Della Corte 2008

An Economic Evaluation of Empirical
Exchange Rate Models
Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

Pasquale Della Corte
University of Warwick
Lucio Sarno
University of Warwick, AXA Investment Managers, and Centre for Economic
Policy Research (CEPR)
Ilias Tsiakas
University of Warwick
This paper provides a comprehensive evaluation of the short-horizon predictive ability

of economic fundamentals and forward premiums on monthly exchange-rate returns in a
framework that allows for volatility timing. We implement Bayesian methods for estimation
and ranking of a set of empirical exchange rate models, and construct combined forecasts
based on Bayesian model averaging. More importantly, we assess the economic value of
the in-sample and out-of-sample forecasting power of the empirical models, and find two
key results: (1) a risk-averse investor will pay a high performance fee to switch from a
dynamic portfolio strategy based on the random walk model to one that conditions on
the forward premium with stochastic volatility innovations and (2) strategies based on
combined forecasts yield large economic gains over the random walk benchmark. These
two results are robust to reasonably high transaction costs. (JEL F31, F37, G11)
Forecasting exchange rates using models that condition on economically mean-

ingful variables has long been at the top of the research agenda in inter-
national finance, and yet empirical success remains elusive. Starting with
the seminal contribution of Meese and Rogoff (1983), a vast body of em-
pirical research finds that models that condition on economic fundamentals
cannot outperform a naive random walk model. Even though there is some
This paper was partly written while Lucio Sarno was a visiting scholar at the International Monetary Fund and
Norges Bank. The authors are indebted for useful conversations or constructive comments to Joel Hasbrouck
(editor), two anonymous referees, David Backus, Luc Bauwens, Sid Chib, Frank Diebold, Massimo Guidolin,
Rich Lyons, Michael Moore, Roel Oomen, Carol Osler, Dagfinn Rime, Simon van Norden, Arnold Zellner
as well as to participants at the 2007 INFER conference keynote speech; the 2007 workshop on “Trading
Strategies and Financial Market Inefficiency” at Imperial College London; the 2007 seminar on Bayesian
Inference in Econometrics and Statistics; the 2007 Royal Economic Society meetings; the 2006 European
Science Foundation workshop on FX Markets at the University of Warwick; the 2006 European meetings of
the Econometric Society; the 2006 Northern Finance Association meetings; the 2006 EC2 Conference on the
“Econometrics of Monetary Policy and Financial Decision Making”; and seminars at the European Central Bank
and the International Monetary Fund. Address correspondence to Lucio Sarno, Finance Group, Warwick Business
School, University of Warwick, Coventry CV4 7AL, UK; lucio.sarno@wbs.ac.uk. Other authors: Pasquale Della
Corte: pasquale.dellacorte@wbs.ac.uk. Ilias Tsiakas: ilias.tsiakas@wbs.ac.uk.

C The Author 2008. Published by Oxford University Press on behalf of The Society for Financial Studies.
All rights reserved. For Permissions, please e-mail: journals.permissions@oxfordjournals.org.
doi:10.1093/rfs/hhn058 Advance Access publication June 13, 2008
The Review of Financial Studies / v 22 n 9 2009
evidence that exchange rates and fundamentals comove over long horizons [e.g.,
Mark (1995) and Mark and Sul (2001)], the prevailing view in international
finance research is that exchange rates are not predictable, especially at short

horizons.
A separate yet related literature finds that forward exchange rates contain
valuable information for predicting spot exchange rates. In theory, the relation
between spot and forward exchange rates is governed by the uncovered interest
parity (UIP) condition, which suggests that the forward premium must be per-
fectly positively related to future exchange rate changes. In practice, however,
this is not the case as we empirically observe a negative relation.1 The result
of the empirical failure of UIP is that conditioning on the forward premium
often generates exchange rate predictability. For example, Backus, Gregory,
and Telmer (1993) and Backus, Foresi, and Telmer (2001) explore this further
and find evidence of predictability using the lagged forward premium as a
predictive variable. Furthermore, Clarida et al. (2003, 2006) and Boudoukh,
Richardson, and Whitelaw (2006) show that the term structure of forward ex-
change (and interest) rates contains valuable information for forecasting spot
exchange rates.
On the methodology side, while there is extensive literature on statistical
measures of the accuracy of exchange rate forecasts, there is little work assess-
ing the economic value of exchange rate predictability. Relevant research to
date comprises an early study by West, Edison, and Cho (1993), which pro-
vides a utility-based evaluation of exchange rate volatility, and more recently,
Abhyankar, Sarno, and Valente (2005), who use a similar method for inves-
tigating long-horizon exchange rate predictability. However, in the context of
dynamic asset allocation strategies, there is no study assessing the economic
value of the predictive ability of empirical exchange rate models that condition
on economic fundamentals or the forward premium while allowing for volatility
timing.
Our empirical investigation attempts to fill this gap and connect the related
literatures that examine the performance of empirical exchange rate models.
On the one hand, we directly investigate whether the statistical rejection of
UIP generates economic value to a dynamically optimizing investor, who ex-
ploits the UIP violation in order to generate excess returns. On the other hand,
our economic evaluation provides a novel way for confirming (or not) the
underwhelming performance of exchange rate models conditioning on eco-
nomic fundamentals that has so far been established by statistical tests. We
do this by employing a range of economic and Bayesian statistical criteria for
performing a comprehensive assessment of the short-horizon, in-sample and
out-of-sample, predictive ability of three sets of models for the conditional
mean of monthly nominal exchange rate returns. These models include the
1
See, for example, Bilson (1981); Fama (1984); Froot and Thaler (1990); and Backus, Foresi, and Telmer (2001).
For a survey of this literature, see Lewis (1995) and Engel (1996) and the references therein.
3492
Economic Evaluation of Empirical Exchange Rate Models
naive random walk model, the monetary fundamentals model (in three vari-
ants), and the spot-forward regression model. Each of the models is studied
under three volatility specifications: constant variance (standard linear regres-

sion), GARCH(1,1), and stochastic volatility (SV). In total, we evaluate the
performance of 15 specifications, which encompass the most popular empirical
exchange rate models studied in prior research. Our analysis employs monthly
returns data ranging from January 1976 to December 2004 for three major US
dollar exchange rates: the UK pound sterling, the Deutsche mark/euro, and the
Japanese yen.
An important contribution of our analysis is the use of economic criteria.
Statistical evidence of exchange rate predictability in itself does not guarantee
that an investor can earn profits from an asset allocation strategy that exploits
this predictability. In practice, ranking models is useful to an investor only
if it leads to tangible economic gains. Therefore, we assess the economic
value of exchange rate predictability by evaluating the impact of predictable
changes in the conditional foreign exchange (FX) returns and volatility on
the performance of dynamic allocation strategies. We employ mean-variance
analysis as a standard measure of portfolio performance and apply quadratic
utility, which allows us to quantify how risk aversion affects the economic
value of predictability, building on empirical studies of volatility timing in
stock returns by Fleming, Kirby, and Ostdiek (2001) and Marquering and
Verbeek (2004).2 Ultimately, we measure how much a risk-averse investor is
willing to pay for switching from a dynamic portfolio strategy based on the
random walk model to one that conditions on either monetary fundamentals or
the forward premium and has a dynamic volatility specification.
The design of the dynamic allocation strategies is based on the mean-variance
setting of West, Edison, and Cho (1993), which adopts quadratic utility and
fixes the investor’s degree of relative risk aversion to a constant value. Quadratic
utility allows us to compute in closed form the utility gains from using the condi-
tional mean and volatility forecasts of one model rather than another. Combined
with the approach of Fleming, Kirby, and Ostdiek (2001), it is then straight-
forward to compute the performance fees, and hence provide an economically
meaningful ranking of competing models for a given degree of relative risk aver-
sion. Despite the well-known shortcomings affecting quadratic utility, there are
a number of reasons that make it an appealing assumption, which we discuss
in detail later in the paper. For example, quadratic utility is necessary to justify
mean-variance optimization for nonnormal return distributions, and therefore
allows the economic evaluation of a larger universe of models within mean-
variance. More importantly, there is evidence that quadratic utility provides a
2
For studies of asset return predictability following this approach, see also Kandell and Stambaugh (1996);
Barberis (2000); Baks, Metrick, and Wachter (2001); Bauer (2001); Shanken and Tamayo (2001); Avramov
(2002); Cremers (2002); and Della Corte, Sarno, and Thornton (2008). Karolyi and Stulz (2003) provide a survey
of asset allocation in an international context.
3493
highly satisfactory approximation to a wide range of more sophisticated utility

functions (e.g., Hlawitschka, 1994).
We assess the statistical evidence on exchange rate predictability in a

Bayesian framework, which requires a choice for the prior distribution of
the model parameters. For example, in the case of the simple linear regres-
sion model, we assume independent Normal-Gamma prior distributions. We
rank the competing model specifications by computing the posterior prob-
ability of each model. The posterior probability is based on the marginal
likelihood and hence it accounts for parameter uncertainty, while imposing
a penalty for lack of parsimony (higher dimension). In the context of this
Bayesian methodology, an alternative approach to determining the best model
available is to form combined forecasts, which exploit information from the
entire universe of model specifications under consideration. Specifically, we
implement the Bayesian model averaging (BMA) method, which weighs all
conditional mean and volatility forecasts by the posterior probability of each
model.
To preview our key results, we find strong economic and statistical evidence
against the naive random walk benchmark with constant variance innovations.
In particular, while we confirm that conditioning on monetary fundamentals
has no economic value either in-sample or out-of-sample, a key result of the
paper is that the predictive ability of forward exchange rate premiums has
substantial economic value in a dynamic allocation strategy. Also, stochastic
volatility significantly outperforms the constant variance and GARCH(1,1)
models irrespective of the conditional mean specification. This leads to the
conclusion that the best empirical exchange rate model is a model that exploits
the information in the forward market for the prediction of conditional exchange
rate returns and allows for stochastic volatility for the prediction of exchange
rate volatility. We also provide evidence that combined forecasts, which are
formed using BMA, substantially outperform the random walk benchmark.
These results are robust to reasonably high transaction costs and hold for all
currencies both in-sample and out-of-sample. Finally, these findings have clear
implications for international asset allocation strategies that are subject to FX
risk.
The remainder of the paper is organized as follows. In the next section,
we briefly review the literature on exchange rate predictability conditioning
on either fundamentals or forward exchange premiums. Section 2 lays out
the competing empirical models for the conditional mean and volatility of
exchange rate returns. In Section 3, we discuss the framework for assessing
the economic value of exchange rate predictability for a risk-averse investor
with a dynamic portfolio allocation strategy. Section 4 provides a sketch of
the Bayesian estimation tools, discusses the approach to model selection, and
explains the construction of combined forecasts using the BMA method. Our
3494
empirical results are reported in Section 5, followed by robustness checks in

Section 6. Finally, Section 7 concludes.

1. Stylized Facts on Exchange Rate Predictability
In this section, we briefly review the theoretical and empirical research that
motivates our conditioning on lagged monetary fundamentals and forward
premiums in the set of empirical exchange rate models.
1.1 Exchange rates and monetary fundamentals
There is extensive literature in international finance that studies the relation
between nominal exchange rates and monetary fundamentals and focuses on
the following predictive variable, xt :
x t = z t − st , (1)

z t = m t − m ∗t − yt − yt∗ , (2)
where st is the log of the nominal exchange rate (defined as the domestic price
of foreign currency); m t is the log of the money supply; yt is the log of national
income; and asterisks denote variables of the foreign country. Note that long-run
money neutrality and income homogeneity are imposed, with the coefficients
on m t − m ∗t and yt − yt∗ both set to unity, as predicted by conventional theories
of exchange rate determination, and z t represents the relative velocity between
the two countries in question. The relation between the exchange rates and
fundamentals defined in Equations (1) and (2) suggests that a deviation of the
nominal exchange rate st from its long-run equilibrium level determined by
the fundamentals z t (i.e., xt = 0) requires the exchange rate to move in the
future so as to converge towards its long-run equilibrium. In other words, the
deviation xt has predictive power on future realizations of the exchange rate.3
Despite the appeal of the theoretical relation between exchange rates and
fundamentals, the empirical evidence is mixed. On the one hand, short-run
exchange rate variability appears to be disconnected from the underlying fun-
damentals (Mark, 1995) in what is commonly referred to as the “exchange rate
disconnect puzzle.” On the other hand, some recent empirical research finds
that fundamentals and nominal exchange rates move together in the long run
(Groen, 2000; Mark and Sul, 2001; Rapach and Wohar, 2002; Sarno, Valente,
and Wohar, 2004). Either way, our study contributes to the empirical literature
on the predictive ability of monetary fundamentals on exchange rates by pro-
viding an economic evaluation of the in-sample and out-of-sample forecasting
power of fundamentals at a short (one-month ahead) horizon.
3
The specification of fundamentals in Equation (2) is common in the relevant empirical literature (e.g., Mark,
1995; Mark and Sul, 2001). Theories of exchange rate determination view z t as the core set of economic
fundamentals that determine the long-run equilibrium exchange rate. These theories include traditional models
based on aggregate demand functions (e.g., Mark, 1995; Engel and West, 2005), and representative-agent general
equilibrium models (e.g., Lucas, 1982; Obstfeld and Rogoff, 1995).
3495
1.2 The spot-forward exchange rate relation

Assuming risk neutrality and rational expectations, UIP is the cornerstone
condition for FX market efficiency. For a one-period horizon, UIP is represented

by the following equation:
∗
E t−1 st = i t−1 − i t−1 , (3)
∗
where i t−1 and i t−1 are the one-period domestic and foreign nominal interest
rates, respectively; and st ≡ st − st−1 .
In the absence of riskless arbitrage, covered interest parity (CIP) holds and
implies
∗
f t−1 − st−1 = i t−1 − i t−1 , (4)
where f t−1 is the log of the one-period forward exchange rate (i.e., the rate
agreed now for an exchange of currencies in one period). Substituting the
∗
interest rate differential i t−1 − i t−1 in Equation (3) by the forward premium
(or forward discount) f t−1 − st−1 , we can estimate the following regression,
which is commonly referred to as the “Fama regression,” (Fama, 1984):
st = α + β ( f t−1 − st−1 ) + u t , (5)
where u t is a disturbance term.

If UIP holds, we should find that α = 0, β = 1, and the disturbance term u t
is uncorrelated with information available at time t − 1. Despite the increasing
sophistication of the econometric techniques implemented and the improving
quality of the data sets utilized, empirical studies estimating the Fama (1984)
regression consistently reject the UIP condition (Hodrick, 1987; Lewis, 1995;
Engel, 1996). As a result, it is now a stylized fact that estimates of β tend
to be closer to minus unity than plus unity (Froot and Thaler, 1990). The
negative value of β is the defining feature of what is commonly referred to as
the “forward bias puzzle,” namely the tendency of high-interest currencies to
appreciate when UIP would predict them to depreciate.4
Attempts to explain the forward bias puzzle using models of risk premiums
have met with limited or mixed success, especially for plausible degrees of risk
aversion [e.g., Engel (1996) and the references therein]. Moreover, it has proved
difficult to explain the rejection of UIP by resorting to a range of proposed ex-
planations, including learning, peso problems, and bubbles (e.g., Lewis, 1995);
consumption-based asset pricing theories, which allow for departures from
both time-additive preferences (Backus, Gregory, and Telmer, 1993; Bansal
4
Exceptions to this puzzle include Bansal (1997), who finds that the forward bias is related to the sign of the
interest rate differential; Bansal and Dahlquist (2000), who document that the forward bias is largely confined
to developed economies and countries where the interest rate is lower than in the United Sates; Bekaert and
Hodrick (2001), who provide a “partial rehabilitation” of UIP by accounting for small-sample distortions; and
Lustig and Verdelhan (2007), who attempt to explain the forward bias puzzle focusing on the cross section of
foreign currency risk premiums.
3496
et al., 1995; and Bekaert, 1996) and from expected utility (Bekaert, Hodrick,
and Marshall, 1997); and using popular models of the term structure of interest
rates adapted to a multicurrency setting (Backus, Foresi, and Telmer, 2001). In

conclusion, even with the benefit of 20 years of hindsight, the forward bias has
not been convincingly explained and remains a puzzle in international finance
research.
In this context, the objective of this paper is neither to find a novel resolution
to the forward bias puzzle nor to discriminate among competing explanations.
Instead, we focus on predicting short-horizon exchange rate returns when con-
ditioning on the lagged forward premium, thus empirically exploiting the for-
ward bias reported in the strand of literature stemming from Bilson (1981);
Fama (1984); Bekaert and Hodrick (1993); and Backus, Gregory, and Telmer
(1993). For example, Bilson (1981) argues that regressions conditioning on the
forward premium can potentially yield substantial economic returns, whereas
arguments based on limits to speculation would suggest otherwise (Lyons,
2001; Sarno, Valente, and Leon, 2006). Furthermore, term structure models
that exploit departures from UIP often yield accurate out-of-sample forecasts
(e.g., Clarida and Taylor, 1997; Clarida et al., 2003; and Boudoukh, Richardson,
and Whitelaw, 2006). However, little attention has been given to the question of
whether the statistical rejection of UIP and the forward bias resulting from the
negative estimate of β offers economic value to an international investor facing
FX risk. Our paper fills this void in the literature by assessing the economic
value of the predictive ability of empirical exchange rate models that condition
on the forward premium in the context of dynamic asset allocation strategies.
2. Modeling FX Returns and Volatility

In this section, we present the candidate models applied to monthly exchange
rate returns in our study of short-horizon exchange rate predictability. We use
a set of specifications for the dynamics of both the conditional mean and
volatility, which are set against the naive random walk benchmark. In short, we
estimate five conditional mean and three conditional volatility specifications
yielding a total of 15 models for each of the three dollar exchange rates under
consideration.
2.1 The conditional mean

We examine five conditional mean specifications in which the dynamics of
exchange rate returns are driven by the following regression:
st = α + βxt−1 + u t , u t = vt εt , εt ∼ N I D (0, 1) . (6)
Our first specification is the naive random walk (RW) model, which sets β = 0.
This model is the standard benchmark in the literature on exchange rate pre-
dictability since the seminal work of Meese and Rogoff (1983).
3497
The next three model specifications condition on monetary fundamentals

(MF). Specifically, MF1 uses the canonical version xt = z t − st as defined in
Equations (1) and (2). This is the most common formulation of the monetary

fundamentals model since Mark (1995). The second variant MF2 corrects for the
deterministic component in the deviation of the exchange rate from monetary
fundamentals by allowing for an intercept and a slope parameter; in other
words, we run the ordinary least squares (OLS) regression st = κ0 + κ1 z t + ζt ,
and set xt = − ζt . The third variant MF3 further corrects for the time trend in
fundamentals deviations; in this case, we run the OLS regression st = κ0 +
κ1 z t + κ2 t + ζt , where t is a simple time trend, and again we set xt = − ζt .5
Finally, the fifth conditional mean specification is the forward premium (FP)
model, which sets xt = f t − st as in Equation (5) resulting in the Fama (1984)
regression. The FP model stems directly from the spot-forward exchange rate
relation derived from UIP. Hence it constitutes the empirical model that exploits
the forward bias and allows us to assess the economic value of conditioning on
the forward premium in the context of dynamic asset allocation strategies. The
forward bias (a negative estimate of the β coefficient in the FP model) implies
that the more the foreign currency at a premium in the forward market, the less
the home currency expected to depreciate. Equivalently, the more domestic
interest rates exceed foreign interest rates, the more the domestic currency
tends to appreciate over the holding period.
2.2 The conditional variance

We model the dynamics of the conditional variance by implementing three
models: the simple linear regression (LR), the GARCH(1,1) model, and the
stochastic volatility (SV) model. The LR framework simply assumes that the
conditional variance of FX return innovations is constant over time (vt2 = v 2 ),
and therefore presents the benchmark against which models with time-varying
conditional variance will be evaluated.
The benchmark GARCH(1,1) model of Bollerslev (1986) is defined as
vt2 = ω + γ1 vt−1
2
+ γ2 u 2t−1 . (7)
Our motivation for studying the simple GARCH(1,1) model is based on the
early study of West, Edison, and Cho (1993), which conducts a utility-based
evaluation of exchange rate volatility and finds that GARCH(1,1) is the best
performing model.
Stochastic volatility models are similar to the GARCH process in that they
capture the persistent and hence predictable component of volatility. Unlike
5
The motivation behind the MF2 and MF3 variants derives from empirical evidence that cointegration between st
and z t may be established only by correcting for the deterministic components (either a constant or a constant
and a time trend) in the cointegrating residual (e.g., Rapach and Wohar, 2002). Note, however, that in the out-of-
sample exercise we estimate the deterministic component recursively as we move through the data sample, and
hence our results do not suffer from “look-ahead bias.”
3498
GARCH models, however, the assumption of a stochastic second moment in-

troduces an additional source of risk.6 According to the plain vanilla SV model,
the persistence of the conditional volatility vt is captured by the dynamics of

the Gaussian stochastic log-variance process h t :
vt = exp (h t /2) , (8)

h t = μ + φ (h t−1 − μ) + σηt , ηt ∼ N I D (0, 1) . (9)
3. Measuring the Economic Value of Exchange Rate Predictability

This section discusses the framework we use in order to evaluate the impact
of predictable changes in both exchange rate returns and volatility on the
performance of dynamic allocation strategies.
3.1 FX models in a dynamic mean-variance framework

In mean-variance analysis, the maximum expected return strategy leads to
a portfolio allocation on the efficient frontier. Consider an investor who
has a one-month horizon and constructs a dynamically rebalanced portfolio
that maximizes the conditional expected return subject to achieving a tar-
get conditional volatility. Computing the time-varying weights of this portfo-
lio requires one-step ahead forecasts of the conditional mean and the con-
ditional variance-covariance matrix. Let rt+1 denote the K × 1 vector of
risky asset returns, μt+1|t = E t [rt+1 ] is the conditional expectation of rt+1 ,
and t+1|t = E t [(rt+1 − μt+1|t )(rt+1 − μt+1|t ) ] is the conditional variance-
covariance matrix of rt+1 . At each period t, the investor solves the following
problem:

max μ p,t+1|t = wt μt+1|t + 1 − wt ι r f ,
wt
2
s.t. σ∗p = wt t+1|t wt , (10)
where wt is the K × 1 vector of portfolio weights on the risky assets, ι is a

K × 1 vector of ones, μ p,t+1|t is the conditional expected return of the portfolio,
σ∗p is the target conditional volatility of the portfolio returns, and r f is the return
on the riskless asset. The solution to this optimization problem delivers the risky
asset weights:
σ∗p −1
wt = √ t+1|t μt+1|t − ιr f , (11)
Ct
6
Market microstructure theories of speculative trading (e.g., Tauchen and Pitts, 1983; Andersen, 1996) provide
rigorous arguments for modeling volatility as stochastic. For details on SV models, see Kim, Shephard, and Chib
(1998) and Chib, Nardari, and Shephard (2002). For an application of SV models to exchange rates, see Harvey,
Ruiz, and Shephard (1994). Finally, for a comparison between GARCH and SV models, see Fleming and Kirby
(2003).
3499
−1
where Ct = (μt+1|t − ιr f ) t+1|t (μt+1|t − ιr f ). The weight on the riskless as-

set is 1 − wt ι.
Constructing the optimal portfolio weights requires estimates of the condi-

tional expected returns, variances, and covariances. We consider five condi-
tional mean strategies (RW, MF1 , MF2 , MF3 , and FP) and three conditional
volatility strategies (LR, GARCH, and SV) for a total of 15 sets of one-step
ahead conditional expected return and volatility forecasts. The conditional co-
variances are computed using the constant conditional correlation (CCC) model
of Bollerslev (1990), in which the dynamics of covariances are driven by the
time-variation in the conditional volatilities.7 By design, in this setting the op-
timal weights will vary across models only to the extent that forecasts of the
conditional mean and volatility will vary, which is precisely what the empirical
models provide. The benchmark against which we compare the model speci-
fications is the random walk model with constant variance (RWLR ). In short,
our objective is to determine whether there is economic value in (1) condi-
tioning on monetary fundamentals and, if so, which of the three specifications
works best; (2) conditioning on the forward premium; (3) using a GARCH
volatility specification; and (4) implementing an SV process for the monthly
FX innovations.
3.2 Quadratic utility

Mean-variance analysis is a natural framework for assessing the economic value
of strategies that exploit predictability in the mean and variance. In particular,
we rank the performance of the competing FX models using the West, Edison,
and Cho (1993) methodology, which is based on mean-variance analysis with
quadratic utility. The investor’s realized utility in period t + 1 can be written
as
λ 2 λWt2 2
U (Wt+1 ) = Wt+1 − Wt+1 = Wt R p,t+1 − R p,t+1 , (12)
2 2
where Wt+1 is the investor’s wealth at t + 1, λ determines his risk preference,
and

R p,t+1 = 1 + r p,t+1 = 1 + 1 − wt ι r f + wt rt+1 (13)
is the period t + 1 gross return on his portfolio.

We quantify the economic value of exchange rate predictability by setting the
investor’s degree of relative risk aversion (RRA) δt = λWt / (1 − λWt ) equal
to a constant value δ. In this case, West, Edison, and Cho (1993) demonstrate
that one can use the average realized utility, U (·), to consistently estimate the
7
In notation local to this footnote, the CCC model of Bollerslev (1990) specifies the covariances as follows:
σi j,t = σi,t σ j,t ρi j , where σi,t and σ j,t are the conditional volatilities implied by either the GARCH(1,1) or the
SV process, and ρi j is the constant sample correlation coefficient. Note that for the out-of-sample results, we
use a rolling correlation estimate updated every time a new observation is added. From a numerical standpoint,
implementing the CCC model is attractive because it eliminates the possibility of t+1|t not being positive
definite.
3500
expected utility generated by a given level of initial wealth. Specifically, the

average utility for an investor with initial wealth W0 is equal to
T −1
δ

U (·) = W0 R p,t+1 − R2 . (14)
t=0
2(1 + δ) p,t+1
Average utility depends on taste for risk. In the absence of restrictions on δ,

quadratic utility exhibits increasing RRA. This is counterintuitive since, for in-
stance, an investor with increasing RRA becomes more averse to a percentage
loss in wealth when his wealth increases. As in West, Edison, and Cho (1993)
and Fleming, Kirby, and Ostdiek (2001), fixing δ implies that expected utility
is linearly homogeneous in wealth: double wealth and expected utility doubles.
Hence we can standardize the investor problem by assuming W0 = $1. Further-
more, by fixing δ rather than λ, we are implicitly interpreting quadratic utility
as an approximation to a nonquadratic utility function, with the approximating
choice of λ dependent on wealth. The estimate of expected quadratic utility
given in Equation (14) is used to implement the Fleming, Kirby, and Ostdiek
(2001) framework for assessing the economic value of our FX strategies in the
context of dynamic asset allocation.
A critical aspect of mean-variance analysis is that it applies exactly only
when the return distribution is normal or the utility function is quadratic.
Hence, the use of quadratic utility is not necessary to justify mean-variance
optimization. For instance, one could instead consider using utility functions
belonging to the constant relative risk aversion (CRRA) class, such as power
or log utility. However, quadratic utility is an attractive assumption because
it allows us to consider nonnormal distributions of returns, while remaining
within the mean-variance framework, as well as providing a high degree of
analytical tractability. Absent Gaussianity, quadratic utility is needed to justify
mean-variance and allows us to use the Fleming, Kirby, and Ostdiek (2001)
framework (also based on quadratic utility) for evaluating the performance of
fat-tailed volatility specifications, such as the tGARCH model of Bollerslev
(1987).
Additionally, quadratic utility may be viewed as a second-order Taylor series
approximation to expected utility. In an investigation of the empirical robustness
of the quadratic approximation, Hlawitschka (1994) finds that a two-moment
Taylor series expansion “may provide an excellent approximation” (p. 713) to
expected utility and concludes that the ranking of common stock portfolios
based on two-moment Taylor series is “almost exactly the same” (p. 714) as
the ranking based on a wide range of utility functions.
3.3 Performance measures

At any point in time, one set of estimates of the conditional mean and variance
is better than a second set if investment decisions based on the first set lead to
higher average realized utility U . Alternatively, the optimal model requires less
3501
wealth to yield a given level of U than a suboptimal model. Following Fleming,

Kirby, and Ostdiek (2001), we measure the economic value of our FX strategies
by equating the average utilities for selected pairs of portfolios. Suppose, for

example, that holding a portfolio constructed using the optimal weights based
on the random walk/linear regression (RWLR ) model yields the same average
utility as holding the forward premium/stochastic volatility ( FPSV ) optimal
portfolio, which is subject to monthly expenses , expressed as a fraction
of wealth invested in the portfolio. Since the investor would be indifferent
between these two strategies, we interpret as the maximum performance fee
he will pay to switch from the RWLR to the FPSV strategy. In other words, this
utility-based criterion measures how much a mean-variance investor is willing
to pay for conditioning on the forward premium under stochastic volatility
innovations. The performance fee will depend on the investor’s degree of risk
aversion. To estimate the fee, we find the value of that satisfies
T −1
∗ δ ∗ 2
R p,t+1 − − R p,t+1 −
t=0
2 (1 + δ)
T −1

δ
= R p,t+1 − R 2p,t+1 , (15)
t=0
2 (1 + δ)
where R ∗p,t+1 is the gross portfolio return constructed using the expected return
and volatility forecasts from the FPSV model, and R p,t+1 is implied by the
benchmark RWLR model.
In the context of mean-variance analysis, a commonly used measure of
economic value is the Sharpe ratio. However, as suggested by Marquering and
Verbeek (2004) and Han (2006), the Sharpe ratio can be misleading because
it severely underestimates the performance of dynamic strategies. Specifically,
the realized Sharpe ratio is computed using the sample standard deviation of
the realized portfolio returns and hence it overestimates the conditional risk
an investor faces at each point in time. Furthermore, the Sharpe ratio cannot
quantify the exact economic gains of the dynamic strategies over the static
random walk strategy in the direct way of the performance fees. Therefore,
our economic analysis of short-horizon exchange rate predictability focuses
primarily on performance fees, while Sharpe ratios of selected models are
reported in the robustness section.8
3.4 The dynamic FX strategies

In this mean-variance quadratic-utility framework, we design the following
global strategy. Consider a US investor who builds a portfolio by allocating his
8
The annualized Sharpe ratios reported in Table 10 are adjusted for the serial correlation in the monthly portfolio
returns generated by the dynamic strategies. Specifically, following Lo (2002), we multiply the monthly Sharpe
ratios by the adjustment factor 12
11
, where ρk is the autocorrelation coefficient of portfolio returns
12+2 k=1 (12−k)ρk
at lag k.
3502
wealth between four bonds: one domestic (United States), and three foreign
bonds (UK, Germany, and Japan). At the beginning of each month, the four
bonds yield a riskless return in local currency. Hence the only risk the US

investor is exposed to is the FX risk. Each month the investor takes two steps.
First, he uses each of the 15 models to forecast the one-month ahead conditional
mean and volatility of the exchange rate returns. Second, conditional on the
forecasts of each model, he dynamically rebalances his portfolio by computing
the new optimal weights for the maximum return strategy. This setup is designed
to inform us whether using one particular conditional mean and volatility
specification affects the performance of a short-horizon allocation strategy in
an economically meaningful way. The yields of the riskless bonds are proxied
by monthly eurodeposit rates.
In the context of this maximum return dynamic strategy, we compute both
the in-sample and the out-of-sample performance fee, , where the out-of-
sample period starts in January 1990 and ends in December 2004. Further-
more, we compare the performance fees for the combinations corresponding
to the following cases: (1) three sets of target annualized portfolio volatilities
σ∗p = {8%, 10%, 12%}; (2) all pairs of 15 models (FPSV versus RWLR ); and
(3) degrees of RRA δ = {2, 6}. We report the estimates of as annualized fees
in basis points.9
3.5 Transaction costs

The impact of transaction costs is an essential consideration in assessing the
profitability of trading strategies. This is especially true in our case because the
trading strategy based on the random walk benchmark is static (independent
of state variables), whereas the remaining empirical models generate dynamic
strategies.10 Furthermore, making an accurate determination of the size of
transaction costs is difficult because it involves three factors: (1) the type of
investor (e.g., individual versus institutional investor); (2) the value of the
transaction; and (3) the nature of the broker (e.g., brokerage firm versus direct
Internet trading). This difficulty is reflected in the wide range of estimates used
in empirical studies. For example, Marquering and Verbeek (2004) consider
three levels of transaction costs, 0.1%, 0.5%, and 1%, to represent low, medium,
and high costs.
Our approach avoids these concerns by calculating the break-even transaction
cost τBE that renders investors indifferent between two strategies (e.g., Han,
2006). Hence, we assume that transaction costs equal a fixed proportion (τ)
of the value traded in each bond: τ|wt − wt−1 1+r 1+rt
p,t
|. In comparing a dynamic
strategy with the static (random walk) strategy, an investor who pays transaction
9
The in-sample period in our economic value results starts in January 1979 due to lack of data for the Japanese
Eurocurrency interest rate. In contrast, for the statistical analysis, the in-sample period starts in January 1976.
10
The random walk model (RWLR ) is the only empirical model that assumes constant mean and variance. Therefore,
the in-sample optimal weights for the RWLR trading strategy remain constant over time. However, the out-of-
sample optimal weights will vary because every month we reestimate the drift and variance of the RWLR model.
3503
costs lower than τBE will prefer the dynamic strategy. We report τBE in monthly
basis points.11

4. Estimation and Forecasting
4.1 Bayesian Markov chain Monte Carlo estimation

Stochastic volatility models are generally less popular in empirical applica-
tions than GARCH despite their parsimonious structure, intuitive appeal, and
popularity in theoretical asset pricing. This is primarily due to the numerical dif-
ficulty associated with estimating SV models using standard likelihood-based
methods because the likelihood function is not available analytically. Bayesian
estimation offers a substantial computational advantage over any classical ap-
proach because it avoids tackling difficult numerical optimization procedures.
In this context, we estimate all three volatility frameworks (LR, GARCH, and
SV) using similar Bayesian Markov chain Monte Carlo (MCMC) estimation
algorithms. This is a crucial aspect of our econometric analysis because it ren-
ders the posterior mean estimates directly comparable across the three volatility
structures. It also allows us to use the same model risk diagnostics for all model
specifications. Finally, a distinct advantage of Bayesian inference is that it pro-
vides the posterior distribution of a regression coefficient conditional on the
data, which holds for finite samples and regardless of whether exchange rates
(and fundamentals) are (co)integrated (e.g., Sims, 1988).12
We estimate the parameters of the SV model using the Bayesian MCMC
algorithm of Chib, Nardari, and Shephard (2002), which builds on the proce-
dures developed by Kim, Shephard, and Chib (1998). The algorithm constructs
a Markov chain whose limiting distribution is the target posterior density of the
SV parameters. The Markov chain is a Gibbs sampler in which all parameters
are drawn sequentially from their full conditional posterior distribution. The
Gibbs sampler is iterated 5000 times and the sampled draws, beyond a burn-
in period of 1000 iterations, are treated as variates from the target posterior
distribution. We design a similar Bayesian MCMC algorithm for estimating
the GARCH(1,1) parameters, which also draws from the insights of Vrontos,
Dellaportas, and Politis (2000). The Bayesian linear regression algorithm im-
plements a simple MCMC assuming an independent Normal-Gamma prior
distribution (for details see Koop, 2003). The MCMC algorithm for each of the
three volatility models is summarized in the Appendix.
The mean of the MCMC parameter draws is an asymptotically efficient esti-
mator of the posterior mean of θ (see Geweke, 1989). The numerical standard
11
In contrast to , which is reported in annual basis points, τBE is reported in monthly basis points because τBE is
a proportional cost paid every month when the portfolio is rebalanced.
12
This is not the case in classical inference, where the small samples, typically employed in the study of exchange
rate predictability combined with the assumption that exchange rates and fundamentals are cointegrated, can
have a critical impact in overstating predictability (e.g., Berkowitz and Giorgianni, 2001).
3504
error (NSE) is the square root of the asymptotic variance of the MCMC esti-
mator:
⎧ ⎫
1 ⎨ ⎬

BI
NSE = ψ0 + 2 j ,
K (z) ψ (16)
I ⎩ ⎭
j=1
where I = 5000 is the number of iterations (beyond the initial burn-in of 1000
j is
iterations), j = 1, . . . , B I = 500 lags is the set bandwidth, z = BjI , and ψ
the sample autocovariance of the MCMC draws for each estimated parameter
cut according to the Parzen kernel K (z).
The likelihood function of the SV models is not available analytically, and
hence must be simulated. The log-likelihood function is evaluated under the
predictive density as

T
T
log
L= log
f (st |t−1 , θ) = log
f t (st | h t , θ) , (17)
t=1 t=1
where θ is taken as the posterior mean estimate from the MCMC simulations.
The key to this calculation is simulating the one-step ahead predictive log-
variance h t |t−1 , θ, which is a nontrivial task as it is sampled using the particle
filter of Pitt and Shephard (1999). The particle filter is summarized in the
Appendix.
4.2 Model risk and posterior probability

Model risk arises from the uncertainty over selecting a model specifica-
tion. Consistent with our Bayesian approach, a natural statistical criterion for
resolving this uncertainty is the posterior probability of each model. Hence, we
rank the competing models using the posterior probability, which has three im-
portant advantages relative to the log-likelihood: (1) it is based on the marginal
likelihood and therefore accounts for parameter uncertainty; (2) it imposes a
penalty for lack of parsimony (higher dimension); and (3) it forms the basis of
the BMA strategy discussed below. Ranking the models using the highest pos-
terior probability is equivalent to choosing the best model in terms of density
forecasts and is a robust model selection criterion in the presence of misspecifi-
cation and non-nested models (e.g., Fernandez-Villaverde and Rubio-Ramirez,
2004).
Consider a set of N models M1 , . . . , M N . We form a prior belief π (Mi ) on
the probability that the ith model is the true model, observe the FX returns
data s, and then update our belief that the ith model is true by computing the
posterior probability of each model defined as follows:
p(s | Mi )π(Mi )
p(Mi | s) = N
, (18)
j=1 p(s | M j )π(M j )
3505
where p (s | Mi ) is the marginal likelihood of the ith model defined as fol-
lows:

p(s | Mi ) = p(s, θ | Mi )dθ = p(s | θ, Mi )π(θ | Mi )dθ. (19)

θ θ
In Equation (18), we set our prior belief to be that all models are equally likely
(i.e., π (Mi ) = N1 ).
Note that the marginal likelihood is an averaged (not a maximized) like-
lihood. This implies that the posterior probability is an automatic “Occam’s
Razor” in that it integrates out parameter uncertainty.13 Furthermore, the
marginal likelihood is simply the normalizing constant of the posterior density
and (suppressing the model index for simplicity) it can be written as
f (s | θ) π (θ)
p (s) = , (20)
π (θ | s)
where f (s | θ) is the likelihood, π (θ) the prior density of the parameter
vector θ, π (θ | s) the posterior density, and θ is evaluated at the posterior
mean. Since θ is drawn in the context of MCMC sampling, the posterior density
π (θ | s) is computed using the technique of reduced conditional MCMC runs
of Chib (1995) and Chib and Jeliazkov (2001).
4.3 Combined forecasts

Assessing the predictive ability of empirical exchange rate models primarily
involves a pairwise comparison of the competing models. However, given that
we do not know which one of the models is true, it is important that we assess the
performance of combined forecasts proposed by the seminal work of Bates and
Granger (1969). Specifically, we design two strategies based on a combination
of forecasts for both the conditional mean and volatility of exchange rate
returns: the BMA strategy and the Bayesian winner (BW) strategy.14
We assess the economic value of combined forecasts by treating the BMA
and BW strategies in the same way as any of the 15 individual empirical
models. For instance, we compute the performance fee for the BMA one-
month ahead forecasts and compare it to the random walk benchmark. We
focus on two distinct universes of models: the restricted universe of the five SV
models (because the five conditional mean specifications with SV innovations
have the highest marginal likelihood), and the unrestricted universe of all 15
empirical exchange rate models.
13
Occam’s Razor is the principle of parsimony, which states that among two competing theories that make exactly
the same prediction, the simpler one is best.
14
See Diebold and Pauly (1990); Diebold (1998, 2004); and Timmermann (2006) for a review of forecast com-
binations. A previous version of the paper also considers a deterministic model average (DMA) method, which
involves taking an equally-weighted average of the conditional mean and volatility forecasts from a given uni-
verse of available models; we find that the BMA and BW combination methods outperform the DMA method
(results available upon request).
3506
4.3.1 The BMA strategy. In the context of our Bayesian approach, it is

natural to implement the BMA method originally discussed in Leamer (1978)
and surveyed in Hoeting et al. (1999). The BMA strategy accounts directly for

uncertainty in model selection, and is in fact easy to implement once we have
the output from the MCMC simulations. Define f i,t as the forecast density of
each of the N competing models at time t. Then, the BMA forecast density is
given by

N
f tBMA = pt (Mi | st ) f i,t , (21)
i=1
where pt (Mi | st ) is the posterior probability of model Mi given the data
st .
It is important to note that (1) the BMA weights vary not only across models
but also across time periods, as does the marginal likelihood of each model, and
(2) we evaluate the BMA strategy ex ante. We do this by lagging the posterior
probability of each model for the following reason. Suppose that we need to
compute the period t BMA forecasts of the conditional mean and volatility for
the four bonds we include in the portfolio. Knowing the mean and volatility
forecasts implied by each model for the three exchange rates is not sufficient.
We also need the realized data point st in order to evaluate the predictive
density f t (st |t−1 , θ). Since the realized data point st is only observed ex
post, the only way to form the BMA weights ex ante is to lag the predictive
density and thus use f t−1 (st−1 |t−2 , θ).
4.3.2 The BW strategy. Under the BW strategy, in each time period we select
the set of one-step ahead conditional mean and volatility from the empirical
model that has the highest marginal likelihood up to that period. In other
words, the BW strategy only uses the forecasts of the “winner” model in terms
of marginal predictive density, and hence discards the forecasts of the rest of
the models. Clearly, there is no model averaging in the BW strategy. Similar
to the BMA, the BW strategy is evaluated ex ante using the lagged marginal
likelihood.
5. Empirical Results
5.1 FX data and descriptive statistics

The data sample consists of 348 monthly observations ranging from January
1976 to December 2004, and focuses on three exchange rates relative to the US
dollar: the UK pound sterling (USD/GBP), Deutsche mark/euro (USD/DEM-
EURO), and Japanese yen (USD/JPY). The spot and one-month forward
exchange rates are taken from Datastream for the period of January 1985 on-
wards, whereas for the period ranging from January 1976 to December 1984,
3507
they are taken from Hai, Mark, and Wu (1997). After the introduction of the euro
in January 1999, we use the euro exchange rate to replace the Deutsche mark
rate.

Data on money supply and income are from the International Monetary
Fund’s International Financial Statistics database. Specifically, we define the
money supply as the sum of money (line code 34) and quasi-money (line
code 35) for Germany and Japan, whereas for the UK we use M0 (line code 19).
Since German exchange rate data are only available until December 1998, we
use the money and quasi-money data of the Euro area for the remaining period
(January 1999 to December 2004). The US data is obtained from the aggregate
M2 of the Board of Governors of the Federal Reserve System. Furthermore,
we use the monthly industrial production index (line code 66) as a proxy
for national income rather than the gross domestic product (GDP), because
the latter is available only at the quarterly frequency.15 We deseasonalize the
money and industrial production indices following the procedure of Gomez and
Maravall (2000). Note that we ignore the complication arising from the fact that
the data we use on monetary fundamentals may not be available in real time
and may not suffer from the measurement errors that characterize real-time
macroeconomic data (Faust, Rogers, and Wright, 2003). This issue will not
affect our main findings on the predictive ability of the forward premium and
stochastic volatility.
We take logarithmic transformations of the raw data to yield time series for
st , f t , m t , m ∗t , yt , and yt∗ . The monetary fundamentals series z t is constructed
as in Equation (2); st is taken as the natural logarithm of the domestic price of
foreign currency, the United States being the domestic country; f t is the natural
logarithm of the US dollar price of a one-month forward contract issued at
time t for delivery of one unit of foreign currency at time t + 1. Finally, in our
economic evaluation of the set of candidate exchange rate models, the proxy
for the riskless domestic and foreign bonds is the end-of-month Euromarket
interest rate with one-month maturity, obtained from Datastream.16
Table 1 reports the descriptive statistics for the monthly percent FX returns
st , the three monetary fundamentals predictors, MF1 , MF2 , and MF3 , also
expressed in percent, and the percent forward premium, f t − st . For our sample
period, the sample means of the FX returns are −0.012% for USD/GBP, 0.165%
for USD/DEM-EURO, and 0.309% for USD/JPY. The FX return standard
deviations are similar across the three exchange rates at about 3% per month.
Finally, the exchange rate return sample autocorrelations are approximately
0.10 but decay rapidly.
15
For all countries, the correlation coefficient between the quarterly industrial production index and GDP over our
sample period is higher than 0.95.
16
We use the eurocurrency deposit rate as a proxy for the riskless rate because these deposits are comparable
across countries in all respects (such as credit risk and maturity) except for currency of denomination; see Levich
(1985).
3508
nloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January
Table 1
Descriptive statistics for monthly FX returns and fundamentals
UK (USD/GBP) Germany (USD/DEM-EURO) Japan (USD/JPY)
st MF1 MF2 MF3 FP st MF1 MF2 MF3 FP st MF1 MF2 MF3 FP
Panel A: Percent Returns

Mean −0.012 0.000 0.000 0.000 −0.159 0.165 0.000 0.000 0.000 0.131 0.309 0.000 0.000 0.000 0.267
Std Dev 2.95 21.52 13.50 11.57 0.266 3.05 33.43 17.56 16.82 0.300 3.32 42.00 29.61 16.88 0.300
Min −10.53 −50.02 −37.41 −24.56 −0.736 −9.71 −56.90 −34.87 −29.02 −0.779 −10.08 −72.24 −45.68 −43.25 −2.12
Max 11.69 52.89 41.86 44.00 1.35 9.21 75.85 45.75 46.78 1.59 11.89 72.29 70.28 30.85 1.31
Skewness −0.101 −0.137 −0.372 0.474 1.49 0.156 0.164 0.240 0.473 −0.162 0.619 0.234 0.575 −0.247 −1.05
Kurtosis 4.04 2.41 3.60 4.24 9.19 3.12 2.05 2.17 2.39 4.46 4.06 1.63 2.20 2.00 13.94
Corr(st ,st−1 ) 0.106 0.988 0.976 0.965 0.786 0.115 0.994 0.983 0.983 0.641 0.147 0.996 0.987 0.980 0.558
Corr(st ,st−3 ) 0.013 0.960 0.920 0.888 0.480 0.040 0.982 0.945 0.943 0.600 0.108 0.989 0.969 0.933 0.510
Corr(st , st−6 ) 0.037 0.911 0.838 0.780 0.208 0.025 0.959 0.886 0.880 0.593 −0.057 0.973 0.938 0.856 0.403
Corr(st , st−12 ) 0.025 0.793 0.674 0.569 0.334 0.030 0.897 0.734 0.719 0.538 0.078 0.935 0.882 0.704 0.267
Panel B: Absolute Percent Returns
Mean 2.25 17.92 10.46 8.57 0.237 2.43 28.14 15.07 14.18 0.271 2.48 37.17 25.65 14.60 0.315
Std Dev 1.90 11.87 8.51 7.76 0.200 1.84 17.99 8.98 9.01 0.184 2.22 19.46 14.71 8.43 0.248
Min 0.000 0.023 0.041 0.050 0.000 0.006 0.320 0.007 0.039 0.000 0.000 0.028 0.477 0.009 0.000
Max 11.69 52.89 41.86 44.00 1.35 9.71 75.85 45.75 46.78 1.59 11.89 72.29 70.28 43.25 2.12
Skewness 1.48 0.687 1.37 1.46 1.55 1.12 0.238 0.547 0.711 1.64 1.52 −0.281 0.729 0.312 1.95
Kurtosis 6.04 2.89 4.64 5.67 6.93 4.38 2.21 3.11 3.40 10.40 5.76 2.11 3.46 2.82 11.33
Corr(|st |, |st−1 |) 0.065 0.960 0.944 0.933 0.878 0.135 0.981 0.939 0.944 0.647 0.129 0.983 0.950 0.927 0.746
Corr(|st |, |st−3 |) 0.149 0.876 0.832 0.797 0.731 0.085 0.943 0.807 0.818 0.466 0.051 0.950 0.884 0.761 0.590
Corr(|st |, |st−6 |) 0.093 0.724 0.690 0.606 0.577 0.051 0.876 0.633 0.641 0.410 0.010 0.882 0.786 0.541 0.500
Corr(|st |, |st−12 |) 0.030 0.431 0.414 0.248 0.385 −0.046 0.685 0.305 0.271 0.212 −0.035 0.703 0.659 0.280 0.330
The table summarizes the descriptive statistics for the spot exchange rate percent returns (st ), the three demeaned percent monetary fundamentals specifications (MF1 , MF2 , MF3 ), and
the percent forward premium (FP). The data sample ranges from January 1976 through December 2004 for a sample size of 348 monthly observations. The exchange rates are defined as
3509
US dollars per unit of foreign currency. For a detailed definition of the three monetary fundamentals specifications, see Section 2.1.
The three specifications of monetary fundamentals predictors display high

volatility and persistence. For instance, the standard deviation of MF1 is about
20% for the UK, 30% for Germany, and 40% for Japan. However, the stan-

dard deviation of MF3 (which is corrected for both the intercept and the time
trend component) is approximately half the value of the canonical monetary
fundamentals MF1 . The three monetary fundamentals predictors exhibit low
skewness, low excess kurtosis, and high serial correlation. Finally, the average
forward premium is negative for the UK, but positive for Germany and Japan.
The standard deviation of f t − st is low across all exchange rates (in fact, about
100 times smaller than MF1 ), but the forward premium exhibits high kurtosis
and its sample autocorrelation is high and decreasing slowly.
5.2 Estimation of exchange rate models

We begin our statistical and economic evaluation of short-horizon exchange
rate predictability by performing a Bayesian estimation of the parameters of our
15 candidate models: the five conditional mean specifications (RW, MF1 , MF2 ,
MF3 , and FP) under the three volatility frameworks (LR, GARCH, and SV). The
posterior mean estimates for the model parameters are presented in Tables 2,
3, and 4. We particularly focus on the size, sign, and statistical significance of
the β estimate because it captures the effect of either monetary fundamentals
or the forward premium in the conditional mean of exchange rate returns. In
our Bayesian MCMC framework, we assess statistical significance using two
diagnostics. First, we report the highest posterior density (HPD) region for each
parameter estimate. For example, the 95% HPD region is the shortest interval
that contains 95% of the posterior distribution. We check whether the 90%,
95%, and 99% HPD regions contain zero, which is equivalent to two-sided
hypothesis testing at the 10%, 5%, and 1% levels, respectively. Second, we
compute the NSE as defined in Equation 16.
Tables 2–4 illustrate that, for the three monetary fundamentals specifications
(MF1 , MF2 , and MF3 ), the in-sample β estimate tends to be a low positive
number, which increases in size as we move from MF1 to MF3 . This suggests
that when st is below (above) its fundamental value z t , it is expected to slowly
rise (decrease) over time. In contrast, the in-sample β estimate for the FP model
has a large negative value. The tables also report the estimates of the conditional
variance parameters. For the LR model, the monthly variance of FX returns
remains largely unchanged across the five conditional mean specifications and
is around 10 (i.e., ν ≈ 3%) for all three currencies. For the GARCH(1,1)
models, the conditional monthly variance is highly persistent since the sum
γ1 + γ2 revolves around 0.96 for all specifications. The SV models exhibit
(1) high persistence (φ) in the conditional monthly log-variance and (2) a
sizeable stochastic component in the conditional monthly log-variance. Finally,
all parameters in both the conditional mean and volatility exhibit very low NSE
values and therefore a high degree of statistical significance.
3510
Table 2
Posterior means for the UK pound sterling (USD/GBP)
Parameter RW MF1 MF2 MF3 FP

Panel A: Bayesian linear regression
α −0.012 −0.012 −0.012 −0.012 −0.110
(0.0005) (0.0005) (0.0005) (0.0005) (0.0005)
β – 0.0079 0.0254∗∗ 0.0254∗ −0.629
(2.4e − 05) (3.7e − 05) (4.4e − 05) (0.0016)
v2 8.72∗∗∗ 10.12∗∗∗ 8.63∗∗∗ 10.06∗∗∗ 8.69∗∗∗
(0.0021) (0.0021) (0.0021) (0.0021) (0.0021)
Panel B: Bayesian GARCH(1,1)
α 0.018 0.027 0.005 0.017 −0.101

(0.0022) (0.0021) (0.0021) (0.0020) (0.0022)
β – 0.0042 0.0215 0.0193 −0.816
(0.0001) (0.0002) (0.0002) (0.0078)
ω 0.331∗∗∗ 0.346∗∗∗ 0.324∗∗∗ 0.329∗∗∗ 0.387∗∗∗
(0.0015) (0.0026) (0.0017) (0.0017) (0.0064)
γ1 0.905∗∗∗ 0.902∗∗∗ 0.903∗∗∗ 0.902∗∗∗ 0.897∗∗∗
(0.0003) (0.0004) (0.0003) (0.0003) (0.0009)
γ2 0.055∗∗∗ 0.056∗∗∗ 0.0572∗∗∗ 0.058∗∗∗ 0.056∗∗∗
(0.0002) (0.0002) (0.0002) (0.0002) (0.0004)
Panel C: Bayesian stochastic volatility
α 0.048 0.046 0.022 0.022 −0.045

(0.0026) (0.0028) (0.0032) (0.0027) (0.0034)
β – 0.0028 0.0211 0.0226 −0.653
(0.0001) (0.0002) (0.0004) (0.0107)
μ 2.01∗∗∗ 2.02∗∗∗ 2.01∗∗∗ 2.01∗∗∗ 2.00∗∗∗
(0.0037) (0.0033) (0.0035) (0.0036) (0.0033)
φ 0.882∗∗∗ 0.878∗∗∗ 0.885∗∗∗ 0.884∗∗∗ 0.871∗∗∗
(0.0012) (0.0014) (0.0014) (0.0013) (0.0015)
σ2 0.093∗∗∗ 0.092∗∗∗ 0.086∗∗∗ 0.090∗∗∗ 0.097∗∗∗
(0.0010) (0.0010) (0.0009) (0.0010) (0.0011)
The table presents the Bayesian MCMC estimates of the posterior means of the linear regression,
GARCH(1,1), and SV model parameters for the USD/GBP monthly percent returns. The MCMC chain
run for 5000 iterations after an initial burn-in of 1000 iterations. The numbers in parentheses indicate the
numerical standard error (NSE). The superscripts *, **, and *** indicate that the 90%, 95%, and 99%
highest posterior density (HPD) regions, respectively, do not contain zero. The HPD region for each MCMC
parameter estimate is the shortest interval that contains 95% of the posterior distribution.
5.3 Evaluating forecasts using statistical criteria

We assess the statistical evidence on short-horizon exchange rate predictability
by ranking our set of 15 candidate models according to their log-likelihood and
posterior probability. The conditional performance of the models is evaluated
in-sample as well as out-of-sample. The in-sample period for the three monthly
exchange rates covers 29 years ranging from January 1976 to December 2004.
The out-of-sample exercise involves two steps: (1) initial parameter estimation
for the 14-year period of January 1976 to December 1989, and (2) sequential
monthly updating of the parameter estimates for the out-of-sample 15-year
period of January 1990 to December 2004. In other words, the forecasts at
any given month are constructed according to a recursive procedure that is
conditional only upon information up to the date of the forecast. The model is
3511
Table 3
Posterior means for the Deutsche mark/euro (USD/DEM-EURO)

α 0.160 0.160 0.160 0.160 0.206
(0.0005) (0.0005) (0.0005) (0.0005) (0.0005)
β – 0.0077 0.0104 0.0148 −0.355
(1.6e − 0.5) (3.0e − 05) (3.1e − 05) (0.0015)
v2 9.30∗∗∗ 9.26∗∗∗
9.30∗∗∗ 9.27∗∗∗ 10.81∗∗∗
(0.0023) (0.0022) (0.0022) (0.0022) (0.0022)
α 0.153 0.164 0.154 0.159 0.216

(0.0024) (0.0022) (0.0023) (0.0023) (0.0023)
β – 0.0068 0.0097 0.0134 −0.463
(7.6e − 05) (0.0001) (0.0001) (0.0072)
ω 0.405∗∗∗ 0.404∗∗∗ 0.400∗∗∗ 0.399∗∗∗ 0.409∗∗∗
(0.0019) (0.0026) (0.0028) (0.0027) (0.0023)
γ1 0.930∗∗∗ 0.929∗∗∗ 0.928∗∗∗ 0.928∗∗∗ 0.929∗∗∗
(0.0003) (0.0003) (0.0003) (0.0003) (0.0003)
γ2 0.028∗∗∗ 0.029∗∗∗ 0.030∗∗∗ 0.030∗∗∗ 0.028∗∗∗
(0.0002) (0.0002) (0.0003) (0.002) (0.0002)
Panel C: Bayesian stochastic volatility
α 0.163 0.176 0.165 0.173 0.219

(0.0031) (0.0010) (0.0028) (0.0028) (0.0030)
β – 0.0074 0.0091 0.0136 −0.440
(9.4e − 05) (0.0002) (0.0002) (0.0084)
μ 2.17∗∗∗ 2.16∗∗∗ 2.17∗∗∗ 2.16∗∗∗ 2.17∗∗∗
(0.0030) (0.0028) (9.4e − 05) (0.0030) (0.0029)
φ 0.746∗∗∗ 0.751∗∗∗ 0.757∗∗∗ 0.755∗∗∗ 0.751∗∗∗
(0.0032) (0.0036) (0.0033) (0.0036) (0.0033)
σ2 0.068∗∗∗ 0.069∗∗∗ 0.068∗∗∗ 0.067∗∗∗ 0.069∗∗∗
(0.0005) (0.0006) (0.0005) (0.0005) (0.0005)
GARCH(1,1), and SV model parameters for the USD/DEM-EURO monthly percent returns. The MCMC
chain run for 5000 iterations after an initial burn-in of 1000 iterations. The numbers in parentheses indicate
the numerical standard error (NSE). The superscripts *, **, and *** indicate that the 90%, 95%, and
99% highest posterior density (HPD) regions, respectively, do not contain zero. The HPD region for each
MCMC parameter estimate is the shortest interval that contains 95% of the posterior distribution.
then successively re-estimated as the date on which forecasts are conditioned

moves through the data set. Hence the design of the out-of-sample exercise is
computationally intensive.
Our analysis of the statistical evidence begins with Table 5, which presents
the log-likelihood values and shows that across volatility models, the SV model
always has higher log-likelihood than both LR and GARCH. This result is
robust as it holds for all currencies both in-sample and out-of-sample. Similarly,
the GARCH(1,1) model always beats the constant variance LR models in terms
of log-likelihood. Across conditional mean specifications, the RW model is
always worse in-sample. Finally, the out-of-sample log-likelihood values lead
to the following conclusions: FP is the best model for the yen, but the RW
model is best for the pound sterling and the Deutsche mark/euro.
3512
Table 4
Posterior means for the Japanese yen (USD/JPY)

α 0.299∗ 0.299∗ 0.299∗ 0.299∗ 0.615∗∗∗
(0.0006) (0.0005) (0.0005) (0.0005) (0.0007)
β – 0.0070 0.0075 0.0189∗ −1.224∗∗
(1.3e − 0.5) (1.9e − 05) (3.4e − 05) (0.0016)
v2 11.0∗∗∗ 10.96 ∗∗∗
10.99∗∗∗ 10.94∗∗∗ 10.79∗∗∗
(0.0027) (0.0026) (0.0026) (0.0026) (0.0026)
α 0.343∗ 0.344∗ 0.342∗ 0.333∗ 0.623∗∗∗

(0.0027) (0.0023) (0.0024) (0.0024) (0.0027)
β – 0.0065 0.0077 0.0164 −1.170∗∗
(6.4e − 05) (8.6e − 05) (0.0002) (0.0069)
ω 0.595∗∗∗ 0.593∗∗∗ 0.599∗∗∗ 0.594∗∗∗ 0.676∗∗∗
(0.0028) (0.0034) (0.0043) (0.0042) (0.0076)
γ1 0.911∗∗∗ 0.911∗∗∗ 0.909∗∗∗ 0.911∗∗∗ 0.898∗∗∗
(0.0004) (0.0004) (0.0005) (0.0005) (0.0011)
γ2 0.037∗∗∗ 0.036∗∗∗ 0.038∗∗∗ 0.036∗∗∗ 0.041∗∗∗
(0.0003) (0.0003) (0.0003) (0.0003) (0.0005)
Panel C: Bayesian Stochastic Volatility
α 0.166 0.166 0.161 0.138 0.532∗∗∗

(0.0039) (0.0039) (0.0041) (0.0040) (0.0038)
β – 0.0055 0.0059 0.0155 −1.763∗∗∗
(8.1e − 05) (0.0001) (0.0002) (0.0109)
μ 2.16∗∗∗ 2.16∗∗∗ 2.16∗∗∗ 2.15∗∗∗ 2.06∗∗∗
(0.0040) (0.0042) (0.0037) (0.0036) (0.0040)
φ 0.814∗∗∗ 0.818∗∗∗ 0.818∗∗∗ 0.816∗∗∗ 0.801∗∗∗
(0.0017) (0.0020) (0.0018) (0.0019) (0.0019)
σ2 0.149∗∗∗ 0.147∗∗∗ 0.146∗∗∗ 0.150∗∗∗ 0.230∗∗∗
(0.0025) (0.0021) (0.0022) (0.0020) (0.0043)
GARCH(1,1), and SV model parameters for the USD/JPY monthly percent returns. The MCMC chain
run for 5000 iterations after an initial burn-in of 1000 iterations. The numbers in parentheses indicate
the numerical standard error (NSE). The superscripts *, **, and *** indicate that the 90%, 95%, and
99% highest posterior density (HPD) regions, respectively, do not contain zero. The HPD region for each
MCMC parameter estimate is the shortest interval that contains 95% of the posterior distribution.
In Table 6 we rank the in-sample and out-of-sample performance of our set

of candidate models according to their posterior probability. The key input to
this statistical criterion is the calculation of the marginal likelihood. Therefore,
Table 6 gives us a distinct statistical perspective on performance because the
marginal likelihood is computed in a way that integrates out parameter uncer-
tainty and imposes a penalty for lack of parsimony (higher dimension). The
results in Table 6 indicate two clear patterns in ranking the models. The first
pattern confirms one of our most robust results: the best models for all three
currencies both in-sample and out-of-sample have SV innovations. The second
pattern is slightly different from the log-likelihood findings: for all three ex-
change rates, both in-sample and out-of-sample, the best model is FPSV , the
second best is RWSV , and third best is one of the three MFSV specifications. The
3513
Table 5
The likelihood of the models
Model RW MF1 MF2 MF3 FP

Panel A: In-sample log-likelihood
USD/GBP
LR −867.47 −866.90 −865.12 −865.76 −866.50
GARCH(1,1) −860.94 −860.51 −858.00 −858.67 −860.28
SV −801.80 −804.07 −802.25 −801.15 −801.33
USD/DEM-EURO
LR −878.71 −877.47 −878.09 −877.55 −878.49

GARCH(1,1) −878.26 −876.96 −877.45 −876.91 −877.97
SV −847.84 −846.23 −846.52 −846.72 −846.64
USD/JPY
LR −907.95 −906.60 −907.17 −906.34 −904.08

GARCH(1,1) −906.97 −905.60 −905.93 −905.37 −903.05
SV −828.49 −827.76 −828.66 −827.61 −791.47
Model RW MF1 MF2 MF3 FP
Panel B: Out-of-Sample log-Likelihood

USD/GBP
LR −439.52 −440.173 −438.97 −439.25 −439.91
GARCH(1,1) −427.58 −427.69 −426.27 −426.31 −428.25
SV −412.08 −412.91 −412.24 −412.63 −412.18
USD/DEM-EURO
LR −451.91 −452.37 −452.50 −453.06 −452.76

GARCH(1,1) −448.47 −448.68 −449.00 −449.36 −449.03
SV −427.120 −433.68 −433.91 −434.69 −434.58
USD/JPY
LR −465.04 −465.58 −465.47 −466.99 −464.84

GARCH(1,1) −458.33 −458.68 −458.68 −460.14 −458.05
SV −433.814 −426.40 −425.50 −425.68 −413.52
The table reports the in-sample and out-of-sample log-likelihood values for the three FX rates (USD/GBP,
USD/DEM-EURO, and USD/JPY), five conditional mean specifications (RW, MF1 , MF2 , MF3 , and FP) and
three volatility frameworks (linear regression, GARCH, and stochastic volatility). The out-of-sample data runs
from January 1990 through December 2004.
single exception is the pound sterling for which RWSV is the best out-of-sample
model. Hence, in contrast to the likelihood evidence, the MF specifications lose
to RW even in-sample. In other words, the penalty the posterior probability im-
poses on the three monetary fundamentals models for lack of parsimony offsets
their log-likelihood advantage.
5.4 Evaluating forecasts using economic criteria

We assess the economic value of short-horizon exchange rate predictability by
analyzing the performance of the dynamically rebalanced portfolios constructed
using our set of 15 candidate models. Our analysis focuses on the performance
3514
Table 6
The models with the highest posterior probability
Best Model Second Best Model Third Best Model

Panel A: The best in-sample models
USD/GBP FPSV RWSV MFSV
3
USD/DEM-EURO FP SV
RW SV
MFSV
2
USD/JPY FP SV
RW SV
MFSV
3
Panel B: The best out-of-sample models
USD/GBP RWSV FPSV MFSV

2
USD/DEM-EURO FPSV RWSV MFSV
2
USD/JPY FPSV RWSV MFSV
3
The table shows the three best models according to the highest in-sample and out-of-sample
posterior probability for the three FX rates (USD/GBP, USD/DEM-EURO, and USD/JPY). The
out-of-sample data runs from January 1990 through December 2004. Ranking the models using the
highest posterior probability is equivalent to choosing the best model in terms of density forecasts
and is a robust model selection criterion in the presence of misspecification and non-nested models.
fee a US investor is willing to pay for switching from one FX strategy to

another. The fees are reported in Table 7, which displays the economic value of
each mean and volatility specification relative to the benchmark random walk
model with constant variance (RWLR ). We present the fees for the degrees of
RRA δ = 2 and δ = 6.
Panel A of Table 7 presents the in-sample performance fees and demonstrates
that the three monetary fundamentals specifications generally have no economic
value, as indicated by the negative values. In contrast, the FP model exhibits
high economic value, especially under stochastic volatility. For example, at the
target portfolio volatility of σ∗p = 10% and for δ = 2, a US investor is willing
to pay a substantial 248 annual basis points (bps) for switching from the RWLR
model to FPSV . Consistent with our statistical evidence, for all conditional mean
specifications there tends to be high economic value associated with stochastic
volatility. However, contrary to our statistical evidence, the performance of the
GARCH(1,1) model is surprisingly poor relative to the constant variance LR
model. For σ∗p = 10% and δ = 2, the in-sample fee for switching from RWLR
to RWGARCH is −24 bps, whereas the fee for switching from RWLR to RWSV is
42 bps.17 Finally, as investors become less risk-averse, the fees tend to increase
in absolute value, strengthening the evidence against RWLR and in favor of the
FPSV specification.
The out-of-sample performance fees are displayed in panel B of Table 7 and
suggest that even for out-of-sample there is still high economic value in both
17
At first sight, the poor performance of the GARCH model in terms of economic value appears rather surprising.
For instance, Fleming and Kirby (2003) find that SV models only marginally outperform GARCH models.
However, there is no study to date that assesses the economic value of GARCH and SV models, especially when
applied to exchange rates. Furthermore, the negative in-sample and out-of-sample performance fees of RWGARCH
are not far from zero.
3515
Table 7
The economic value of the empirical exchange rate models
Panel A: In-sample performance for models versus RWLR

MFLR
1 MFLR
2 MFLR
3 FPLR
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) −26 − −58 − −129 − −144 − −127 − −167 − 144 120 145 118
10 (%) −37 − −90 − −164 − −190 − −165 − −228 − 180 120 181 117
12 (%) −51 − −129 − −200 − −239 − −205 − −299 − 217 120 218 117
MFGARCH
1 MFGARCH
2 MFGARCH
3 FPGARCH
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) 3 14 −13 − −120 − −128 − −128 − −151 − 132 101 131 98

10 (%) 1 9 −25 − −152 − −165 − −164 − −202 − 165 101 163 96
12 (%) −3 − −41 − −184 − −205 − −201 − −257 − 198 101 195 96
MFSV
1 MFSV
2 MFSV
3 FPSV
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) 102 166 38 59 −55 − −91 − −79 − −144 − 203 136 166 108
10 (%) 118 151 15 18 −74 − −132 − −108 − −213 − 248 132 190 97
12 (%) 130 137 −21 − −95 − −180 − −142 − −296 − 291 129 206 87
RWGARCH RWSV
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) −19 − −19 − 36 556 9 51

10 (%) −24 − −24 − 42 471 16 150
12 (%) −28 − −29 − 48 404 19 267
Panel B: Out-of-sample performance for models versus RWLR
MFLR
1 MFLR
2 MFLR
3 FPLR
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) −173 − −178 − 23 26 11 9 −154 − −149 − 61 34 56 31

10 (%) −217 − −224 − 27 23 9 2 −192 − −184 − 76 34 68 30
12 (%) −261 − −271 − 30 21 4 1 −229 − −218 − 90 34 79 29
MFGARCH
1 MFGARCH
2 MFGARCH
3 FPGARCH
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) −124 − −120 − −1 − −2 − −195 − −185 − 55 28 63 32

10 (%) −154 − −147 − −1 − −2 − −242 − −227 − 70 29 82 34
12 (%) −184 − −172 − −1 − −3 − −289 − −267 − 86 29 103 36
MFSV
1 MFSV
2 MFSV
3 FPSV
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) −18 − −68 − 14 −3 −33 − 55 40 26 19 218 92 174 73

10 (%) −30 − −108 − 10 −11 −64 − 65 37 19 10 266 90 197 65
12 (%) −45 − −158 − 4 −19 −105 − 73 34 6 2 311 87 212 58
3516
Table 7
(Continued)
RWGARCH RWSV
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6

8 (%) −27 − −17 − 105 342 76 246
10 (%) −32 − −16 − 127 321 82 205
12 (%) −37 − −13 − 147 304 82 166
The table presents the in-sample and out-of-sample performance fees () and break-even transaction costs (τBE )
for selected models against the RWLR benchmark for three target portfolio volatilities (8%, 10% and 12%). Each
maximum return strategy builds an efficient portfolio by investing in the monthly return of four bonds from the
United States, UK, Germany, and Japan and using the three exchange rates to convert the portfolio return in US
dollars. The fees denote the amount an investor with quadratic utility and a degree of relative risk aversion equal
to either 2 or 6 is willing to pay for switching from RWLR to another model (such as FPSV ). The performance fee
is expressed in annual basis points. The transaction cost τBE is defined as the minimum monthly proportional
cost that cancels out the utility advantage (and hence positive performance fee) of a given strategy. The τBE
values are expressed in monthly basis points and are reported only when is positive. The in-sample period
starts in January 1979 and the out-of-sample data runs from January 1990 through December 2004.
the forward premium and stochastic volatility. This is a new and important
result, which adds to the existing literature that is anchored around the seminal
contribution of Meese and Rogoff (1983). Specifically, at σ∗p = 10% and δ = 2,
the annual performance fees for switching from RWLR to another model are
127 bps for RWSV and 266 bps for FPSV . We can therefore conclude that there is
substantial economic value both in-sample and out-of-sample against the naive
random walk model and in favor of conditioning on the forward premium with
stochastic volatility. This finding is in fact consistent with the large profits made
by financial institutions that engage in sophisticated multicurrency forward bias
strategies. For example, Galati and Melvin (2004) show that simple carry trades
aiming at exploiting the forward bias constitute a significant source of the surge
in FX trading in recent years.
In addition to the results associated with individual models, even stronger
economic evidence is found for the combined forecasts reported in Table 8,
which compares BMA and BW to the RWLR benchmark for two cases: (1) the
restricted universe of the five SV models (because the SV models generally
perform better), and (2) the unrestricted universe of all 15 models. A purely
agnostic approach to forecast combination would use the full set of 15 models
(case 2). The results in Table 8 provide robust evidence against the naive random
walk model as all performance fees based on the BMA and BW are positive
and high, both in-sample and out-of-sample. For example, when selecting
among the SV models and setting σ∗p = 10% and δ = 2, the annual in-sample
performance fee for switching away from the benchmark RWLR is 255 bps
for BMA and 235 bps for BW. The out-of-sample fees are even higher at 317
bps for BMA and 340 bps for BW. In short, therefore, there is clear in-sample
and out-of-sample economic evidence on the superiority of combined forecasts
relative to the naive random walk benchmark.
In conclusion, Figure 1 offers a visual description of the time variation in
the weights investing in the three risky assets: the UK, German, and Japanese
3517
Table 8
The economic value of combined forecasts
Panel A: In-sample performance
All Models (versus RWLR )

BMA BW
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) 207 145 156 120 192 117 156 93
10 (%) 254 141 172 109 235 114 177 83
12 (%) 299 138 178 100 276 111 191 74
Stochastic volatility models (versus RWLR )

BMA BW
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) 208 146 157 120 192 117 156 93
10 (%) 255 142 173 110 235 114 178 83
12 (%) 300 139 179 100 276 111 191 74
Panel B: Out-of-sample performance

All models (versus RWLR )
BMA BW
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) 250 130 206 108 268 128 222 107
10 (%) 306 127 237 100 329 125 255 99
12 (%) 360 124 259 91 386 122 279 90
Stochastic volatility models (versus RWLR )

BMA BW
σ∗p 2 τBE
2 6 τBE
6 2 τBE
2 6 τBE
6
8 (%) 259 134 215 112 277 131 231 111
10 (%) 317 131 249 104 340 129 267 103
12 (%) 373 128 273 95 400 126 294 94
The table reports the in-sample and out-of-sample performance fees () and break-even transaction costs
(τBE ) for all maximum return strategies based on combined forecasts for three target portfolio volatilities
(8%, 10%, and 12%). BMA denotes Bayesian model average and BW is Bayesian winner. The combined
forecasts are shown for two cases: (1) the unrestricted universe of all 15 models, and (2) the restricted
universe of only the five stochastic volatility models. The fees denote the amount an investor with quadratic
utility and a degree of relative risk aversion equal to either 2 or 6 is willing to pay for switching from
the RWLR benchmark to the BMA or BW strategy. τBE is defined as the minimum monthly proportional
cost that cancels out the utility advantage (and hence positive performance fee) of a given strategy. The
transaction costs are only reported when is positive. The performance fees are expressed in annual basis
points, and the transaction costs in monthly basis points. The in-sample period starts in January 1979 and
the out-of-sample data runs from January 1990 through December 2004.
bonds. The figure displays the weights for four cases: the benchmark RWLR
model, the best performing individual model FPSV , the BMA, and the BW
combined forecast strategies. As expected, the weights are very smooth over
time for RWLR , and remain reasonably smooth for the FPSV model and the two
combined forecast strategies.18
18
However, the dynamic weights appear to be more volatile in the beginning of the sample before they stabilize.
We believe that the initial instability in the weights is due to the high exchange-rate volatility around the 1992
crisis of the Exchange Rate Mechanism that forced the UK to abandon the target zone system.
3518

Figure 1
The out-of-sample dynamic weights for selected models
This is the out-of-sample time variation in the weights investing in the three risky assets (the UK, Japanese,
and German bonds) at a target portfolio volatility of 10% and a degree of relative risk aversion of 2. The figure
presents four cases: the benchmark random walk model with constant variance (upper left), the forward premium
model with stochastic volatility (upper right), the Bayesian model average strategy (lower left), and the Bayesian
winner strategy (lower right).
5.5 Transaction costs

If transaction costs are sufficiently high, the period-by-period fluctuations in
the dynamic weights of an optimal strategy will render the strategy too costly
to implement relative to the static random walk model. We address this concern
by computing the break-even transaction cost τBE as the minimum monthly
proportional cost that cancels out the utility advantage (and hence positive
performance fee) of a given strategy. In comparing a dynamic strategy with the
static random walk strategy, an investor who pays a transaction cost lower than
τBE will prefer the dynamic strategy. The τBE values are expressed in monthly
basis points and are reported only when is positive.
The in-sample break-even transaction costs are reported in panel A of Table 7,
which demonstrates that for the forward premium and stochastic volatility the
values of τBE are positive and high; they tend to be higher than 100 bps and can
be as high as 556 bps. For instance, at σ∗p = 10% and δ = 2, a US investor will
switch back to the RWLR model if he is subject to a proportional transaction
cost of at least 120 bps for FPLR , 101 bps for FPGARCH , 132 bps for FPSV , and
471 bps for RWSV . In other words, at the reasonably high transaction cost of
50 bps (e.g., Marquering and Verbeek, 2004), there is still significant in-sample
3519
economic value in empirical models that condition on the forward premium,

especially under stochastic volatility.
Determining the out-of-sample robustness to transaction costs is one of the

most important considerations in assessing the forecasting performance of
empirical exchange rate models. Panel B of Table 7 shows that conditioning
on the forward premium and stochastic volatility leads to reasonably high τ B E
values. Specifically, at σ∗p = 10% and δ = 2, the break-even transaction cost
that would eliminate the performance fee of 266 bps of the FPSV model relative
to the RWLR benchmark is 90 bps. Furthermore, the τBE for RWSV versus RWLR
is a very large: 321 bps.
The evidence on the τBE of combined forecasts displayed in Table 8 is
even stronger. Compared to the benchmark RWLR , a combined forecast of all
15 models exhibits an in-sample τBE of 141 bps for BMA and 114 bps for BW.
Panel B of Table 8 shows that the out-of-sample τBE values for combined fore-
casts are generally as high as in sample. In short, as the τBE values are generally
positive and reasonably high, we conclude that the in-sample and out-of-sample
economic value we have reported is robust to reasonably high transaction costs
for empirical exchange rate models conditioning on the forward premium, for
models with SV innovations, and for combined forecasts.
5.6 Summary of results

The statistical and economic evidence on short-horizon exchange rate pre-
dictability supports the following four results: (1) the forward premium model
unequivocally beats the random walk, (2) conditioning on monetary funda-
mentals has no economic value, (3) the stochastic volatility process always
leads to superior portfolio performance, and (4) the combined forecasts con-
sistently outperform the constant variance random walk benchmark. All these
results hold both in-sample and out-of-sample and are robust to reasonably
high transaction costs.
6. Robustness and Extensions

This section discusses directions in which one can possibly extend the analysis
of the paper. First, we perform an additional robustness test by evaluating the
out-of-sample performance of the empirical models in three five-year subsam-
ples. Recall that the full sample period at our disposal covers 29 years ranging
from January 1976 to December 2004. We use data from January 1976 to
December 1989 for in-sample estimation, whereas the out-of-sample period is
of 15 years ranging from January 1990 to December 2004. The out-of-sample
results we report in Tables 5–8 are for the entire 15-year out-of-sample period.
In addition, panel A of Table 9 presents the performance fees for selected mod-
els and for three subsamples: 1990–1994, 1995–1999, and 2000–2004. We find
that the economic value of conditioning on the forward premium and stochas-
tic volatility is positive in all periods but is substantially higher in the last two
3520
Table 9
Out-of-sample robustness
Panel A: Subsample analysis for selected models versus RWLR
(σ∗p = 10%, δ = 2)

FPSV BMA BW
Subsample 2 τBE
2 2 τBE
2 2 τBE
2
1990–1994 40 12 196 50 260 56
1995–1999 539 347 519 346 523 357
2000–2004 229 83 227 93 224 96
1995–2004 381 193 363 208 364 281
1990-2004 266 90 306 127 329 125
Panel B: The performance of tGARCH models versus RWLR (δ = 2)

MFtGARCH
1 MFtGARCH
2 MFtGARCH
3 FPtGARCH RWtGARCH
σ∗p 2 τBE
2 2 τBE
2 2 τBE
2 2 τBE
2 2 τBE
2
8 (%) −34 − −32 − −29 − 110 49 21 78
10 (%) −43 − −40 − −36 − 140 50 28 82
12 (%) −50 − −48 − −42 − 169 51 35 88
The table provides an analysis of out-of-sample robustness for the performance fees () and
break-even transaction costs (τBE ) of selected models against the RWLR benchmark. Panel A
conducts a subsample analysis and panel B examines the performance of the tGARCH(1,1) model
with Student-t innovations. BMA denotes Bayesian model average and BW is Bayesian winner.
All maximum return strategies build an efficient portfolio by investing in the monthly return of
four bonds from the United States, UK, Germany, and Japan and using the three exchange rates to
convert the portfolio return in US dollars. The fees denote the amount an investor with quadratic
utility and a degree of relative risk aversion equal to 2 is willing to pay for switching from RWLR
to (say) FPSV . The target portfolio volatility in panel A is set at 10%. τBE is defined as the
minimum monthly proportional cost that cancels out the utility advantage (and hence positive
performance fee) of a given strategy. The transaction costs are only reported when is positive.
The performance fees are expressed in annual basis points, and the transaction costs in monthly
basis points. The combined forecasts are for the universe of all 15 models. The out-of-sample
period runs from January 1990 through December 2004.
subsamples. This is consistent with the well-known fact in the literature that the
forward bias is small in the early 1990s (e.g., Flood and Rose, 2002).19 For all
models, the best subsample period is 1995–1999. Furthermore, it is important
to note that the combined forecast strategies substantially outperform the ran-
dom walk benchmark in all three subsamples and display similar performance
fees to FPSV for the last two subsamples. However, for the first subsample when
the forward bias is small, the BMA and BW strategies significantly outperform
FPSV by optimally using predictive information from the entire universe of
models, including monetary fundamentals.
Second, our analysis of the conditional variance of exchange rate returns
includes the GARCH(1,1) specification because this is the benchmark model
in the seminal study of West, Edison, and Cho (1993). As a further robustness
check, we examine the out-of-sample performance of the tGARCH(1,1) model
of Bollerslev (1987) to determine whether departing from the assumption of
19
In a separate experiment, we start the out-of-sample exercise in 1985 and find significant economic value in the
forward premium and stochastic volatility for the 1985–1989 period. However, starting the out-of-sample period
in 1985 leaves too few in-sample observations for initial parameter estimation. Therefore, the tables present the
out-of-sample results for the period starting in 1990.
3521
Table 10
Sharpe ratios for selected models
σ∗p RWLR FPSV BMA BW

Panel A: In-sample
8 (%) 0.88 1.09 1.11 1.13
10 (%) 0.91 1.14 1.15 1.17
12 (%) 0.94 1.17 1.19 1.21
Panel B: Out-of-sample
8 (%) 0.76 0.98 1.06 1.11
10 (%) 0.76 0.98 1.06 1.12
12 (%) 0.76 0.98 1.06 1.12
The table presents the in-sample and out-of-sample annualized Sharpe ratios for selected
models. BMA denotes Bayesian model average and BW is Bayesian winner. The Sharpe
ratios are adjusted for the serial correlation in the monthly portfolio returns generated by the
dynamic strategies (e.g., Lo, 2002). All maximum return strategies build an efficient portfolio
by investing in the monthly return of four bonds from the United States, UK, Germany, and
Japan and using the three exchange rates to convert the portfolio return in US dollars. The
maximum return strategies are evaluated at three target portfolio return volatilities: 8(%),
10(%), and 12(%). The in-sample period starts in January 1979 and the out-of-sample data
runs from January 1990 through December 2004.
conditional normality improves the performance of the GARCH model.20 The

results from this exercise (a subset of which is reported in panel B of Table 9) re-
veal that using a Student-t distribution leads to substantial performance gains in
the GARCH framework.21 In particular, we find that the out-of-sample perfor-
mance fees of the tGARCH(1,1) model are much higher than for GARCH(1,1),
especially for the forward premium and random walk conditional mean speci-
fications. For instance, setting σ∗p = 10% and δ = 2 and comparing the results
in Tables 7 and 9 indicates that the out-of-sample fees for switching from the
RWLR model to the forward premium models are as follows: 76 bps for FPLR ,
70 bps for FPGARCH , 140 bps for FPtGARCH , and 266 bps for FPSV . Similarly,
when switching from the random walk with constant variance, RWLR , to a
random walk with time-varying volatility, the fees are: −32 bps for RWGARCH ,
28 bps for RWtGARCH , and 127 bps for RWSV . Therefore, we can conclude that
in terms of economic value the tGARCH model performs better than GARCH,
although the SV model outperforms both normal and Student-t GARCH spec-
ifications. Hence, our main conclusions remain qualitatively the same.
Third, Table 10 presents the in-sample and out-of-sample annualized Sharpe
ratios for selected models. The Sharpe ratio values are generally in agreement
with the performance fees and hence confirm our conclusions. Specifically,
FPSV , BMA, and BW consistently outperform RWLR both in-sample and out-
of-sample. For example, at σ∗p = 10% , the out-of-sample Sharpe ratios are:
0.76 for RWLR , 0.98 for FPSV , 1.06 for BMA, and 1.12 for BW.
20
In estimating the tGARCH model, we implement an algorithm similar to the GARCH case as described in the
Appendix, with an additional Metropolis-Hastings step for sampling the degrees of freedom parameter ν.
21
Note that the degrees of freedom parameter estimate revolves around ν = 10 for the UK pound sterling, ν = 25
for the Deutsche mark/euro, and ν = 7 for the Japanese yen (not reported). This indicates that the unconditional
distribution of exchange rate returns is not normal, especially for the UK pound sterling and the Japanese yen.
3522
Fourth, this paper explores the predictability in exchange rates by focusing

on the frequency and horizon of one month. On the one hand, adopting the
monthly frequency is a natural choice because this is the highest frequency at

which monetary fundamentals are observed. On the other hand, our motivation
for investigating predictability at the one-month horizon is founded on the
prevailing view in this literature that exchange rates are not predictable at short
horizons. It is clear, therefore, that one possible direction in extending the
analysis of this paper is to study the predictability of the forward premium,
stochastic volatility, and combined forecasts for higher frequencies and longer
horizons. We leave this for future research.
Finally, we study short-horizon exchange rate predictability by estimating a
set of univariate conditional mean and volatility models. However, in assess-
ing the economic value of exchange rate predictability we build multivariate
dynamic asset allocation strategies. Specifically, the optimal weights of the
dynamically rebalanced portfolios are computed using the conditional mean
forecasts, the conditional volatility forecasts, and the dynamic covariances im-
plied by the CCC model of Bollerslev (1990). In the CCC model, the dynamics
of covariances are driven by the time-variation in the conditional volatilities.
By design, therefore, the advantage of this setting is that the optimal weights
will vary across models only to the extent that forecasts of the conditional mean
and volatility will vary, which is precisely what the empirical models provide.
Indeed, introducing multivariate stochastic volatility models for capturing the
dynamic heteroskedasticity of the covariances of exchange rate returns re-
mains an important extension to this line of research. Multivariate stochastic
volatility models are high dimensional and their estimation is computation-
ally challenging (e.g., Chib, Nardari, and Shephard, 2006). Additionally, the
dynamic conditional correlation (DCC) model of Engle (2002) has yet to be
examined in a Bayesian framework. We will revisit this issue in future research.
7. Conclusion
This paper draws from three separate yet related strands of international finance
literature. A large body of empirical research finds that models that condition on
monetary fundamentals cannot outperform the naive random walk model in out-
of-sample forecasting of exchange rates. Despite the increasing sophistication
of the econometric techniques implemented and the improving quality of the
data sets utilized, evidence of exchange rate predictability remains elusive. A
second and related research strand indicates that the rejection of the risk-neutral
FX efficient market hypothesis implies that exchange rate movements can be
predicted using information contained in forward premiums. Finally, financial
economists agree that exchange rate volatility is predictable by specifying either
GARCH or stochastic volatility innovations.
Prior research in this area has largely relied on standard statistical measures
of forecast accuracy. In this paper, we complement this approach in two critical
3523
aspects. First, in assessing the predictive performance of the set of empirical

exchange rate models, we implement a Bayesian methodology that explicitly
accounts for parameter and model uncertainty. Second, we provide a com-

prehensive economic evaluation of the models in the context of dynamic asset
allocation strategies. In doing so, our study contributes to the growing empirical
literature on exchange rate predictability in the following manner. We assess
the economic value of exchange rate forecasts derived from empirical models
that condition on information contained in either monetary fundamentals or
forward premiums. This is done in a framework that allows for time-varying
volatility. The empirical exchange rate models are set against the naive random
walk benchmark. Finally, we evaluate the performance of combined forecasts
based on BMA.
Our results provide robust evidence against the random walk (no predict-
ability) benchmark, and therefore our empirical findings reinforce the notion
that exchange rates are predictable. Specifically, we find that the predictive abil-
ity of the forward premium has substantial economic value in a dynamic port-
folio allocation context and that stochastic volatility significantly outperforms
the constant variance and GARCH(1,1) models irrespective of the conditional
mean specification. Combined forecasts formed using BMA also substantially
outperform the random walk. These results are robust to reasonably high trans-
action costs and hold for all currencies both in-sample and out-of-sample. In
short, these findings suggest that the random walk hypothesis as applied to ex-
change rates might have been overstated, while at the same time they justify the
widespread use of forward bias and volatility timing strategies in the practice
of currency management.
Appendix: Bayesian MCMC Estimation

We perform Bayesian MCMC estimation of the parameters of the empirical exchange rate models
by constructing a Markov chain whose limiting distribution is the target posterior density. This
Markov chain is a Gibbs sampler in which all parameters are drawn sequentially from their full
conditional posterior distribution. The chain is then iterated and the sampled draws, beyond a
burn-in period, are treated as variates from the target posterior distribution.
The linear regression algorithm

In the Bayesian LR model, we need to estimate θ = {θ1 , θ2 }, where θ1 = {α, β} is the set of the
conditional mean parameters, and θ2 = {v −2 } is the constant precision defined as the inverse of the
variance. We define the following priors: for θ1 = {α, β} we assumea Normalprior N (θ1 , V ), where
ν 2s −2
θ1 = 02 and V = I2 ; for θ2 = {v −2 } we assume a prior Gamma 2 , ν with mean s −2 = 1,
and degrees of freedom ν = 2. The posterior distributions are summarized in the following simple
Gibbs algorithm (for more details see Koop, 2003):
1. Initialize θ2 . −1 −1

2. −1 θ1 from θ1 | s, θ2 ∼ N θ1 , V , where V = V + θ2 X X
Sample and θ1 =
V V θ1 + θ2 X s .
3524

ν 2s −2
3. Sample θ2 from θ2 | s, θ1 ∼ Gamma 2, ν , where ν= T +ν and s2 =
(s−X θ1 ) (s−X θ
1 )+νs 2
ν .
4. Go to Step 2 and iterate 100, 000 times beyond a burn-in of 20, 000 iterations.

The GARCH(1,1) algorithm
In the Bayesian GARCH(1,1) model, we need to estimate θ = {θ1 , θ2 }, where θ1 = {α, β} is the set
of the conditional mean parameters, and θ2 = {ω, γ1 , γ2 } are the conditional variance parameters.
We ensure that the conditional
variance is covariance stationary by specifying ω as a lognormal
prior: ω ∼ logN w, W , with w = −1 and W = 2. The prior specification is completed by

assuming γ1 ∼ Beta g 1 , G 1 and γ2 ∼ Beta g 2 , G 2 , where g 1 = 40, G 1 = 5, g 2 = 2, and
G 2 = 40. These hyperparameters imply a mean of 0.89 and 0.05 for γ1 and γ2 , respectively. The
algorithm is summarized below:
1. Initialize θ1 and transform the data intost∗ = (st − α − βxt−1 ) .

2. Sample the variance parameters θ2 from their full conditional posterior density: θ2 | s ∗ , θ1 .
This posterior density is not available analytically. We compute the log-likelihood of the
transformed data st∗ as function of θ2 (conditional on θ1 ) and then we optimize the conditional
log-posterior. We generate a proposal from a t-distribution t (m, V, ξ) , where m is the mode,
V is the inverse of the negative Hessian, and ξ a tuning parameter. The proposal is then
accepted according to the independence chain Metropolis-Hasting algorithm (e.g., Chib and
Greenberg, 1995).
3. Sample all the conditional mean coefficients θ1 | s, θ2 using a precision-weighted average
of a set of normal priors and the normal likelihood conditional on θ2 .
4. Update the data st∗ = (st − α − βxt−1 ) .
5. Go to Step 2 and iterate 5000 times beyond a burn-in of 1000 iterations.
The stochastic volatility algorithm

In the Bayesian SV model, we need to estimate θ = {θ1 , θ2 }, where θ1 = {α, β} is the set of the
conditional meanparameters,
and θ2 = {μ, φ, σ2 } are the conditional log-variance parameters. The
prior for μ is N m, M with m = 1 and M = 25. Following Kim, Shephard, and Chib (1998),
we formulate the prior for φ in terms of φ = 2φ∗ − 1, where φ∗ is distributed as Beta( f , F).
This implies that the prior on φ ∈ (−1, 1) is p(φ) = κ{0.5(1 + φ} f −1 {0.5(1 − φ} F−1 , f , F > 0.5,
( f +F)
where κ = 0.5 ( f )+(F) . Specifying f = 20 and F = 1.5 yields a mean of 0.86 with a variance of

0.01. For σ, the prior is inverse gamma IG s, S with s = 3 and S = 2.5 so that the distribution has
a mean of 0.20 with a variance of 0.006. The parameters of the SV model are estimated using the
Bayesian MCMC algorithm of Chib, Nardari, and Shephard (2002), which builds on the procedures
developed by Kim, Shephard, and Chib (1998), and is summarized below:

1. Initialize θ, mx, and transform the data into st∗ = ln (st − α − βxt−1 )2 + c , c = 0.001
to put the model in state-space form. The “offset” constant c eliminates the inlier problem.
2. Sample the log-variance parameters θ2 from their full conditional posterior density: θ2 | s ∗ ,
mx. This posterior density is not available analytically. We use the Kalman filter to compute the
log-likelihood of the transformed data st∗ as a function of θ2 (conditional on mxt ) and then
optimize the conditional log-posterior. We generate a proposal from a t-distribution t (m, V, ξ),
where m is the mode, V is the inverse of the negative Hessian, and ξ a tuning parameter. The
proposal is then accepted according to the independence chain Metropolis-Hastings algorithm
(e.g., Chib and Greenberg, 1995).
3. Sample the log-variance vector {h t } in one block from the posterior distribution: h | s ∗ ,
mx, θ2 . This step uses the de Jong and Shephard (1995) simulation smoother, which is an
3525
algorithm designed for efficient sampling of the state vector in a state-space model. See also
Carter and Kohn (1994).
4. Sample all the conditional mean coefficients θ1 from θ1 | s, h using a precision-weighted
average of a set of normal priors
and the normal likelihood
conditional on h. Then update the

transformed data st∗ = ln (st − α − βxt−1 )2 + c , c = 0.001.
5. Finally, sample the mixture indicator variable mx | s ∗ , h, θ directly from its posterior:

Pr mxt | st∗ , h t ∝ Pr (mxt ) f N st∗ | h t + m mxt , υ2mxt , t≤T

where m mxt , υ2mxt are the means and variances of the seven-component mixture of normal
densities that are used to approximate the log χ2 (1) distribution (see Kim, Shephard, and
Chib, 1998).
6. Go to Step 2 and iterate 5000 times beyond a burn-in of 1000 iterations.
The particle filter

The particle filter of Pitt and Shephard (1999) generates a sample from the density h t | Ft , θ. This
is a nontrivial task performed by an auxiliary sampling–importance resampling algorithm. The SV
application of the algorithm is detailed in Chib, Nardari, and Shephard (2002) and sketched below:

∗j
from (h t−1 | Ft−1 , θ) calculate:
j
1. Given a sample h 1t−1 , . . . , h t−1 M h t = μ + φ h t−1 − μ ,

∗j
ω j = N st | α + βxt−1 , exp( h t ) , for j = 1, . . . , M. Sample R = 10, 000 times the in-

tegers 1, 2, . . . , M = 2000 with probability proportional to ω j . Let the sampled indices be
∗k1 ∗k R
k1 , . . . , k R and associate these with h t , . . . , h t .
2. For each value of k j from Step 1, simulate the values h 1t , . . . , h tR from the volatility process
k
∗j j j
as: h t = μ + φ h t−1 j
− μ + σηt , j = 1, . . . , R, where ηt ∼ N (0, 1).
∗1
3. Resample the values h t , . . . , h ∗R t M times with replacement using probabilities propor-

∗j
N st |α+βxt−1 ,exp h t
tional to
∗k j
, for j = 1, . . . , R, to produce the desired filtered sample
N st |α+βxt−1 ,exp h t

h 1t , . . . , h t
M from (h t | Ft , θ).
References
Abhyankar, A., L. Sarno, and G. Valente. 2005. Exchange Rates and Fundamentals: Evidence on the Economic
Value of Predictability. Journal of International Economics 66:325–48.
Andersen, T. G. 1996. Return Volatility and Trading Volume: An Information Flow Interpretation of Stochastic
Volatility. Journal of Finance 51:169–204.
Avramov, D. 2002. Stock Return Predictability and Model Uncertainty. Journal of Financial Economics 64:423–
58.
Backus, D. K., A. W. Gregory, and C. I. Telmer. 1993. Accounting for Forward Rates in Markets for Foreign
Currency. Journal of Finance 48:1887–1908.
Backus, D. K., S. Foresi, and C. I. Telmer. 2001. Affine Term Structure Models and the Forward Premium
Anomaly. Journal of Finance 56:279–304.
Baks, K., A. Metrick, and J. A. Wachter. 2001. Should Investors Avoid All Actively Managed Mutual Funds? A
Study in Bayesian Performance Evaluation. Journal of Finance 56:45–85.
Bansal, R. 1997. An Exploration of the Forward Premium Puzzle in Currency Markets. Review of Financial
Studies 10:369–403.
3526
Bansal, R., and M. Dahlquist. 2000. The Forward Premium Puzzle: Different Tales from Developed and Emerging
Economies. Journal of International Economics 51:115–44.
Bansal, R., A. R. Gallant, R. Hussey, and G. Tauchen. 1995. Nonparametric Estimation of Structural Models for
High Frequency Currency Market Data. Journal of Econometrics 66:251–87.

Barberis, N. 2000. Investing for the Long Run When Returns Are Predictable. Journal of Finance 55:225–64.
Bates, J. M., and C. W. J. Granger. 1969. The Combination of Forecasts. Operations Research Quarterly
20:451–68.
Bauer, G. H. 2001. Conditional Currency Hedging and Asset Market Shocks. Working Paper, University of
Rochester.
Bekaert, G. 1996. The Time-Variation of Risk and Return in Foreign Exchange Markets: A General Equilibrium
Perspective. Review of Financial Studies 9:427–70.
Bekaert, G., and R. J. Hodrick. 1993. On Biases in the Measurement of Foreign Exchange Risk Premiums.
Journal of International Money and Finance 12:115–38.
Bekaert, G., and R. J. Hodrick. 2001. Expectations Hypotheses Tests. Journal of Finance 56:115–38.
Bekaert, G., R. J. Hodrick, and D. Marshall. 1997. The Implications of First-Order Risk Aversion for Asset
Market Risk Premiums. Journal of Monetary Economics 40:3–39.
Berkowitz, J., and L. Giorgianni. 2001. Long-Horizon Exchange Rate Predictability? Review of Economics and
Statistics 83:81–91.
Bilson, J. F. O. 1981. The “Speculative Efficiency” Hypothesis. Journal of Business 54:435–51.
Bollerslev, T. 1986. Generalized Autoregressive Conditional Heteroskedasticity. Journal of Econometrics

31:307–27.
Bollerslev, T. 1987. A Conditionally Heteroskedastic Time Series Model for Speculative Prices and Rates of
Return. Review of Economics and Statistics 69:542–47.
Bollerslev, T. 1990. Modeling the Coherence in Short-Run Nominal Exchange Rates: A Multivariate Generalized
ARCH Model. Review of Economics and Statistics 72:498–505.
Boudoukh, J., M. Richardson, and R. F. Whitelaw. 2006. The Information in Long-Maturity Forward Rates:
Implications for Exchange Rates and the Forward Premium Anomaly. Working Paper, New York University.
Carter, C. K., and R. Kohn. 1994. On Gibbs Sampling for State Space Models. Biometrika 81:541–53.
Chib, S. 1995. Marginal Likelihood from the Gibbs Output. Journal of the American Statistical Association
90:1313–21.
Chib, S., and E. Greenberg. 1995. Understanding the Metropolis-Hastings Algorithm. American Statistician
49:327–35.
Chib, S., and I. Jeliazkov. 2001. Marginal Likelihood from the Metropolis-Hastings Output. Journal of the
American Statistical Association 96:270–81.
Chib, S., F. Nardari, and N. Shephard. 2002. Markov Chain Monte Carlo Methods for Stochastic Volatility
Models. Journal of Econometrics 108:281–316.
Chib, S., F. Nardari, and N. Shephard. 2006. Analysis of High Dimensional Multivariate Stochastic Volatility
Models. Journal of Econometrics 134:341–71.
Clarida, R. H., L. Sarno, M. P. Taylor, and G. Valente. 2003. The Out-of-Sample Success of Term Structure
Models as Exchange Rate Predictors: A Step Beyond. Journal of International Economics 60:61–83.
Clarida, R. H., L. Sarno, M. P. Taylor, and G. Valente. 2006. The Role of Asymmetries and Regime Shifts in the
Term Structure of Interest Rates. Journal of Business 79:1193–1225.
3527
Clarida, R. H., and M. P. Taylor. 1997. The Term Structure of Forward Exchange Premiums and the Forecastability
of Spot Exchange Rates: Correcting the Errors. Review of Economics and Statistics 79:353–61.
Cremers, M. 2002. Stock Return Predictability: A Bayesian Model Selection Perspective. Review of Financial
Studies 15:1223–49.

de Jong, P., and N. Shephard. 1995. The Simulation Smoother for Time Series Models. Biometrika 82:339–50.
Della Corte, P., L. Sarno, and D. L. Thornton. 2008. The Expectations Hypothesis of the Term Structure of Very
Short-Term Rates: Statistical Tests and Economic Value. Journal of Financial Economics, forthcoming.
Diebold, F. X. 1998. The Past, Present and Future of Macroeconomic Forecasting. Journal of Economic Per-
spectives 12:175–92.
Diebold, F. X. 2004. Elements of Forecasting, 3rd ed. Louiseville, Canada: South-Western College Publishing.
Diebold, F. X., and P. Pauly. 1990. The Use of Prior Information in Forecast Combination. International Journal
of Forecasting 6:503–08.
Engel, C. 1996. The Forward Discount Anomaly and the Risk Premium: A Survey of Recent Evidence. Journal
of Empirical Finance 3:123–92.
Engel, C., and K. D. West. 2005. Exchange Rates and Fundamentals. Journal of Political Economy 113:485–517.
Engle, R. F. 2002. Dynamic Conditional Correlation: A Simple Class of Multivariate Generalized Autoregressive
Conditional Heteroskedasticity Models. Journal of Business and Economic Statistics 20:339–50.
Fama, E. F. 1984. Forward and Spot Exchange Rates. Journal of Monetary Economics 14:319–38.
Faust, J., J. H. Rogers, and J. Wright. 2003. Exchange Rate Forecasting: The Errors We’ve Really Made. Journal
of International Economics 60:35–59.
Fernandez-Villaverde, J., and J. F. Rubio-Ramirez. 2004. Comparing Dynamic Equilibrium Models to Data:
A Bayesian Approach. Journal of Econometrics 123:153–87.
Fleming, J., and C. Kirby. 2003. A Closer Look at the Relation Between GARCH and Stochastic Autoregressive
Volatility. Journal of Financial Econometrics 1:365–419.
Fleming, J., C. Kirby, and B. Ostdiek. 2001. The Economic Value of Volatility Timing. Journal of Finance
56:329–52.
Flood, R. B., and A. K. Rose. 2002. Uncovered Interest Parity in Crisis. International Monetary Fund Staff
Papers 49:252–66.
Froot, K. A., and R. H. Thaler. 1990. Foreign Exchange. Journal of Economic Perspectives 4:179–92.
Galati, G., and M. Melvin. 2004. Why Has FX Trading Surged? Explaining the 2004 Triennial Survey. BIS
Quarterly Review 4:67–74.
Geweke, J. 1989. Bayesian Inference in Econometric Models Using Monte Carlo Integration. Econometrica
57:1317–39.
Gomez, V., and A. Maravall. 2000. Seasonal Adjustment and Signal Extraction in Economic Time Series, in:
D. Pena, G. C. Tiao, and T. R. Tsay (eds.), A Course in Time Series Analysis. New York: Wiley.
Groen, J. J. J. 2000. The Monetary Exchange Rate Model as a Long-Run Phenomenon. Journal of International
Economics 52:299–319.
Hai, W., N. C. Mark, and Y. Wu. 1997. Understanding Spot and Forward Exchange Rate Regressions. Journal
of Applied Econometrics 12:715–34.
Han, Y. 2006. Asset Allocation with a High Dimensional Latent Factor Stochastic Volatility Model. Review of
Financial Studies 19:237–71.
3528
Harvey A. C., E. Ruiz, and N. Shephard. 1994. Multivariate Stochastic Variance Models. Review of Economics
Studies 61:247–64.
Hlawitschka, W. 1994. The Empirical Nature of Taylor-Series Approximations to Expected Utility. American
Economic Review 84:713–19.

Hodrick, R. J. 1987. The Empirical Evidence on the Efficiency of Forward and Futures Foreign Exchange
Markets. London: Harwood.
Hoeting, J. A., D. Madigan, A. E. Raftery, and C. T. Volinsky. 1999. Bayesian Model Averaging: A Tutorial.
Statistical Science 14:382–417.
Kandell, S., and R. F. Stambaugh. 1996. On the Predictability of Stock Returns: An Asset-Allocation Perspective.
Journal of Finance 51:385–424.
Karolyi, G. A., and R. M. Stulz. 2003. Are Financial Assets Priced Locally or Globally? in G. Constantinides,
M. Harris, and R. M. Stulz (eds.), Handbook of Economics of Finance. Amsterdam: Elsevier.
Kim, S., N. Shephard, and S. Chib. 1998. Stochastic Volatility: Likelihood Inference and Comparison with ARCH
Models. Review of Economic Studies 65:361–93.
Koop, G. 2003. Bayesian Econometrics. New York: Wiley.
Leamer, E. E. 1978. Specification Searches. New York: Wiley.
Levich, R. M. 1985. Empirical Studies of Exchange Rates: Price Behavior, Rate Determination, and Market
Efficiency, in R. W. Jones and P. B. Kenen (eds.), Handbook of International Money and Finance. Amsterdam:
Elsevier.
Lewis, K. K. 1995. Puzzles in International Financial Markets, in G. M. Grossman and K. Rogoff (eds.),
Handbook of International Economics. Amsterdam: Elsevier.
Lo, A.W. 2002. The Statistics of Sharpe Ratios. Financial Analysts Journal 58:36–52.
Lucas, R. E. 1982. Interest Rates and Currency Prices in a Two-Country World. Journal of Monetary Economics
10:335–59.
Lustig, H., and A. Verdelhan. 2007. The Cross-Section of Currency Risk Premia and US Consumption Growth
Risk. American Economic Review 97:89–117.
Lyons, R. K. 2001. The Microstructure Approach to Exchange Rates. Cambridge, MA: MIT Press.
Mark, N. C. 1995. Exchange Rates and Fundamentals: Evidence on Long-Horizon Predictability. American
Economic Review, 85, 201–218.
Mark, N. C., and D. Sul. 2001. Nominal Exchange Rates and Monetary Fundamentals: Evidence from a Small
Post-Bretton Woods Panel. Journal of International Economics 53:29–52.
Marquering, W., and M. Verbeek. 2004. The Economic Value of Predicting Stock Index Returns and Volatility.
Journal of Financial and Quantitative Analysis 39:407–29.
Meese, R. A., and K. Rogoff. 1983. Empirical Exchange Rate Models of the Seventies: Do They Fit Out of
Sample? Journal of International Economics 14:3–24.
Obstfeld, M., and K. Rogoff. 1995. Exchange Rate Dynamics Redux. Journal of Political Economy 103:624–60.
Pitt, M. K., and N. Shephard. 1999. Filtering via Simulation Based on Auxiliary Particle Filters. Journal of the
American Statistical Association 94:590–99.
Rapach, D. E., and M. E. Wohar. 2002. Testing the Monetary Model of Exchange Rate Determination: New
Evidence from a Century of Data. Journal of International Economics 58:359–85.
Sarno, L., G. Valente, and H. Leon. 2006. Nonlinearity in Deviations from Uncovered Interest Parity: An
Explanation of the Forward Bias Puzzle. Review of Finance 10:443–82.
3529
Sarno, L., G. Valente, and M. E. Wohar. 2004. Monetary Fundamentals and Exchange Rate Dynamics Under
Different Nominal Regimes. Economic Inquiry 42:179–93.
Shanken, J., and A. Tamayo. 2001. Risk, Mispricing, and Asset Allocation: Conditioning on Dividend Yield.
NBER Working Paper 8666.

Sims, C. A. 1988. Bayesian Skepticism on Unit Root Econometrics. Journal of Economic Dynamics and Control
12:463–74.
Tauchen, G. E., and M. Pitts. 1983. The Price Variability-Volume Relationship on Speculative Markets. Econo-
metrica 51:485–505.
Timmermann, A. 2006. Forecast Combinations, in G. Elliott, C. W. J. Granger, and A. Timmermann (eds.),

Handbook of Economic Forecasting. Amsterdam, The Netherlands: Elsevier.
Vrontos, I. D., P. Dellaportas, and D. N. Politis. 2000. Full Bayesian Inference for GARCH and EGARCH
Models. Journal of Business and Economic Statistics 18:187–98.
West, K. D., H. J. Edison, and D. Cho. 1993. A Utility-Based Comparison of Some Models of Exchange Rate
Volatility. Journal of International Economics 35:23–45.
3530

Della Corte 2008

Uploaded by

Copyright:

Available Formats

Della Corte 2008

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Della Corte 2008

Uploaded by

Copyright:

Available Formats

An Economic Evaluation of Empirical

Exchange Rate Models

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

This paper provides a comprehensive evaluation of the short-horizon predictive ability

Forecasting exchange rates using models that condition on economically mean-

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

highly satisfactory approximation to a wide range of more sophisticated utility

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

empirical results are reported in Section 5, followed by robustness checks in

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

1.2 The spot-forward exchange rate relation

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

st = α + β ( f t−1 − st−1 ) + u t , (5)

where u t is a disturbance term.

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

2. Modeling FX Returns and Volatility

2.1 The conditional mean

st = α + βxt−1 + u t , u t = vt εt , εt ∼ N I D (0, 1) . (6)

The next three model specifications condition on monetary fundamentals

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

2.2 The conditional variance

GARCH models, however, the assumption of a stochastic second moment in-

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

vt = exp (h t /2) , (8)

3. Measuring the Economic Value of Exchange Rate Predictability

3.1 FX models in a dynamic mean-variance framework

where wt is the K × 1 vector of portfolio weights on the risky assets, ι is a

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

3.2 Quadratic utility

is the period t + 1 gross return on his portfolio.

expected utility generated by a given level of initial wealth. Specifically, the

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

Average utility depends on taste for risk. In the absence of restrictions on δ,

3.3 Performance measures

wealth to yield a given level of U than a suboptimal model. Following Fleming,

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

3.4 The dynamic FX strategies

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

3.5 Transaction costs

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

4.1 Bayesian Markov chain Monte Carlo estimation

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

4.2 Model risk and posterior probability

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

4.3 Combined forecasts

4.3.1 The BMA strategy. In the context of our Bayesian approach, it is

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

5.1 FX data and descriptive statistics

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

Panel A: Percent Returns

Panel B: Absolute Percent Returns

The three specifications of monetary fundamentals predictors display high

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

5.2 Estimation of exchange rate models

Downloaded from https://academic.oup.com/rfs/article-abstract/22/9/3491/1569619 by Iowa State University user on 23 January 2019

Panel B: Bayesian GARCH(1,1)

α 0.018 0.027 0.005 0.017 −0.101

Panel C: Bayesian stochastic volatility

α 0.048 0.046 0.022 0.022 −0.045