (1996), for example, construct priors assuming that the investor has seen, in
addition to the actual data, a hypothetical prior sample of the data such that the sample
means, variances, and covariances of returns and predictor variables are the same in the

ση =0
1
ση =0.04
ση =0.08
ση =5
0.8
ση =10
ση =100

0.6
P(R2>k)

0.4

0.2

0
0 0.01 0.02 0.03
// 0.5 1
k
Figure 2
The prior probability that the R2 exceeds a value k implied by various prior beliefs. Prior beliefs are indexed by s! , the prior
standard deviation of the normalized coefficient on the predictor variable. The dogmatic prior is given by s! ¼ 0; the diffuse prior
by s! ¼ 1. Intermediate priors express some skepticism over return predictability. Note that left portion of the x-axis of the
graph is scaled differently from the right portion.

186 Wachter
 hypothetical prior sample as in the actual sample. However, in the hypothetical sample, the
R2 is exactly equal to zero (see also Avramov 2002, 2004). Cremers (2002) constructs
informative priors assuming the investor knows sample moments of the predictive variable.
These constructions raise the question of how the investor knows the sample moments of
returns and predictive variables (note that it is not sufficient for the investor to make a guess
that is close to the sample values). If it is by seeing the data, the prior and the posterior are
equal and the problem reduces to the full-information case. An alternative is to assume that
the investor has somehow intuited the correct values. According to this latter (somewhat
awkward) interpretation, to be consistent these moments would have to be treated as
constants (namely conditioned on) throughout the analysis, which they are not.
Figure 2 suggests that priors of the form (Equations 23 and 24) with small s! have more
reasonable economic properties than uninformative priors. Wachter & Warusawitharana
(2009a) investigate the quantitative implications of these priors for portfolio allocation.
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

Not surprisingly, because the posterior mean of b shrinks toward zero, the portfolio
allocation under these priors exhibits less dependence on the dividend yield.
Several papers critique the evidence in favor of predictability based on out-of-sample
by University of Pennsylvania on 11/26/10. For personal use only.

performance: Bossaerts & Hillion (1999) find no evidence of out-of-sample return pre-
dictability using a number of predictors, whereas Goyal & Welch (2008) find that pre-
dictive regressions often perform worse than using the sample mean when it comes to
predicting returns. For the researcher, these studies raise the question of how the Bayesian
asset allocation strategies perform out of sample. Note, however, that from the point of
view of the Bayesian investor, such additional information is irrelevant. The predictive
distribution for returns, as generated from the likelihood and the prior, is the sole determi-
nant of the portfolio strategy.
Wachter & Warusawitharana (2009a) examine the out-of-sample performance implied by
various priors. They show that asset allocation, using the results of OLS regression without
taking parameter uncertainty into account, indeed delivers worse out-of-sample performance
than a strategy implied by a dogmatic belief in no predictability. Relative to the OLS bench-
mark, the strategy implied by the uninformative Jeffreys prior (16) performs better, but still
worse than the no-predictability prior. Across various specifications, the best-performing prior
is an intermediate one, representing some weight on the data and some weight on an econom-
ically reasonable view that, if predictability should exist, the R2 should be relatively small.
Campbell & Thompson (2008) adopt a second approach to improving out-of-sample
performance. They show that the out-of-sample performance improves when weak eco-
nomic restrictions are imposed on the return forecasts, thereby requiring that the expected
excess return be positive and that the predictor variable has the theoretically expected sign.
The Campbell and Thompson paper is non-Bayesian, but it would not be difficult to
incorporate these prior views into a Bayesian setting.

2.5. Additional Sources of Uncertainty

One of the objectives of adopting Bayesian decision theory into the asset allocation prob-
lem is to better capture the uncertainty faced by investors. However, despite the uncertain
nature of the predictive relation, estimation risk appears to play a minor role in the
empirical findings. The disconnect between these results and our intuition may be due to
the fact that assuming the model given by Equations 3–6 still, to a large degree, understates
the uncertainty actually faced by investors. Although investors do not know the parameters

www.annualreviews.org ' Asset Allocation 187

 of the system, they know that returns and predictor variables obey such a system. With the
available data, this information is enough to estimate the parameters precisely. In reality,
investors do not know that returns obey such a system. That is, whereas Equation 3 is
unrestrictive in the sense that one could always regress returns on the lagged dividend yield,
the system itself is restrictive. For instance, it requires not only that utþ1 is an error in the
traditional regression sense of being uncorrelated with the right-hand-side variable, but also
that it is a shock, namely independent of any variable known at time t. The possibility of
other likelihood functions is something that would occur to real-world investors.
Pastor & Stambaugh (2009a, 2009b) confront this problem by assuming that returns
obey a predictive system:
ytþ1 ¼ mt þ utþ1
xtþ1 ¼ (I " A)Ex þ Axt þ vtþ1 ð25Þ
mtþ1 ¼ (1 " r)Er þ rmt þ wtþ1 ,
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

where u, v, and w are iid (across time) and jointly normally distributed. Here, m (unobserved)
is the true expected excess return, and the agent learns about m by observing x and y. Under
by University of Pennsylvania on 11/26/10. For personal use only.

this predictive system, one could still regress ytþ1 on the observable xt. However, the error in
the regression would be correlated with time-t variables. Pastor & Stambaugh find that this
distinction between mt and xt , and particularly the fact that the autocorrelation of x need not
equal the autocorrelation of m, has important consequences for investors.
One could expand the uncertainty faced by investors in other ways. Recent studies
(Avramov 2002, Cremers 2002, Wachter & Warusawitharana 2009b) explore the possibil-
ity that an investor assigns some prior probability to alternative models. Although this
represents a form of “model uncertainty,” the agent is still Bayesian in the sense that he
assigns probabilities. One could go further and assume that there are some forms of
uncertainty that investors simply cannot quantify. Gilboa & Schmeidler (1989) define a
set of axioms on preferences that distinguish between risk (in which the agent assigns
probabilities to states of nature) and uncertainty (in which probabilities are not assigned).
They show that aversion to uncertainty leads investors to maximize the minimum over the
set of priors that may be true. Uncertainty aversion, also called ambiguity aversion, has
been the subject of a fast-growing literature in recent years, much of which has focused on
asset allocation (see Chamberlain 1999, Chen & Epstein 2002, Chen et al. 2009, Garlappi
et al. 2007, Hansen 2007, Maenhout 2006).
This notion of additional uncertainty facing investors is likely to be a subject of contin-
ued active debate. As discussed above, there are a number of complementary approaches,
such as the predictive system, model uncertainty with probabilities over the models, and
model uncertainty such that the agent need not formulate probabilities over the models.
The contention of the previously discussed models is that periods of low valuation (e.g.,
when the dividend yield is high) represent, to some uncertain extent, a readily available
opportunity for the investor. However, another possibility is that the excess returns earned
by this market-timing strategy are a compensation for a type of risk that does not appear in
the sample, i.e., the risk of a rare event.4

4
Yet another possibility is that the excess returns represent compensation for greater volatility. Shanken & Tamayo
(2005) evaluate this claim directly in a Bayesian setting and find little support for it. A large literature debates the
extent to which changes in volatility are linked to changes in expected returns; based on available evidence, however,
it does not appear that the fluctuations in expected returns captured by the dividend yield correspond to changes in
volatility. See Campbell (2003) for a discussion of this literature.

188 Wachter
 In Wachter (2008), I show that predictability in excess returns can be captured by a
model with a representative investor with recursive preferences (see below), in which there
is a time-varying probability of a rare event. Times when this rare-event probability are
high correspond to times when the dividend yield is also high. Most of the time, the rare
event does not happen, implying higher than average realized returns. Occasionally, the
rare event does happen, in which case high dividend yields are followed by quite low
returns. The representative agent holds a constant weight in equities (as is required by
equilibrium) despite the fact that excess returns vary in a predictable fashion. Strategies
that attempt to time the market, according to this view, are risky, though this risk would be
difficult to detect in the available time series.

3. DYNAMIC MODELS
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

I now consider the investor who has a horizon beyond one period and, at each time point,
faces a consumption and portfolio choice decision. I start with a general specification that
by University of Pennsylvania on 11/26/10. For personal use only.

allows for multiple risk assets and state variables. Let Ct denote the investor’s consumption
at time t, zt the N +1 vector of allocations to risky assets, and Wt the investor’s wealth.
Samuelson (1969) models this problem as

XT
C1"g
max E e"bt t ð26Þ
c,z
t¼0
1"g

subject to the budget constraint

Wtþ1 ¼ (Wt " Ct )Rf ,tþ1 þ Wt zt> (Rtþ1 " Rf ,tþ1 ) ð27Þ
"bt C1"g
and terminal condition WT , 0. Here e represents period utility (for simplicity, I have
t
1"g
assumed that the investor does not have a bequest motive). An alternative is to consider the
1"g
problem without the utility flow from consumption, namely the investor maximizes W1"g T
.
This is not as realistic, but it is sometimes a helpful simplification. Another helpful sim-
plification is to take the limit of Equation 26 as T goes to infinity.
The problem above can be solved by backward induction using the Bellman equation
(see Duffie 1996, ch. 3). Let Xt denote an n+1 vector of state variables that determine the
distribution of returns. Let t ¼ T"t denote the horizon. Define the value function as the
remaining utility:
t
X
J(Wt , Xt , t) ¼ max E e"bs u(ctþs ):
c,z
s¼0

Then it follows that J can be defined through backward induction as

J(Wt , Xt , t) ¼ u(ct ) þ e"b Et ½ J(Wtþ1 , Xtþ1 , t " 1)) ð28Þ

with the boundary condition J(W, X, 0) ¼ u(W). See Brandt (2009) for further discussion
of the value function and its properties.
Although Equation 28 reduces the multiperiod problem (Equation 26) to a series of one-
period problems, these one-period problems may look quite different from the problem con-
sidered in Section 2 because of the interaction between the state variables X and wealth W.
Indeed, when there is no X (so returns are iid), Samuelson (1969) shows that Equation 26

www.annualreviews.org ' Asset Allocation 189

 reduces to a series of one-period problems. In this article, however, I am primarily inter-
ested in the case where returns are not iid. Merton (1973) characterizes the solution to this
general problem. For technical reasons that are discussed below, it is easier to do this if one
assumes that time is continuous.

3.1. Return Distribution and the Value Function

Let Bt denote a d+1 vector of independent Brownian motions. Let

l(X) ¼ ½l1 (X), . . . , lN (X))>

denote the N+1 vector of instantaneous excess returns and

s(X) ¼ ½s1 (X)> , . . . , sN (X) > )>

Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

denote the N+d matrix of loadings on the Brownian motions. Assume that the price
process for asset i, i ¼ 1, . . ., N is given by
by University of Pennsylvania on 11/26/10. For personal use only.

dP(ti)
¼ (li (Xt ) þ rf (Xt )) dt þ si (Xt ) dBt , ð29Þ
P(ti)
where rf ¼ logRf . I assume Xt follows a Markov process:

dXt ¼ b(Xt ) dt þ a(Xt ) dBt : ð30Þ

Assumptions in Equations 29 and 30 imply that the current value of the state variables
at time t fully determine the investment opportunities that are available to the investor.
That is, they determine the investment opportunity set.
Merton (1971) shows that under the assumptions above, wealth follows the process
% &
dWt ¼ Wt zt> l(Xt ) þ Wt rf (Xt ) " Ct dt þ Wt zt> s(Xt ) dBt : ð31Þ

Merton (1973) derives a partial differential equation characterizing the value function J.
Moreover, he shows that the first-order condition with respect to z leads to the following
characterization of z in terms of derivatives of J:

JW
(ss > ) l " 1 (ss>) sa> JXW ,
"1 "1
z¼" ð32Þ
JWW W JWW W
where JW , JWW , and JXW refer to first and second partial derivatives of J. Here and in
what follows, I eliminate time subscripts and function arguments when not required
for clarity. I show in the Appendix (section below) that the value function takes the
form

I(X, t)1"g W 1"g

J(W, X, t) ¼ : ð33Þ
1"g
Applying Equation 33, it follows that the allocation can be rewritten as
>
1
(ss>) l þ 1 " g (ss>) sa> IX :
"1 "1
z¼ ð34Þ
g g I
Equation 34 (and, more generally, Equation 32) provides a gateway to understanding
portfolio choice in this rich dynamic context. There are two terms in Equation 34, only one

190 Wachter
 of which depends on the process for X. Note that in the discrete-time setting when one
period remains, the value function depends only on wealth, not on X. The same is true in
continuous time; in the limit, as the horizon approaches 0, the value function’s dependence
on X also approaches zero. Therefore, as the horizon approaches 0, only the first term
remains. As a result, Merton (1973) refers to this term as what the investor would choose
if he behaved myopically, namely if, similar to the discrete-time investor with one period
left, he took into account only the very immediate future and did not look beyond.
Given that myopic demand captures, in a limiting sense, the desired allocation of a one-
period investor, how does it compare with the results derived in Section 2? Consider for
simplicity the case of a single risky asset. In this case, l corresponds to the (instantaneous)
expected excess return on the asset and ss > to the (instantaneous) variance. Indeed, Ito’s
Lemma implies that for an asset with price Pt
# $
1
d logPt ¼ l þ rf " ss > dt þ s dBt ,
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

2
so that, assuming units are the same, Et ½ytþ1 ) * l " 12 ss > and Vart ½ytþ1 ) * ss > . Myopic
by University of Pennsylvania on 11/26/10. For personal use only.

demand therefore closely resembles Equation 10. The main difference is that Equation 10
is approximate, whereas Equation 34 is exact. Recall that in the setting of Section 2 (indeed
in any discrete-time setting) power preferences rule out levered positions or short positions
in the stock at any horizon. However, when trading is continuous, the agent can exit these
positions in time to avoid negative wealth. This property, which is not without controversy,
plays a key role in making the continuous-time model tractable.
Myopic demand, then, is the continuous-time analog of the static portfolio choice
described in Section 2. In contrast, the second term in Equation 34 is completely new. As
Merton (1973) shows, this term represents the agent’s efforts to hedge future changes in the
investment opportunity set. There are two offsetting motives: On the one hand, the inves-
tor would like more wealth in states with superior investment opportunities, all the better
to take advantage of them. On the other hand, the investor would like more wealth in
states with poorer investment opportunities, so as to lessen the overall risk to long-term
wealth. The former is a substitution effect; the latter is an income effect.
To see how these motives are represented by Equation 34, consider the case with a
single state variable. Note that the sign of JX equals the sign of IX. Define an increase in X
to indicate an improvement in investment opportunities if and only if it increases the
agent’s utility, namely if and only if JX > 0. [Merton (1973) discusses hedging motives in
terms of the consumption-wealth ratio rather than the value function. I explore the link to
the consumption-wealth ratio in what follows.] If an asset positively covaries with the
stock, hedging demand is negative so long as g is greater than 1 and positive so long as g is
less than 1. In effect, the agent with g > 1 reduces his investment to an asset that pays off in
states with superior investment opportunities (the income effect dominates), whereas the
agent with g < 1 increases his investment to such an asset (the substitution effect domi-
nates). Logarithmic utility (g ¼ 1) corresponds to the knife-edge case when these effects
cancel each other out.
To go further, it is necessary to learn more about the function I(X,t). This function
depends on the parameters in Equations 29 and 30, so it will embody an empirical statement
about the distribution of returns. Applying the theory above to estimated processes for
returns is one way the literature has built on the insights in Merton (1973). A second source
of innovation is in the type of utility function considered (see next section).

www.annualreviews.org ' Asset Allocation 191

 3.2. Recursive Utility
One limitation of the assumption in Equation 26 is its implication that an identical param-
eter, g, controls both the agent’s attitudes toward the smoothness of consumption over time
and the agent’s attitudes toward the smoothness of consumption over states, namely her
attitudes toward risk. Building on the work of Kreps & Porteus (1978), Epstein & Zin
(1989, 1991) and Weil (1990) develop a class of utility functions that retains the attractive
scale invariance of power utility but that allows for a separation between the concepts of
risk aversion and the willingness to substitute over time. Such a separation implies that the
agent has preferences over the timing of the resolution of uncertainty, which may itself be
attractive. The resulting utility function lies outside out of the expected-utility framework
in the sense that the utility cannot be written explicitly as an expectation of future con-
sumption. Rather, utility is defined recursively.5
I use the continuous-time formulation of the Epstein & Zin (1989) utility function
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

developed by Duffie & Epstein (1992a, 1992b). Let Vt denote the remaining utility.
Following Duffie and Epstein, I use the notation V to denote the utility process and the
by University of Pennsylvania on 11/26/10. For personal use only.

notation J to denote optimized utility as a function of wealth, the state variables, and the
horizon. At the optimum, Vt ¼ J(Wt , Xt , T " t). Duffie and Epstein specify Vt as follows:
Z T
Vt ¼ Et f (Cs ,Vs ) ds, ð35Þ
t

where 8 0 1 1
>
> 1 !1" c
>
> b % & B % & " C
>
> (1 " g)V @ C (1 " g)V 1"g " 1A c 6¼ 1
>
>
<1 " 1
>
f (C,V) ¼ c ð36Þ
>
> 0 1
>
>
>
> 1
>
> b((1 " g)V)@log C " log((1 " g)V)A c ¼ 1:
>
: 1"g

Duffie & Epstein (1992a) show that the parameter c > 0 can be interpreted as the
elasticity of intertemporal substitution (EIS) and g > 0 can be interpreted as relative risk
aversion. When g ¼ 1/c, power preferences given in Equation 26 are recovered (note that
the resulting formulation of Vt may not take the same form as Equation 26 but will imply
the same underlying preferences and therefore the same choices).
Results in Duffie & Epstein (1992a) show that the first-order condition for portfolio
allocation (Equation 32) and the first-order condition for consumption fc ¼ Jw derived by
Merton (1973) are valid in this more general setting. Below, I use these results to charac-
terize optimal consumption and investment behavior, considering the case of c 6¼ 1 and
c ¼ 1 separately.6

5
Kihlstrom (2009) develops an alternative approach to separating the inverse of the elasticity of substitution and risk
aversion within an expected-utility framework.
6
Interesting questions of existence and uniqueness of solutions are beyond the scope of this study. Schroder &
Skiadas (1999) provide such results assuming bounded investment opportunities and a utility function that generalizes
the recursive utility case considered here. Wachter (2002) proves existence in the return predictability case (Section 3.3)
under power utility with risk aversion greater than 1. Dybvig & Huang (1988) and Dybvig et al. (1999) provide further
existence results under power utility.

192 Wachter
 3.2.1. Characterizing the solution when the EIS does not equal 1. As shown in the Appen-
dix (see below), so long as c 6¼ 1, the form of the value function (Equation 33), and
therefore the form of optimal allocation (Equation 34), still holds. Myopic demand takes
the same form as under power utility: It is determined by g alone. The parameter g also
determines whether the income or substitution effect dominates in the portfolio decision.
These results support the interpretation of g as risk aversion in this more general model.
It is also instructive to consider the consumption policy. Define a function H as follows:

H(X, t) ¼ b"c I(X, t)"(1"c) : ð37Þ

It follows from the first-order condition for consumption (fc ¼ JW) that the wealth-
consumption ratio is equal to H:

Wt
¼ H(Xt , T " t): ð38Þ
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

Ct
It follows from Equation 37 that
by University of Pennsylvania on 11/26/10. For personal use only.

IX 1 HX
¼" : ð39Þ
I 1"c H
Recall that the sign of IX equals the sign of JX, the derivative of the value function
with respect to the state variables. As in the asset allocation decision, there are two
effects that changes in investment opportunities could have on consumption behavior.
On the one hand, an improvement could lead investors to consume less out of wealth, to
better take advantage of the opportunities (the substitution effect). On the other, an
improvement raises wealth in the long run, allowing the investor to consume more today
(the income effect). Equation 39 shows that, for investors who are relatively willing to
substitute intertemporally (c>1), consumption falls relative to wealth when investment
opportunities rise (the substitution effect dominates). For investors who are relatively
unwilling to substitute intertemporally (c<1), consumption rises (the income effect
dominates). These results support the interpretation of c as the elasticity of intertemporal
substitution.
Substituting into the Bellman, Equation 53 leads to the following differential equation
for H:

1 Ht 1 HX 1 1 > % >&"1 1 " g 1 HX > % >&"1

" bþ l ss l" as ss l
1"c H 1"c H 2g 1"cg H
0 1 0 1
1 1 @1 " g H> H 1 1 % &1
þ rf þ þ 1Atr@a> X X aA " tr a> HXX a
21 " c 1 "c H H 2 1"c H
0 12
ð40Þ
1 1 " g A 1 HX > % >&"1 > HX>
þ @ as ss sa
2 1"c g H H
0 1"1
1 1
" H "1 " b@1 " A ¼ 0,
1"c c

with boundary condition H(X, 0) ¼ 0. Equation 40 is useful in considering the special cases
below.

www.annualreviews.org ' Asset Allocation 193

 Constant investment opportunities. In the special case of constant investment opportuni-
ties, portfolio choice is myopic (as explained above). The wealth-consumption ratio can
also be derived in closed form. The differential equation for H (which is now a function of
t alone) is given by
# $ !
1 0 1 1 > % > &"1 1 "1 1
H þ l ss l þ rf " b 1 " H" ¼ 0: ð41Þ
1"c 2g c 1"c

The solution is
1 ' (
H(t) ¼ 1 " e"k(1"c)t , ð42Þ
k(1 " c)
where
# $
1 1 > % > &"1 1 "1
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

k¼ l ss l þ rf " b 1 " :
2g c
by University of Pennsylvania on 11/26/10. For personal use only.

The first two terms in k provide a measure of the quality of investment opportunities.
For c > 1, H(t) is increasing in k. This follows from the fact that H(0) ¼ 0 and that H0 (t) is
increasing in k for any (fixed) t > 0. As such, the greater the investment opportunities are,
the less the investor consumes out of wealth. Note that the discount rate enters k with a
negative sign: Whereas an increase in investment opportunities causes the investor to
consume less out of wealth, an increase in the discount rate causes the investor to consume
more. For c < 1, H(t) is decreasing in k. The greater the investment opportunities are, the
more the investor consumes out of wealth. Also, an increase in investment opportunities
and an increase in the discount both lead the investor to consume more and save less as a
percentage of wealth.

Power utility and complete markets. In this setting without trading restrictions, markets
are complete if and only if the diffusion terms for asset prices span the diffusion terms
"1
for X. The term as> (ss >) represents the projection of the diffusion terms for X on the
(i)
diffusion terms for P ; therefore, markets are complete if and only if
"1
as> (ss > ) s ¼ a,
namely if the projection recovers the diffusion terms on X. Further note that, because
tr(AB) ¼ tr(BA) for conforming matrices,

tr(a > HX> HX a) ¼ tr(HX aa > HX>) ¼ HX aa > HX> :

Therefore, Equation 40 reduces to the much simpler

1g " 1 > "1 g"1 "1

Ht " HX b þ l (ss > ) lH þ HX as> (ss >) l
2 g2 g
g"1 1 1 ð43Þ
þ r H " tr(a > HXX a) " 1 þ bH ¼ 0:
g f 2 g

Equation 43 has a solution of the form

Z t
H(X, t) ¼ F(X, s) ds ð44Þ
0

194 Wachter
 with F(X, 0) ¼ 1. To see this, note that it follows from integration by parts that
Z t
@F
ds ¼ F(X, t) " F(X, 0) ¼ Ht " 1:
0 @s

Substituting in, I find that F satisfies

@F 1 g " 1 > % > &"1 g"1 % &"1
" FX b þ l ss lF þ FX as> ss> l
@t 2 g2 g
g"1 1 % & 1 ð45Þ
þ r F " tr a > FXX a þ bF ¼ 0:
g f 2 g
It is no accident that the differential equation simplifies under the case of power utility
and complete markets. This is the case when the conceptually simpler martingale method
of Cox & Huang (1989), Karatzas et al. (1987), and Pliska (1986) is straightforward to
apply (for an example, see Wachter 2002). Although this method can be extended to
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

incomplete markets (He & Pearson 1991, Cuoco 1997) and to recursive utility (Duffie &
Skiadas 1994, Schroder & Skiadas 1999, Skiadas 2007), it is less straightforward in these
by University of Pennsylvania on 11/26/10. For personal use only.

cases.

3.2.2. Characterizing the solution when the EIS equals 1. In the case of c ¼ 1, the value
function takes the form
"bt
W (1"g)(1"e ) G(X, t)1"g
J(W, X, t) ¼ :
1"g
The differential equation for G is given in the Appendix (see below). Equation 32 still holds
(see Duffie & Epstein 1992a), implying that the portfolio allocation is given by
1 % > &"1 1"g % > &"1 > GX>
z¼ ss lþ ss sa : ð46Þ
1 " (1 " g)(1 " e"bt ) 1 " (1 " g)(1 " e"bt ) G
As in the case of c 6¼ 1, the portfolio allocation separates into two terms, the first of which
can be interpreted as myopic demand (because it does not depend on future investment
opportunities) and the second as hedging demand.
Myopic demand is horizon dependent when c ¼ 1. When the horizon is large (as t ! 1),
myopic demand approaches the myopic demand in Equation 34, namely it is determined
by g only. However, for finite horizons, myopic demand is determined by a weighted
average of c"1 (¼1) and g, with the horizon determining the weights:
1 " (1 " g)(1 " e"bt ) ¼ e"bt þ g(1 " e"bt ):
The first-order condition fc ¼ JW applied to Equation 46 implies that the wealth-
1"e"bt
consumption ratio is given by WC ¼ b . Unlike in the c 6¼ 1 case, the wealth-consumption
ratio does not depend on investment opportunities. Unit EIS corresponds to the knife-edge
case where the substitution and income effects cancel each other out, as far as consumption
behavior is concerned. In the limiting case of an infinite horizon, the wealth-consumption
ratio is constant and equal to b"1.

3.3. Time-Varying Risk Premia

I now consider a special case of price dynamics in which there is a single risky asset and a
single state variable Xt. This example is meant to be illustrative; the solution technique can

www.annualreviews.org ' Asset Allocation 195

 be extended to other forms of affine dynamics (see Schroder & Skiadas 1999, Liu 2007).
Specifically, assume rf, s, and a ¼ sX are constants, and let
% &1
l(X) ¼ ss> 2 X
b(X) ¼ "k(X " X):
This specification implies that the expected excess return on the stock is time varying
and depends linearly on a variable Xt that follows a mean-reverting process. It is therefore
the continuous-time equivalent of the process assumed in Section 2 [Wachter (2002) makes
this explicit]. Note that Xt is the Sharpe ratio on the risky asset.

3.3.1. When are exact solutions available? More explicit solutions for the value func-
tion, and therefore for portfolio and consumption choices, are available in two special
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

cases of the above analysis: (a) when the c is equal to 1 (b) when power utility obtains
(g ¼1/c) and markets are complete. Schroder & Skiadas (1999) (who also assume
by University of Pennsylvania on 11/26/10. For personal use only.

complete markets) and Campbell et al. (2004) (who also assume an infinite horizon)
consider the first case. Here I further consider this case, allowing markets to be incom-
plete and the horizon to be finite. The second case is the subject of Wachter (2002). In
a related contribution, Kim & Omberg (1996) show that one can also obtain closed-
form solutions for portfolio choice when the investor maximizes power utility over
terminal wealth.
Indeed, when c ¼ 1, the value function is given by Equation 60, with G taking the
form
) *
X2
G(X, t) ¼ exp A(11) (t) þ A(21) (t)X þ A(31) (t) ð47Þ
2
(1)
and where Ai satisfy a system of ordinary differential equations with boundary conditions
Aði 1Þ (0) ¼ 0. For power-utility and complete markets, the wealth-consumption ratio H(X, t)
is given by Equation 44, where
) *
(2) X2 (2) (2)
F(X, t) ¼ exp A1 (t) þ A2 (t)X þ A3 (t) : ð48Þ
2
(2)
Substituting into Equation 45 results in a set of ordinary differential equation for Ai with
(2)
boundary conditions Ai ¼ 0.

3.3.2. An approximate-solution technique. The affine dynamics above lend themselves to

an approximate-solution technique developed by Campbell & Viceira (1999) based on
earlier work by Campbell (1993). Campbell and Viceira propose log-linearizing the budget
constraint around the mean consumption-wealth ratio. They then derive an approximate
analytical solution to the above problem, assuming an infinite horizon. So long as the
consumption-wealth ratio is not too variable (i.e., the EIS is not far from 1), the approxi-
mation error will be small.
Chacko & Viceira (2005) show how to implement this approximation in a continuous-
time setting. Consider the differential question for the wealth-consumption ratio (Equation
40). Following Chacko and Viceira, I assume that the horizon is infinite, and look for a
stationary solution, namely a solution with Ht ¼ 0. Let h ¼ log H and consider a first-order
approximation of e"h around the mean of "h:

196 Wachter
 % &
e"h * eE½"h) þ "h "E½"h) eE½"h) : ð49Þ
Let h1 ¼ eE½"h) and h0 ¼ h1 (1 " log h1 ). Then Equation 49 implies

H"1 * h0 " h1 log H: ð50Þ

Substitute Equation 50 and Ht ¼ 1 into Equation 40 implies

1 HX 1 1 > % > &"1 1 " g 1 HX > % > &"1
" bþ l ss l" as ss l
1"c H 2g 1"cg H
# $ # $
1 1 1"g H > HX 1 1 % &1
þrf þ þ 1 tr a > X a " tr a > HXX a
21" c 1" c H H 21 " c H
# $2 ð51Þ
1 1"g 1 HX > % > &"1 > HX >
þ as ss sa
2 1"c g H H
# $"1
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

1 1
" (h0 " h1 log H) " b 1 " * 0:
1"c c
by University of Pennsylvania on 11/26/10. For personal use only.

Observe that this differential equation is similar in form to the one for the value
function in the c ¼ 1 case (given by Equation 65). In fact, it is simpler in that there is no
time dependence. It follows that the approximation method can be implemented in any
setting where the c ¼ 1 yields an exact solution. Under the above assumptions on the asset-
return process and state-variable processes,
) *
X2
H(X,t) * exp A(13) þ A(23) X þ A(33) ,
2

where A(i 3) can be determined by matching coefficients. Campbell & Viceira (1999) use this
approximation to show that, in the infinite-horizon problem, portfolio decisions are
driven, almost entirely, by risk aversion g.

3.3.3. Numerical results. When calibrated to reasonable values, what do these dynamic
considerations add to the asset allocation problem? In what follows, I present results from
Wachter (2002); the near-perfect negative correlation between the dividend yield and the
stock return makes it reasonable to assume that markets are complete. I calibrate this
model using the same parameters as used by Barberis (2000) and assume a risk aversion
of 5, so the results are quantitatively comparable to those discussed in Section 2.7 Similar
results are found using alternative specifications and methods (e.g., Brennan et al. 1997,
Brandt 1999, Balduzzi & Lynch 1999).
Figure 3 shows the optimal allocation as a function of horizon for various levels of the
dividend yield. As in the static case, there are substantial horizon and market timing
effects. However, in this case, rather than decreasing (slowly) in the horizon, the degree to
which the allocation varies with the dividend yield is even more marked for long-term
investors than for short-term investors. This greater dependence results from hedging
demand. Because the dividend yield (proportional to X) is negatively correlated with stock
returns, hedging demand leads the investor to allocate more money to stocks for g > 1.

7
Wachter (2002) provides details on how the discrete-time results are used to calibrate the continuous-time model.
However, that paper calibrates the model mistakenly assuming that the process in Barberis (2000) applies to the net
return on equities rather than the excess return. These results correct that mistake.

www.annualreviews.org ' Asset Allocation 197

 4

Allocation to stocks
2

1
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

0
by University of Pennsylvania on 11/26/10. For personal use only.

0 2 4 6 8 10
Horizon (years)

Figure 3
Dynamic allocation as a function of horizon assuming return predictability and that the investor can
trade continuously. The solid line corresponds to the optimal allocation when the dividend yield is at
its sample mean (3.75%). The dash-dotted lines correspond to the allocations when the dividend yield
is one standard deviation above or below its mean (2.91% and 4.59%, respectively). The dotted lines
correspond to the allocations when the dividend yield is two standard deviations above or below its
mean (2.06% and 5.43%, respectively). The agent has power utility over consumption (lines with
circles) or over terminal wealth (lines without circles) with risk aversion equal to five. Note that the
allocations increase as a function of the dividend yield. The model is estimated over monthly data from
1952 to 1995.

The greater the dividend yield is, the more the investor cares about this hedge, which is
why hedging demand makes market timing more extreme. Unlike the simpler horizon
effect in Section 2, this effect reverses for g < 1. The long-horizon investor with g < 1 holds
less in stock than does the short-horizon investor.

3.4. Parameter Uncertainty and Learning in Dynamic Models

So far in this section I have assumed that the investor has full knowledge of the parameters.
I now consider the case of parameter uncertainty in dynamic models. One obvious differ-
ence between the dynamic and static settings is that the degree to which parameters are
uncertain varies over time; that is, the agent learns more about the distribution as time goes
on. Less obviously, learning introduces hedging demands for investors with risk aversion
not equal to one. A well-studied special case is when returns are iid and the investor learns
about the average excess return (see the discussion in Pastor & Veronesi 2009). Given the
assumption of iid returns, it is natural to assume that only the mean is uncertain as, in a
continuous-time setting, the volatility can be estimated with effectively infinite precision
(Merton 1980). Moreover, as the time horizon shortens, the role of uncertainty around the
mean goes to zero (Detemple 1986, Gennotte 1986). The estimated mean will differ from
the true mean; however, the uncertainty around this estimated mean has no effect on the

198 Wachter
 allocation. Given this limiting result in continuous time, it is perhaps not surprising that
estimation risk should have little effect at short horizons as shown in Section 2.
What does have an effect, and a large one, is learning. Hedging demand induced by
learning is negative and can be substantial (Brennan 1998). The reason is that the investor’s
estimate of the average return (effectively a state variable) is positively correlated with
realized returns. When a positive shock to prices occurs, the investor updates his beliefs
about the average return, estimating it to be higher than before. Thus, stocks are less
attractive to an investor with g > 1 (see Equation 34).
Uncertainty about parameters other than the mean is harder to address because it does
not lend itself to closed-form solutions. Studies, therefore, have explored this question
using numerical methods. Xia (2001) allows the investor to be uncertain about the degree
of predictability (the coefficient b in Equation 3) and assumes the other parameters are
known. She decomposes hedging demand into the component to hedge learning about b
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

and the component to hedge changes in Xt. Learning-induced hedging demand decreases in
the difference between the dividend yield and its mean. Moreover, it switches in sign:
It is positive when the dividend yield is below its long-run mean, zero when it is at the
by University of Pennsylvania on 11/26/10. For personal use only.

long-run mean, and negative when it is above the long-run mean. As Xia (2001) shows,
these properties make the overall allocation less variable compared to the no-learning case.
However, the allocation is still more variable than implied by the myopic strategy.
Brandt et al. (2005) and Skoulakis (2007) undertake solving the asset allocation prob-
lem when there is uncertainty about the full set of parameters. The lack of closed-form
solutions and the high dimensionality of the problem make this a formidable technical
challenge. These studies show that, in addition to the effect noted by Xia (2001), uncer-
tainty about the mean (as in Brennan 1998) exerts an important influence, driving down
the average allocation relative to that discussed above. Although the net effect of hedging
demand is under dispute, the market-timing effect remains alive and well.

4. CONCLUDING REMARKS
In this study, I review the literature on static and dynamic asset allocation, with a focus on
the implications of return predictability for long-run investors. For both buy-and-hold and
dynamically trading investors, the optimal allocation to stocks is greater the longer the
horizon, given reasonable assumptions on preferences. This similarity should not obscure
some key differences. In the static case, the effect of any stationary variable on the alloca-
tion will diminish as the horizon grows. In the dynamic case, there is no reason for this to
happen, and indeed the opposite may be true. In effect, for investors who dynamically
trade, even short-term variables can have long-term implications.
This survey also highlights efforts to introduce parameter uncertainty into the agent’s
decision process. This, in theory, serves to pass on some of the uncertainty faced by the
econometrician to the agent; the agent now incorporates this estimation risk into his deci-
sions. Empirically, however, estimation risk appears to have very little effect, except at long
buy-and-hold horizons (at least for the specifications explored herein). This is not to say that
the perfect- and imperfect-information cases are identical. Indeed, learning can induce impor-
tant hedging demands in the dynamic setting. Furthermore, I show in the static setting that
the choice of prior and likelihood can have a large impact on the results. The notion of
uninformative priors is less than clear in a predictive regression setting. Moreover, economic
theory indicates a possible role for unapologetically informative priors that take this theory

www.annualreviews.org ' Asset Allocation 199

 into account. Although these results do not arise from estimation risk per se, they do
incorporate the small-sample nature of the evidence into the decision problem. Our data on
financial markets is unavoidably finite; this should influence agents in economic models just
as it influences the economists doing the modeling. Despite the progress reported here, it is
fair to say that much work along these lines remains to be done.

APPENDIX: SOLVING FOR THE VALUE FUNCTION IN THE DYNAMIC

RECURSIVE UTILITY MODEL
In this Appendix, I use the more general form of the aggregator suggested by Duffie &
Epstein (1992a) for the c ¼ 1 case. The formulas in the text result from taking the limit
as x ! 0.
8 0 1
> !1" 1
> " 1 c
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

>
> b % &B % & 1"g C
>
> (1 " g)V @ C (1 " g)V " 1A c 6¼ 1
>
< 1 " c1
f (C,V) ¼ ð52Þ
by University of Pennsylvania on 11/26/10. For personal use only.

>
>
> %
> & 1
>
> 1"g
log(x1"g þ (1 " g)V))
:b x
> þ (1 " g)V (log C "
1"g
c ¼ 1:

Duffie & Epstein (1992b) derive the continuous-time Bellman equation:

% & 1 % &
"Jt þ JX b þ JW Wz > l þ Wrf " C þ tr(S) þ f C, J(X,W,t) ¼ 0, ð53Þ
2
where
" #> " #" #
a JXX JXW a
S¼ :
Wz > s JWX JWW Wz > s

The first-order condition for consumption is

fC ¼ JW :
Substituting in from Equaiton 52, I find (with some abuse of notation), C as a function of
W, X, and t:
8
>
< c "c % &1"gc
b JW (1 " g)J 1"g c 6¼ 1
C(W, X, t) ¼ ð54Þ
>
: "1 % 1"g &
bJW x þ (1 " g)J c ¼ 1:

It follows from Equation 54 that

8 1"gc
>
> # $"1 # $
>
> 1 % & 1"g 1 "1
>
> c 1"c
>
> b 1 " J W (1 " g)J " b 1 " (1 " g)J c 6¼ 1
>
> c c
>
<
f (C, V) ¼ % & ' "1 % 1"g &( ð55Þ
>
> b x1"g þ (1 " g)J log bJW x þ (1 " g)J "
>
>
>
>
>
> b % 1"g & % &
>
> x þ (1 " g)J log x1"g þ (1 " g)J c ¼ 1:
>
: 1"g

200 Wachter
 Furthermore, note that
8
>
< 1"c % &1 " gc
bc JW (1 " g)J 1 " g c 6¼ 1
CJW ¼ ð56Þ
>
: % 1"g &
b x þ (1 " g)J c ¼ 1:

For c 6¼ 1, substituting into Equation 53 from Equations 55, 56, and 32 implies
2 % &"1
1 JW JW >
"Jt þ JX b " l > ss > l" J as > (ss)"1 l
2 JWW JWW XW

1 1 1 >
þ JW Wrf þ tr(a > JXX a) " J as > (ss> )"1 sa > JXW ð57Þ
2 2 JWW XW
# (
1 1"c % &1"gc 1 "1
bc JW
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

" (1 " g)J 1"g " b 1 " (1 " g)J ¼ 0:

1"c c
by University of Pennsylvania on 11/26/10. For personal use only.

The form of J (Equation 33), combined with Equation 37 implies

% &" 1"g
J(W, X, t) ¼ bc H(X, t) 1"c : ð58Þ

Substituting Equation 58 into Equation 57 leads to Equation 40.

The remainder of this section assumes the c ¼ 1 case. Substituting into Equation 53
from Equations 55, 56, and 32 implies
2 % &"1 % &"1
1 JW JW >
"Jt þ JX> b " l > ss > l" J as> ss> l
2 JWW JWW XW
1 % & 1 1 > % &"1 >
þJW Wrf þ tr a > JXX a " J as > ss > sa JXW
2 2 JWW XW
ð59Þ
0 1
1
"b(x1"g þ (1 " g)J) þ b(x1"g þ (1 " g)J)@1 " Alog(x1"g þ (1 " g)J)
1"g
% "1 &
þb(x1"g þ (1 " g)J)log bJW ¼ 0:

Guess
1
log(x1"g þ (1 " g)J(W, X, t)) ¼ q(t)log W þ log G(X,t): ð60Þ
1"g
Derivatives of J can be found by implicitly differentiating on both sides of Equation 60:
# $
% & Gt
Jt ¼ x1"g þ (1 " g)J q0 logW þ
G
% & 1
JW ¼ x1"g þ (1 " g)J q ð61Þ
W
% & GX
JX ¼ x1"g þ (1 " g)J :
G

Second derivatives follow from Equation 61:

www.annualreviews.org ' Asset Allocation 201

 % & GX
JWX ¼ x1"g þ (1 " g)J q(1 " g)W "1
G
% 1"g & "2 % &
JWW ¼ x þ (1 " g)J W " q þ q2 (1 " g) ð62Þ
# $
% & GX GX> GXX
JXX ¼ x1"g þ (1 " g)J " g þ :
G2 G

Substituting Equations 60–62 into Equation 59 and dividing by x1"g þ (1 " g)J leads to the
following:

Gt GX 1 q % &"1
"q0 logW " þ bþ l> ss> l
G G 2 1 " q(1 " g)
q(1 " g) GX > % > &"1
þ as ss l
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

1 " q(1 " g) G ð63Þ

# $
1 a> GX> GX a a> GXX a 1 q(1 " g)2 GX > % > &"1 > GX>
by University of Pennsylvania on 11/26/10. For personal use only.

þqrf þ tr " g 2
þ þ as ss sa
2 G G 2 1 " q(1 " g) G G

"b " bg(qlogW þ log G) þ b log(bq"1 ) þ blogW " b(1 " g)(qlogW þ log G) ¼ 0:

Matching coefficients on log W implies that

"q0 þ b " bq ¼ 0,

with boundary condition q(0) ¼ 0. Therefore,

q(t) ¼ 1 " e"bt : ð64Þ

The resulting differential equation for G is as follows:

Gt GX 1 q
" þ bþ l> (ss> )"1 l
G G 2 1 " q(1 " g)
q(1 " g) GX >
þ as (ss> )"1 l
1 " q(1 " g) G ð65Þ
# $
1 a> GX> GX a> a > GXX a 1 q(1 " g)2 GX > % > &"1 > GX>
þqrf þ tr " g 2
þ þ as ss sa
2 G G 2 1 " q(1 " g) G G

T " b " bglog G þ blog(bq"1 ) " b(1 " g)log G ¼ 0:

DISCLOSURE STATEMENT
The author is not aware of any affiliations, memberships, funding, or financial holdings
that might be perceived as affecting the objectivity of this review.

ACKNOWLEDGMENTS
I thank Itamar Drechsler, Lubos Pastor, and Moto Yogo for helpful comments and Jerry
Tsai for excellent research assistance.

202 Wachter
 LITERATURE CITED
Ait-Sahalia Y, Hansen LP, eds. 2009. Handbook of Financial Econometrics: Volume 1—Tools and
Techniques. Amsterdam: North Holland. 808 pp.
Andrews DWK. 1993. Exactly median-unbiased estimation of first order autoregressive/unit root
models. Econometrica 61:139–65
Ang A, Bekaert G. 2007. Stock return predictability: Is it there? Rev. Financ. Stud. 20:651–707
Avramov D. 2002. Stock return predictability and model uncertainty. J. Financ. Econ. 64:423–58
Avramov D. 2004. Stock return predictability and asset pricing models. Rev. Financ. Stud. 17:699–738
Avramov D, Zhou G. 2010. Bayesian portfolio analysis. Annu. Rev. Financ. Econ. 2: In press
Balduzzi P, Lynch AW. 1999. Transaction costs and predictability: some utility cost calculations.
J. Financ. Econ. 52:47–78
Barberis N. 2000. Investing for the long run when returns are predictable. J. Financ. 55:225–64
Berger JO. 1985. Statistical Decision Theory and Bayesian Analysis. New York: Springer
Bossaerts P, Hillion P. 1999. Implementing statistical criteria to select return forecasting models: What
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

do we learn? Rev. Financ. Stud. 12:405–28

Boudoukh J, Michaely R, Richardson M, Roberts MR. 2007. On the importance of measuring payout
by University of Pennsylvania on 11/26/10. For personal use only.

yield: implications for empirical asset pricing. J. Financ. 62:877–915

Brandt MW. 1999. Estimating portfolio and consumption choice: a conditional Euler equations
approach. J. Financ. 54:1609–45
Brandt MW. 2009. Portfolio choice problems. See Ait-Sahalia & Hansen 2009, pp. 269–336
Brandt MW, Goyal A, Santa-Clara P, Stroud JR. 2005. A simulation approach to dynamics portfolio
choice with an application to learning about return predictability. Rev. Financ. Stud. 18:831–73
Brennan M. 1998. The role of learning in dynamic portfolio decisions. Eur. Financ. Rev. 1:295–306
Brennan MJ, Schwartz ES, Lagnado R. 1997. Strategic asset allocation. J. Econ. Dyn. Control
21:1377–403
Campbell JY. 1993. Intertemporal asset pricing without consumption data. Am. Econ. Rev.
83:487–512
Campbell JY. 2003. Consumption-based asset pricing. In Handbook of the Economics of Finance,
Vol. IB, ed. G Constantinides, RM Stulz, M Harris. Amsterdam: North-Holland
Campbell JY. 2006. Household finance. J. Financ. 61:1553–604
Campbell JY. 2008. Viewpoint: estimating the equity premium. Can. J. Econ. 41:1–21
Campbell JY, Chacko G, Rodriguez J, Viceira LM. 2004. Strategic asset allocation in a continuous-
time VAR model. J. Econ. Dyn. Control 28:2195–214
Campbell JY, Shiller RJ. 1988. The dividend-price ratio and expectations of future dividends and
discount factors. Rev. Financ. Stud. 1:195–228
Campbell JY, Thompson SB. 2008. Predicting excess stock returns out of sample: Can anything beat
the historical average? Rev. Financ. Stud. 21:1509–31
Campbell JY, Viceira LM. 1999. Consumption and portfolio decisions when expected returns are time-
varying. Q. J. Econ. 114:433–95
Campbell JY, Yogo M. 2006. Efficient tests of stock return predictability. J. Financ. Econ. 81:27–60
Chacko G, Viceira L. 2005. Dynamic consumption and portfolio choice with stochastic volatility in
incomplete markets. Rev. Financ. Stud. 18:1369–402
Chamberlain G. 1999. Econometric applications of maxmin expected utility. Work. Pap., Harvard
Univ.
Chen H, Ju N, Miao J. 2009. Dynamic asset allocation with ambiguous return predictability.
Presented at Am. Financ. Assoc. Meet., Atlanta, GA
Chen Z, Epstein L. 2002. Ambiguity, risk and asset returns in continuous time. Econometrica
70:1403–43
Chib S, Greenberg E. 1995. Understanding the Metropolis-Hastings algorithm. Am. Stat. 49:327–35
Cochrane JH. 1992. Explaining the variance of price-dividend ratios. Rev. Financ. Stud. 5:243–80

www.annualreviews.org ' Asset Allocation 203

 Cochrane JH. 1999. Portfolio advice for a multifactor world. Econ. Perspect. Fed. Reserv. Bank
Chicago 23:59–78
Cochrane JH. 2008. The dog that did not bark: a defense of return predictability. Rev. Financ. Stud.
21:1533–75
Cox JC, Huang C-F. 1989. Optimal consumption and portfolio policies when asset prices follow
a diffusion process. J. Econ. Theory. 49:33–83
Cremers KJM. 2002. Stock return predictability: a Bayesian model selection perspective. Rev. Financ.
Stud. 15:1223–49
Cuoco D. 1997. Optimal consumption and equilibrium prices with portfolio constraints and stochas-
tic income. J. Econ. Theory. 72:33–73
Curcuru SE, Heaton J, Lucas D, Moore D. 2009. Heterogeneity and portfolio choice: theory and
evidence. See Ait-Sahalia & Hansen 2009, pp. 337–82
Detemple JB. 1986. Asset pricing in a production economy with incomplete information. J. Financ.
41:383–91
Duffie D. 1996. Dynamic Asset Pricing Theory. Princeton, NJ: Princeton Univ. Press
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

Duffie D, Epstein LG. 1992a. Asset pricing with stochastic differential utility. Rev. Financ. Stud.
5:411–36
by University of Pennsylvania on 11/26/10. For personal use only.

Duffie D, Epstein LG. 1992b. Stochastic differential utility. Econometrica 60:353–94

Duffie D, Skiadas C. 1994. Continuous-time asset pricing: a utility gradient approach. J. Math. Econ.
23:107–32
Dybvig PH, Huang C-F. 1988. Nonnegative wealth, absence of arbitrage, and feasible consumption
plans. Rev. Financ. Stud. 1:377–401
Dybvig PH, Rogers LCG, Back K. 1999. Portfolio turnpikes. Rev. Financ. Stud. 12:165–95
Epstein L, Zin S. 1989. Substitution, risk aversion and the temporal behavior of consumption and asset
returns: a theoretical framework. Econometrica 57:937–69
Epstein LG, Zin SE. 1991. Substitution, risk aversion, and the temporal behavior of consumption and
asset returns: an empirical analysis. J. Polit. Econ. 99:263–86
Fama EF, French KR. 1989. Business conditions and expected returns on stocks and bonds. J. Financ.
Econ. 25:23–49
Fama EF, Schwert WG. 1977. Asset returns and inflation. J. Financ. Econ. 5:115–46
Garlappi L, Uppal R, Wang T. 2007. Portfolio selection with parameter and model uncertainty:
a multi-prior approach. Rev. Financ. Stud. 20:41–81
Gelman A, Carlin JB, Stern HS, Rubin DB. 1996. Bayesian Data Analysis. London: Chapman & Hall/
CRC
Gennotte G. 1986. Optimal portfolio choice under incomplete information. J. Financ. 41:733–46
Gilboa I, Schmeidler D. 1989. Maxmin expected utility with non-unique prior. J. Math. Econ. 18:141–53
Goetzmann WN, Jorion P. 1993. Testing the predictive power of dividend yields. J. Financ. 48:663–79
Goyal A, Welch I. 2008. A comprehensive look at the empirical performance of equity premium
prediction. Rev. Financ. Stud. 21:1455–508
Graham B, Dodd DL. 1934. Security Analysis. New York: McGraw Hill
Hamilton JD. 1994. Time-Series Analysis. Princeton, NJ: Oxford Univ. Press
Hansen LP. 2007. Beliefs, doubts and learning: valuing macroeconomic risk. Am. Econ. Rev. 97:1–30
He H, Pearson ND. 1991. Consumption and portfolio policies with incomplete markets and short-sale
constraints: the infinite dimensional case. J. Econ. Theory 54:259–304
Hodrick RJ. 1992. Dividend yields and expected stock returns: alternative procedures for inference
and measurement. Rev. Financ. Stud. 5:357–86
Jeffreys H. 1961. Theory of Probability. New York: Oxford Univ. Press
Johannes M, Polson N. 2006. MCMC methods for financial econometrics. In Handbook of Financial
Econometrics: Volume 2—Applications, ed. Y. Ait-Sahalia, L. Hansen, pp. 1–72. 384 pp.
Kandel S, Stambaugh RF. 1996. On the predictability of stock returns: an asset allocation perspective.
J. Financ. 51:385–424

204 Wachter
 Karatzas I, Lehoczky JP, Shreve SE. 1987. Optimal portfolio and consumption decisions for a small
investor on a finite horizon. SIAM J. Contr. Optim. 25:1557–86
Keim DB, Stambaugh RF. 1986. Predicting returns in the stock and bond markets. J. Financ. Econ.
17:357–90
Kihlstrom R. 2009. Risk aversion and the elasticity of substitution in general dynamic portfolio
theory: consistent planning by forward looking, expected utility maximizing investors. J. Math.
Econ. 45:634–63
Kim TS, Omberg E. 1996. Dynamic nonmyopic portfolio behavior. Rev. Financ. Stud. 9:141–61
Kothari S.P, Shanken J. 1997. Book-to-market, dividend yield, and expected market returns: a time-
series analysis. J. Financ. Econ. 44:169–203
Kreps D, Porteus E. 1978. Temporal resolution of uncertainty and dynamic choice theory.
Econometrica 46:185–200
Lettau M, Ludvigson SC. 2001. Consumption, aggregate wealth and expected stock returns. J. Financ.
56:815–49
Lewellen J. 2004. Predicting returns with financial ratios. J. Financ. Econ. 74:209–35
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

Liu J. 2007. Portfolio selection in stochastic environments. Rev. Financ. Stud. 20:1–39
Maenhout P. 2006. Robust portfolio rules and detection-error probabilities for a mean-reverting risk
by University of Pennsylvania on 11/26/10. For personal use only.

premium. J. Econ. Theory. 128:136–63

Markowitz H. 1952. Portfolio selection. J. Financ. 7:77–91
Merton RC. 1971. Optimum consumption and portfolio rules in a continuous-time model. J. Econ.
Theory 3:373–413
Merton RC. 1973. An intertemporal capital asset pricing model. Econometrica 41:867–87
Merton RC. 1980. On estimating the expected return on the market: an exploratory investigation.
J. Financ. Econ. 8:323–61
Pastor L, Stambaugh RF. 2009a. Are stocks really less volatile in the long run? Work. Pap., NBER
Pastor L, Stambaugh RF. 2009b. Predictive systems: living with imperfect predictors. J. Financ.
64:1583–628
Pastor L, Veronesi P. 2009. Learning in financial markets. Annu. Rev. Financ. Econ. 1:361–81
Pliska SR. 1986. A stochastic calculus model of continuous trading: optimal portfolios. Math. Oper.
Res. 11:371–82
Poterba JM, Summers LH. 1988. Mean reversion in stock prices: evidence and implications. J. Financ.
Econ. 22:27–59
Samuelson PA. 1969. Lifetime portfolio selection by dynamic stochastic programming. Rev. Econ.
Stat. 51:239–46
Schroder M, Skiadas C. 1999. Optimal consumption and portfolio selection with stochastic differen-
tial utility. J. Econ. Theory. 89:68–126
Shanken JA, Tamayo A. 2005. Dividend yield, risk, and mispricing: a Bayesian analysis. Work. Pap.,
Emory Univ./Lond. Bus. Sch.
Siegel JJ. 1994. Stocks for the Long Run: A Guide to Selecting Markets for Long-Term Growth. Burr
Ridge, IL: Irwin
Sims CA, Uhlig H. 1991. Understanding unit rooters: a helicopter tour. Econometrica 59:1591–99
Skiadas C. 2007. Dynamic portfolio choice and risk aversion. In Handbooks in Operations Research
and Management Science: Financial Engineering, ed. J Birge, V Linetsky. Amsterdam: North-
Holland
Skoulakis G. 2007. Dynamic portfolio choice with Bayesian learning. Work. Pap., Univ. Maryland
Stambaugh RF. 1999. Predictive regressions. J. Financ. Econ. 54:375–421
Uhlig H. 1994. On Jeffreys prior when using the exact likelihood function. Econ. Theory 10:633–44
Wachter JA. 2002. Portfolio and consumption decisions under mean-reverting returns: an exact solu-
tion for complete markets. J. Financ. Quant. Anal. 37:63–91
Wachter JA. 2008. Can time-varying risk of rare disasters explain aggregate stock market volatility?
Work. Pap. 14386, NBER

www.annualreviews.org ' Asset Allocation 205

 Wachter JA, Warusawitharana M. 2009a. Predictable returns and asset allocation: Should a skeptical
investor time the market? J. Econom. 148:162–78
Wachter JA, Warusawitharana M. 2009b. What is the chance that the equity premium varies over
time? Evidence from predictive regressions. Work. Pap., Board Gov., Fed. Reserve/Univ. Penn.
Weil P. 1990. Nonexpected utility in macroeconomics. Q. J. Econ. 105:29–42
Xia Y. 2001. Learning about predictability: the effects of parameter uncertainty on dynamic
asset allocation. J. Financ. 56:205–46
Zellner A. 1971. An Introduction to Bayesian Inference in Econometrics. New York, NY: Wiley &
Sons
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org
by University of Pennsylvania on 11/26/10. For personal use only.

206 Wachter
 Annual Review of
Financial Economics

Volume 2, 2010 Contents

Portfolio Theory: As I Still See It

Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org

Harry M. Markowitz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
by University of Pennsylvania on 11/26/10. For personal use only.

Bayesian Portfolio Analysis

Doron Avramov and Guofu Zhou . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Cross-Sectional Asset Pricing Tests
Ravi Jagannathan, Ernst Schaumburg, and Guofu Zhou . . . . . . . . . . . . . . 49
CEO Compensation
Carola Frydman and Dirk Jenter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Shareholder Voting and Corporate Governance
David Yermack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
An Informal Perspective on the Economics and Regulation of
Securities Markets
Chester S. Spatt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Privatization and Finance
William Megginson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Asset Allocation
Jessica A. Wachter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Investment Performance Evaluation
Wayne E. Ferson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Martingale Pricing
Kerry Back. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Limits of Arbitrage
Denis Gromb and Dimitri Vayanos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
Stochastic Processes in Finance
Dilip B. Madan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277

vi
 Ambiguity and Asset Markets
Larry G. Epstein and Martin Schneider . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Risk Management
Philippe Jorion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347

Errata
An online log of corrections to Annual Review of Financial Economics articles
may be found at http://financial.annualreviews.org
Annu. Rev. Fin. Econ. 2010.2:175-206. Downloaded from www.annualreviews.org
by University of Pennsylvania on 11/26/10. For personal use only.

Contents vii