Jofi 13190

THE JOURNAL OF FINANCE • VOL. , NO.
0 • NOVEMBER 2022
Pricing Currency Risks
MIKHAIL CHERNOV, MAGNUS DAHLQUIST, and LARS LOCHSTOER*
ABSTRACT
The currency market features a small cross-section, and conditional expected returns
can be characterized by few signals: interest differential, trend, and mean reversion.
We exploit these properties to construct the ex ante mean-variance efficient portfolio
of individual currencies. The portfolio is updated in real time and prices all promi-
nent currency trading strategies, conditionally and unconditionally. The fraction of
risk in these assets that does not affect their risk premiums is at least 85%. Extant
explanations of carry strategies based on intermediary capital or global volatility are
related to these unpriced components, while consumption growth is related to the
priced component of returns.
IN THIS PAPER, WE ARGUE that research on cross-sectional (CS) currency

pricing can depart fruitfully from the factor mining path established by the
equity literature. That is, because a direct solution of the mean-variance
optimization problem in the presence of conditional information is feasible
in the exchange-rate setting. The value of this approach, if it is empirically
successful, is obvious: one obtains a measure of the so-called uncondi-
tional mean-variance efficient (UMVE) portfolio of currencies that prices all
* Mikhail Chernov is at UCLA Anderson School. Magnus Dahlquist is at Stockholm School of
Economics. Lars Lochstoer is at UCLA Anderson School. We thank the Editor Stefan Nagel, the
Associate Editor, and two referees for their valuable feedback. We also thank Craig Burnside,
Kent Daniel, Xiang Fang, Valentin Haddad, Zhengyang Jiang, Chris Jones, Serhiy Kozak, Lukas
Kremens, Francis Longstaff, Sydney Ludvigson, Hanno Lustig, Thomas Maurer, Tyler Muir, Nick
Roussanov, Andrea Vedolin, Adrien Verdelhan, and Irina Zviadadze, as well as participants in
seminars and conferences sponsored by AQR, the 2021 EFA, LSE, the 2021 Spring NBER AP
meeting, the 2021 SFS Cavalcade, Stockholm School of Economics, UCLA, University of Vienna,
USC, the 2021 University of Connecticut Finance Conference, the 2021 Vienna Symposium on For-
eign Exchange Markets, the 2021 WFA, and the 2021 World Symposium on Investment Research
for their comments on earlier drafts. We thank Alessia Menichetti and Felix Wilke for excellent
research assistance. Dahlquist gratefully acknowledges support from the Jan Wallander and Tom
Hedelius Foundation, and the Swedish House of Finance at the Stockholm School of Economics.
We have read The Journal of Finance disclosure policy and have no conflicts of interest to disclose.
Correspondence: Magnus Dahlquist, Stockholm School of Economics; Drottninggatan 98, SE
11160 Stockholm, Sweden; e-mail: magnus.dahlquist@hhs.se
This is an open access article under the terms of the Creative Commons Attribution-
NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided
the original work is properly cited, the use is non-commercial and no modifications or adaptations
are made.
DOI: 10.1111/jofi.13190
© 2022 The Authors. The Journal of Finance published by Wiley Periodicals LLC on behalf of
American Finance Association
1
2 The Journal of Finance®
admissible dynamic trading strategies in these currencies unconditionally. As

a result, as Hansen and Richard (1987) show, the UMVE portfolio correctly
prices the full cross-section of currencies and trading strategies associated
with them conditionally as well. Measurement of the UMVE portfolio can
therefore help direct research that seeks to understand currency pricing.
UMVE portfolio construction requires two critical ingredients: estimates of
the conditional mean and the conditional covariance matrix of currency ex-
cess returns. This is where exchange rates have an advantage over equities.
First, the size of the cross-section of exchange rates is small, not exceeding
40, as compared to thousands of stocks. This feature dramatically affects the
precision of the covariance matrix regardless of the specifics of the estima-
tion method. Second, exchange rates have a particular economic structure
that helps one hypothesize the functional form of the conditional means. Re-
search that goes back decades suggests the importance of three key ingredients
of conditional expectations. Specifically, Covered Interest Parity (CIP, Keynes
(1923)), Uncovered Interest Parity (UIP, Porter (1971)), and Random Walk Hy-
pothesis (RWH, Meese and Rogoff (1983)) lead us to the interest rate differen-
tial. Purchasing Power Parity (PPP, Cassel (1918)) suggests a measure of mean
reversion. And research on weak-form market efficiency in nominal exchange
rates (Cornell and Dietrich (1978)) indicates a measure of trend. These three
factors remain the pillars of the modern currency trading strategies under the
names of carry, value, and momentum, respectively.
We work with a sample of G10 currencies, the most commonly used data
in the literature, with monthly returns based on currency forward rates from
January 1985 to May 2020. We construct monthly conditional expected excess
returns using the three aforementioned signals. The loading on the signal re-
lated to the interest rate differential is fixed to be consistent with the RWH,
while the loadings on the other signals are estimated on an expanding basis
via panel regressions. We construct a conditional covariance matrix of currency
excess returns using daily data within each month. Importantly, the approach
uses data that are available to investors in real time.
If the conditioning information that drives these estimates is sufficiently
rich, the resulting UMVE portfolio will in theory price not only the excess re-
turns on individual currencies but also any dynamic strategy, including those
considered in the literature. Yet, because these estimates are out-of-sample
(OOS) forecasts and because we have to assume specific forecasting models, it
is not given that the resulting UMVE portfolio would perform well (e.g., have
a high Sharpe ratio [SR] or price the test assets). Therefore, to validate our
UMVE construction, we test the implied one-factor model using a broad set
of test assets and show that it is not rejected, while competing alternatives
are rejected.
We consider nine leading trading strategies in addition to individual cur-
rencies as test assets. The dollar strategy represents the currency version of
“the market.” The dollar carry strategy goes long or short the dollar strat-
egy depending on the average interest rate differential, or, more precisely, the
average currency forward discount. Next, we evaluate the CS and time-series
Pricing Currency Risks 3
(TS) carry strategies. These strategies are followed by two versions of CS and
TS momentum depending on whether one uses the past month or year as the
measure of past performance. We also study CS value, which sorts currencies
based on the real exchange rate (RER). As emphasized by Asness, Moskowitz,
and Pedersen (2013) and Koijen et al. (2018), besides being the most common
strategies in the foreign exchange market, they are more broadly considered
in other asset classes. Lastly, because the Hansen and Richard (1987) result
requires correct unconditional pricing of literally all trading strategies, we con-
sider additional 81 strategies that are constructed by changing the base cur-
rency from U.S. dollar (USD) to one of the nine other currencies. For instance,
the Japanese yen (JPY) strategy would involve averaging returns with respect
to the JPY (and then expressing these returns in USD terms). We also con-
sider the effect of changing the numeraire (e.g., converting the dollar strategy
returns to JPY).
The UMVE is tradeable in real time, has an SR in excess of one, and, conve-
niently, can be compared directly to existing trading strategies using standard
regression-based tests. We verify that it prices both excess returns on nine
individual currencies versus the USD and the nine above-mentioned trading
strategies via the unconditional and conditional pricing tests of Gibbons, Ross,
and Shanken (1989) and Chernov, Lochstoer, and Lundeby (2022), respectively.
We also use the above-mentioned additional strategies with a non-USD base
or measurement currency and again fail to reject the estimated UMVE.
As an alternative, we consider a variance-based approach to construct con-
ditional expectations of currency returns. This approach is motivated by the
work of Lustig, Roussanov, and Verdelhan (2011) and Verdelhan (2018), which
implies that the dollar and carry strategies explain (co)variances of currency
returns in both the cross-section and the timeseries. This observation implies
that the dollar and carry betas should predict currency returns. We demon-
strate that the resulting OOS forecast is dramatically different from our advo-
cated conditional mean. The resulting UMVE portfolio has a low SR of 0.3 and
is rejected by our tests.
Having validated the UMVE portfolio, we turn to evaluating its properties.
First, we find that it explains only a small fraction of the time-series vari-
ation in strategy returns. For instance, less than 1% of the variation in the
dollar strategy returns (the currency “market portfolio”) are priced, while the
same for carry strategies is at most 15%.1 Thus, there are large unpriced
1 To interpret the evidence, it is convenient to think of the SDF as consisting of two components.
The first is the UMVE expressed in terms of a numeraire that is not currency-specific (e.g., a
world currency index). The second reflects the change in numeraire to a given currency. We show
that because the SR of the UMVE is slightly greater than one while the volatility of depreciation
rates is about 0.1 (per annum), it must be the case that most of the valuation effect stems from
the UMVE component of the stochastic discount factor (SDF). Furthermore, strategy returns are
exposed to common sources of variation in currencies that are not related to the UMVE. These
common sources are important for explaining the variation in strategy returns (e.g., Verdelhan
(2018)), but not their risk premiums as per the above. As a result, regressions of strategy returns
on the UMVE result in low R2 s. See the Internet Appendix for details. The Internet Appendix is
available in the online version of the article on The Journal of Finance website.
components in strategy returns and we show that hedging out these unpriced
components in an OOS manner dramatically increases the SR of the hedged
strategy returns. As an OOS model check, we verify that the average returns
to the unpriced components are indeed not significantly different from zero.
Related, we document that the portfolios that are the most important for
time-series covariation among currencies and the currency trading strategies
have little relation to the UMVE portfolio. In other words, the main factors
that drive the covariance matrix of returns are not an important source of risk
for the marginal currency investor. For instance, the first two principal compo-
nents (PCs) of individual currency returns capture 75% of the return variation
and are strongly associated with the dollar and CS carry strategies, consistent
with Verdelhan (2018).2 However, these PCs explain only around 5% of the
variation in UMVE returns. In fact, all nine of these PCs only explain around
10% of the UMVE returns, which indicates that conditional currency risk pre-
miums vary strongly and underscores the importance of correct conditioning.
The currency trading strategies we consider involve substantial timing of
individual currencies and therefore potentially account for these conditional
dynamics. We investigate whether the strategies’ PCs are capable of explain-
ing variation in the UMVE. The answer is no, although there is some improve-
ment. The first three PCs of the currency trading strategies explain around
20% of the variation in the UMVE returns. Using all of the PCs, this increases
to 50%. This evidence is consistent with the uncovered large unpriced compo-
nent of strategy returns. It also echoes the rejection of the UMVE constructed
using variance-based conditional expectations. Taken together, these results
suggest a large role of optimal timing implicit in the construction of the UMVE
above and beyond the timing already implicit in existing trading strategies.
General equilibrium models have struggled to explain exchange rate mo-
ments. The current generation of models based on habit (Verdelhan (2010)),
long-run risks (Bansal and Shaliastovich (2013), Colacito et al. (2018)), and
rare disasters (Farhi and Gabaix (2016)) tend to focus on explaining the for-
ward premium puzzle (and to less extent the carry trade). However, as our
analysis shows, the most popular carry strategy explains only about 15% of
the variation in our pricing factor. Further, the UMVE portfolio, which is per-
fectly negatively correlated with the SDF, has approximately zero skewness of
−0.05 and modest excess kurtosis of 3.07. This suggests a relatively modest
effect of disaster risk pricing in our sample. Thus, there is much left to explain
and our empirical model can help guide the specification and tests of future
models on currency risk and return.
We provide initial evidence along these lines by first regressing the UMVE
returns on various candidate explanatory factors. We find that quarterly and
three-year consumption growth, a proxy for long-run consumption risk, are
both significantly positively related to the UMVE returns, consistent with
Lustig and Verdelhan (2007) and Zviadadze (2017). We next find that the
2Aloosh and Bekaert (2022) and Greenaway-McGrevy et al. (2018) offer alternative approaches
to the analysis of the time-series behavior of exchange rates.
factors in the Fama and French (2015) five-factor model are only weakly re-
lated to the currency UMVE portfolio, with only the value (HML) and invest-
ment (CMA) factors having a significant relationship. However, here the R2 s
are only about 1%, so economically, there appears to be no relation between
priced risks in the equity and currency markets. These results extend those
in Burnside (2012), who finds a similar weak relation between carry trade re-
turns and equity factors, to the UMVE portfolio. We also find that intermedi-
ary capital factors and shocks to equity and currency variance are unrelated to
the UMVE returns. This may appear puzzling, given that previous literature
highlights these factors as potential explanations for the carry trade, but we
show that this correlation comes from the large unpriced component in carry
strategy returns.
Finally, we show that the conditional maximum SR (MSR) has a downward
trend over the sample, largely caused by a trend decline in interest differen-
tials across countries. The cyclical variation in the MSR is negatively related
to measures of the conditional variance of consumption and equity returns and
to currency depreciation. In sum, our paper provides a rich set of facts that can
help guide future equilibrium models of currency market risk and return.
A number of papers consider conditional mean-variance efficient portfolios
(CMVE) using the forward discount alone as a signal (Baz et al. (2001), Ack-
ermann, Pohl, and Schmedders (2017), Daniel, Hodrick, and Lu (2017), Mau-
rer, To, and Tran (2020), Maurer, To, and Tran (2022)).3 With the exception
of Maurer, To, and Tran (2022), these papers do not examine whether CMVE
explains other versions of carry or individual currency returns. We reject this
version of our model with respect to our full model, which includes mean re-
version and trend signals, using the Barillas and Shanken (2017) test. We also
reject it using the conditional pricing test, even when only using carry-based
strategies as test assets. Della Corte, Sarno, and Tsiakas (2009) depart from
the lone forward discount signal by considering three monetary fundamentals
variables. They perform an OOS analysis of exchange rate predictability in a
mean-variance framework where the allocation choice is between the risk-free
asset and one of the three currencies they consider.
In contemporaneous and independent work, Nucera, Sarno, and Zinna
(2022) use the risk-premium PC analysis of Lettau and Pelger (2020) to ex-
tract latent factors of a currency SDF. They find that the implied SDF includes
a strong dollar factor and two weak carry and momentum factors, whereas
evidence of a value factor is scant. This is in contrast to our results: we find
that the dollar strategy contains a small fraction of priced risk, the momentum
strategy contains a larger fraction, and the carry and value strategies contain
the largest fractions.
A large literature implements model-free SDF projections in the con-
text of unconditional pricing of assets. Many authors use Hansen and
3 As detailed later in this paper, there is an infinite number of CMVE portfolios, with UMVE
being one of them. UMVE is the only CMVE portfolio that satisfies the unconditional linear beta
pricing relation, which is a staple of cross-sectional asset pricing tests.
Jagannathan (1991) bounds as a diagnostic in their models.4 A variation on

the unconditional version of the Hansen and Jagannathan (1991) approach is
that of entropy minimization as advocated by Stutzer (1996) in the context
of derivatives pricing. Ghosh, Julliard, and Taylor (2019) develop and apply
this framework to cross-sectional asset pricing. Korsaye, Trojani, and Vedolin
(2020), Orlowski, Sokolovski, and Sverdrup (2021), and Sandulescu, Trojani,
and Vedolin (2021) are examples of international applications of this frame-
work. We depart from this work by constructing real-time conditional SDF
projections. Our approach allows for an evaluation of any trading strategies in
the set of assets involved in the projection.
Our strategies that are used as test assets have origins in the large litera-
ture on international asset pricing; such work includes Asness, Moskowitz, and
Pedersen (2013), Burnside, Eichenbaum, and Rebelo (2011), Burnside et al.
(2011), Daniel, Hodrick, and Lu (2017), Koijen et al. (2018), Lustig, Roussanov,
and Verdelhan (2011) (2014), Menkhoff et al. (2012b), and Moskowitz, Ooi, and
Pedersen (2012).
This paper is organized as follows. Section I discusses implementation and
testing of general linear factor models and the specific application to exchange
rates. Section II describes all the empirical results. Section III concludes.
I. Linear Factor Models and Exchange Rates

This section describes the theoretical underpinnings of the UMVE portfolio,
how we estimate its weights in the context of exchange rates, and how the
estimated UMVE portfolio can be tested.
A. The UMVE Portfolio

We seek to correctly price currency risks both conditionally and uncondition-
ally. As pointed out by Hansen and Richard (1987) and Jagannathan (1996),
one can achieve this by constructing the UMVE portfolio.
Suppose that we have N basis assets with an N × 1 vector of excess returns
e
Rt+1 . The conditional mean of this vector is μt = Et (Rt+1
e
) and its conditional
covariance matrix is t = Vt (Rt+1 ). An admissible trading strategy p in the
e
basis assets has an N × 1 vector of weights w pt that are determined only by

information available up until time t. The resulting excess portfolio return is
then R p,t+1 = wpt Rt+1
e
.
4 The SDF frontier associated with the unconditional mean-variance frontier could be used as
informationally efficient SDF bounds. An important literature, starting with Gallant, Hansen,
and Tauchen (1990), explores such frontiers in-sample. For instance, Ferson and Siegel (2003,
2009) explore efficient bounds based on the unconditional frontier. Bekaert and Liu (2004) argue
that using the optimal weights reported is robust to misspecification of the conditional moments
of returns since the solution still describes a valid portfolio strategy that expands the boundary
in-sample. In contrast, we do not explore the connection of the UMVE portfolio, a single point on
the UMV frontier, to these bounds, and our analysis is conducted OOS.
The UMVE portfolio is a dynamic trading strategy in these assets and

obtains the MSR, both conditionally and unconditionally. Ferson and Siegel
(2001) and Jagannathan (1996) show that the UMVE portfolio weights are
1
wt∗ = t−1 μt . (1)
1+ μt t−1 μt
∗
Denote the excess return on this portfolio by Rt+1 = wt∗ Rt+1
e
.
The UMVE portfolio accounts for all risks conditionally in the sense that the
following conditional linear beta pricing relationship holds for any admissible
strategy p:
∗
Et (R p,t+1 ) = β pt Et (Rt+1 ), (2)
∗ ∗
where β pt = Covt (R p,t+1 , Rt+1 )/Vt (Rt+1 ). The UMVE portfolio also implies the
unconditional linear beta pricing relation
∗
E(R p,t+1 ) = β p E(Rt+1 ), (3)
∗ ∗
where β p = Cov(R p,t+1 , Rt+1 )/V (Rt+1 ), for any p.
The latter result may seem surprising in light of Jagannathan and Wang
(1996), who point out that taking unconditional expectations of equation (2)
∗ ∗
yields E(R p,t+1 ) = E(β pt )E(Rt+1 ) + Cov(β pt , Et (Rt+1 )). That is, in general, a con-
ditional linear beta-pricing model does not imply an unconditional linear beta-
pricing model. However, the specific case of the UMVE portfolio is an exception.
Indeed, portfolios with weights that are proportional to t−1 μt are all CMVE
and thus satisfy equation (2). There is an infinite number of such portfolios be-
cause leverage does not affect the conditional SR. It is the particular time vari-
ation in leverage (market timing) implied by the scalar term (1 + μt t−1 μt )−1
in equation (1) that delivers the UMVE portfolio and the validity of the un-
conditional linear beta pricing in equation (3). Thus, opposite from what the
language might seem to imply, the UMVE is also CMVE, while the converse
need not be true.
There is a tight link between the UMVE and the SDF. Specifically,
∗ ∗ ∗
Mt+1 = 1 − (Rt+1 − E(Rt+1 )) (4)
has the property that

∗
Et (Mt+1 R p,t+1 ) = 0 (5)
for all admissible trading strategies p in the set of basis assets. This SDF can
be viewed as a conditional projection of the true SDF, Mt+1 , onto the payoff
space of the basis assets. Alternatively, it can be viewed as an unconditional
linear projection of the true SDF onto the payoff space of all admissible trading
strategies in the basis assets. Since our focus is on excess returns only, we
normalize the unconditional mean of this SDF to one. See Chernov, Lochstoer,
and Lundeby (2022) for a recent distillation of these results.
As is well known, even though the true SDF is strictly positive under no-
arbitrage, the projected SDF can take negative values. However, this does
not matter for the pricing of the assets the projection is made onto. In fact,
∗
Mt+1 = Mt+1 + ηt+1 , where η has the property that Et (ηt+1 R p,t+1 ) = 0 for all ad-
missible trading strategies in the set of basis assets. See Cochrane (2005) for a
treatment of these relations.
B. Exchange Rates
In practice, estimating μt and t could be difficult or infeasible. For in-
stance, for equities, it is difficult to estimate t because N is large, and there
is not much consensus on estimating μt . As a result, the equity literature
has proposed various characteristic-sorted return factors that are presumed
to span the UMVE portfolio (see, e.g., Fama and French (1993)). Real-time
implementability of the trading strategies implied by these characteristics is
paramount because a look-ahead in factor construction biases tests of the null
hypothesis that a given factor model prices the cross-section of assets (see, e.g.,
Black, Jensen, and Scholes (1972), Lo and MacKinlay (1990)).
Starting with Lustig, Roussanov, and Verdelhan (2011), the literature on
currency risk and returns has followed the approach in equities by creating
factors based on characteristics such as the interest rate differential (carry), or
an equal-weighted portfolio of currencies (dollar), assuming that these factors
span the UMVE portfolio. This literature has also emphasized the real-time
implementability of the factors.
We note that exchange rates differ from equities in two very important re-
spects. First, the cross-section of “assets” is limited. Indeed, the largest data
panel considered in the literature does not exceed 40 currencies at any given
point in time (see, e.g., Hassan and Mano (2019), Lustig, Roussanov, and
Verdelhan (2011), Menkhoff et al. (2012a)), while most studies limit their at-
tention to 10 currencies representing developed economies. Second, a volumi-
nous literature on the forecasting of exchange rates suggests that few variables
outperform the RWH OOS.
∗
These observations prompt us to directly construct the returns Rt+1 using
the theoretical definition of the UMVE portfolio as given in equation (1). That
is, we estimate the conditional expected return and covariance matrix of the
∗
currency returns and use these inputs to construct Rt+1 . We do so in an OOS
fashion, where we only use data up to time t to estimate μt and t , so our
UMVE portfolio is implementable in real time.
C. Implementing the UMVE for Exchange Rates

Let the USD be the measurement (numeraire) currency, that is, all exchange
rates are expressed in USD per unit of foreign currency. Let Sti and Fti denote
the spot exchange rate and the one-month forward exchange rate of country
i, respectively.
The payoff of a forward contract (when buying one unit of the foreign cur-
i
rency) is St+1 − Fti . One common way to scale this payoff to define excess re-
turns is to divide by Fti :
ei
Rt+1 = (St+1
i
− Fti )/Fti . (6)
This definition implies that the amount of foreign currency bought is one “for-
ward” USD. Thus, this is an excess return to a trading strategy regardless of
whether CIP holds or not.
Next, we estimate t . First, for each month t, we compute the realized co-
variance matrix based on daily depreciation rates. Specifically, defining X j,t as
the vector of percentage changes in spot exchange rates over day j in month t,
we calculate the realized covariance matrix as

Dt

ˆt = X j,t X j,t ,
j=1
where Dt is the number of trading days in the month. Given a potentially

large cross-section and short time series, we apply the shrinkage method of
Ledoit and Wolf (2020) to ˆ t to improve the mean-squared error of this sam-
ple covariance estimate. Denote the resulting shrunk matrix by ˜ t . We then
apply a simple exponentially weighted average to t to arrive at the estimate
˜
of the conditional covariance matrix for month t + 1 currency percentage price
changes:
t = (1 − λ)
˜ t + λt−1 ,
where we set λ = 0.94 following, for example, the RiskMetrics model. Finally,
to get to the conditional covariance matrix of currency excess returns, we pre-
multiply and postmultiply this matrix by a diagonal matrix with the (i, i)th
element set to Sti /Fti .
Our starting point for the estimation of μt is the RWH for spot exchange
rates. The reason is simple: the hypothesis leads to an exceptionally robust
forecaster of exchange rates (Meese and Rogoff (1983)). This is also a natural
baseline given research in finance, starting with Baz et al. (2001), which uses
RWH to develop trading strategies.
The RWH implies that expected excess currency returns are given by

μit ≡ Et (Rt+1
ei
) = γ · Sti /Fti − 1 ,
with γ = 1. This is a particular violation of UIP, which posits γ = 0. We refer

to Sti /Fti − 1 as the (normalized) forward discount.
Next, we add mean reversion and trend signals for exchange rate forecast-
ing. See Bilson (1984) and Sweeney (1986), respectively, for early contributions
in the context of trading strategies. Thus, the use of the RWH, as well as mean
reversion and trend signals, are all concepts that were in the public domain at
the start of our sample in 1985.
Our mean-reversion signal is motivated by the literature on the role of RER

in forecasting and capturing risk premiums. The RER is defined as
Qti = Sti · Pti /Pt , (7)
where Pt and Pti are the U.S. and foreign consumer price index (CPI), re-
spectively. The weak form of PPP implies mean reversion in the RER. Thus,
when the RER is far from its long-run mean, it should forecast the currency
depreciation.
Combining the RER-based signal with the RWH goes as far back as Bilson
(1984). More recently, Chernov and Creal (2021) and Dahlquist and Pénasse
(2022) show that PPP implies that the RER forecasts nominal depreciation
rates and that the currency risk premium depends on the RER. Eichenbaum,
Johannsen, and Rebelo (2021) demonstrate that the RER outperforms the ran-
dom walk in the forecasting of exchange rates at horizons beyond one year.
As Jorda and Taylor (2012) emphasize, the RER’s long-run mean is not
a clearly defined object empirically. We divide each RER by its five-year
smoothed lag (specifically, the average RER from 4.5 to 5.5 years ago) as a
way to remove the dependence on the long-run mean while still preserving the
long-run nature of mean-reversion signals:
⎛ ⎞−1

6
i ≡ Qi · ⎝ 1
Q i
Qt−60+ ⎠ .
t t j
13
j=−6
In a last step, we cross-sectionally demean the signal at each time t to create a

cross-sectional ranking of “cheap” and “expensive” currencies. Specifically, our
signal is

N
i − 1
ziQt ≡ Q i .
Q (8)
t t
N
i=1
This definition has the virtue of removing any time and currency fixed effects.
The specific form of the reference point is motivated by the cross-sectional
model of Asness, Moskowitz, and Pedersen (2013), who select the specifics to
be comparable with the equity and commodity literatures.
Our trend signal is motivated by the academic and practitioner literature
on the use of moving averages in trading and forecasting exchange rates.
In contrast to the macro literature, the finance literature suggests that us-
ing past performance could be fruitful in improving trading performance (see,
e.g., Sweeney (1986), Kho (1996), Okunev and White (2003), and references
therein). Following this work, we add a simple trend signal, namely, a one-year
depreciation rate.
In summary, we forecast excess returns OOS via

μit = γti · Sti /Fti − 1 + δti · ziQt + φti · Sti /St−12
i
−1 . (9)
We set γti = 1 to match the RWH baseline. The coefficients δti and φti are rees-
timated each month t using historical exchange rates up to time t, as de-
tailed below. Estimating γti in a similar fashion does not materially change the
results.
An alternative approach to estimating conditional expectations of currency
returns would be to assume that currency risk premiums are driven by expo-
sures to the main drivers of currency return variance. Lustig, Roussanov, and
Verdelhan (2011) argue that the carry and dollar factors explain the cross-
section of expected currency returns. Verdelhan (2018) demonstrates that
these two risk factors capture most of the time-series variation in currency
returns and are closely related to the first two PCs of currency returns. Thus,
the conditional pricing relation Et (Reit+1 ) = βit λt suggests that one could fore-
cast currency returns via
μit = λtcarry · βitcarry + λtdollar · βitdollar , (10)
where a factor (carry or dollar) is a portfolio of the currencies, RFp,t+1 =

(wtF ) Rt+1
e
. The Internet Appendix describes computation of the betas. We esti-
carry
mate λt and λtdollar using data only up to time t. We refer to this specification
of the forecast as the variance-based expected return forecast, and consider
this in our analysis as well.
D. Testing Linear Factor Models

D.1. Tests
Equation (2) holds mechanically if the conditional expectations on both the
left- and right-hand sides are computed using our estimates of μt and t , re-
gardless of whether these estimates are close to the true objects. Thus, because
we do not observe the true conditional means and covariances, we cannot test
our model using equation (2) directly.
The analysis of Hansen and Richard (1987) helps resolve this problem. They
show that one can evaluate all of the conditional implications of equation (2)
by testing equation (3) using all admissible trading strategies as test assets.
Because our model’s factor represents a return on a traded asset (the UMVE),
the model implies that α p = 0 in the TS regression
∗
R p,t+1 = α p + β p Rt+1 + ε p,t+1 (11)
for each test asset p (e.g., Cochrane (2005), Section 12.1).

Averages of returns and estimated β p ’s do not equal their expected values in
a finite sample. This implies that point estimates of α p are nonzero, even if our
estimates of μt and t equal their true values.
Further, it is not clear a priori that our estimates of μt and t , which are
constructed on an OOS basis, are correct, that is, there is no guarantee that
the resulting UMVE portfolio would have the ex ante MSR. This is another
source of nonzero α p . We therefore validate our estimates of μt , t , and the
resulting UMVE weights by performing standard Gibbons, Ross, and Shanken

(1989) (GRS) joint tests of α p = 0 across a large set of trading strategies p
proposed in the literature.
We further follow Chernov, Lochstoer, and Lundeby (2022) and use multi-
horizon returns (MHR) on the selected assets to generate additional test assets
that are endogenous to the model of the SDF. Specifically, using the projected
SDF associated with the UMVE in equation (4), we test whether
∗
E(Mt,t+h i
Rt,t+h )=1
∗

∗
for a range of horizons h. Here, the multihorizon SDF is Mt,t+h = hj=1 Mt+ j.

h f
The multihorizon gross return is Rt,t+h = j=1 Rt+ j , where Rt+1 = Rt+1 + Rt ,
i i i ei
with the latter denoting the one-month U.S. gross funding rate.
The set of test assets includes the same currency strategies as well as the
UMVE portfolio. Chernov, Lochstoer, and Lundeby (2022) show that uncondi-
tional MHR pricing allows for testing most, if not all, aspects of conditional
model misspecification.
D.2. Trading Strategies

Next, we describe the nine leading trading strategies proposed in the litera-
ture. We use the USD as the base and measurement currency. In the Internet
Appendix, we discuss implications of using the strategies with another base
and/or measurement currency.
The portfolio excess return of a trading strategy is given by

N
R p,t+1 = wipt Rt+1
ei
, (12)
i=1
where wipt is the portfolio’s weight in currency i at time t and N is the number
of currencies. The weight of a given portfolio p can be based on a signal, zipt ,
and chosen such that the portfolio has an exposure to the USD or not. We use
so-called rank and sign weights based on these signals. See Asness, Moskowitz,
and Pedersen (2013), Koijen et al. (2018), and Moskowitz, Ooi, and Pedersen
(2012) for further discussions of such weights.
We use rank weights for CS strategies. The rank of a currency is based on
the signal, and the weight is based on the rank according to

N
−1
w pt = κ rank(z pt ) − N
i i
rank(z pt ) ,
i
(13)
i=1
where the scaling constant κ makes the portfolio one USD long and one USD
short (and hence USD neutral). With nine currencies versus the USD, the pos-
sible weight values are +0.4, +0.3, +0.2, +0.1, 0.0, −0.1, −0.2, −0.3, and −0.4.
Note that the weights depend on the currency ranks, the long and short posi-
tions sum to +1 and −1, respectively, and the net exposure to the USD is zero.
We use sign weights for TS strategies. The weights are then +1 or −1, de-
pending on the sign of the signal, and the net exposure to the USD can be
positive or negative. We further scale these sign weights by N to get a portfolio
volatility similar to those of the CS strategies. However, this scaling does not
affect inference about a strategy’s risk-adjusted performance.
Note that the dollar strategy differs from both CS and TS approaches as it
is an equal-weighted average of the individual currency returns (Lustig, Rous-
sanov, and Verdelhan (2011)).5 The dollar strategy can be seen as an equal-
weighted market portfolio of currencies. It simply goes long all currencies ver-
sus the USD.
We consider three carry strategies. The dollar carry strategy uses the av-
erage forward discount (across all currencies) as a signal. Specifically, it goes
long (short) all currencies versus the USD when the average forward discount
is positive (negative) (Lustig, Roussanov, and Verdelhan (2014)). Hence, the
dollar carry strategy is a conditional version of the dollar strategy above: when
the average forward discount is positive, it goes long the dollar strategy; when
the forward discount is negative, it goes short the dollar strategy.
The CS-carry strategy uses an individual currency’s forward discount as a
signal and the ranking weights as described above. Currencies with relatively
high forward discounts have positive weights, while currencies with relatively
low forward discounts have negative weights (similar to Lustig, Roussanov,
and Verdelhan (2011), who construct a high-minus-low carry portfolio rather
than using the rank weights). Recall that the CS strategies are USD neutral.
The TS-carry uses the sign of the individual currency’s forward discount as a
signal. It goes long (short) currencies with a positive (negative) discount (Burn-
side et al. (2011), Daniel, Hodrick, and Lu (2017)). At each point in time, a vary-
ing number of currencies may have a positive or negative forward discount, so
there is a time-varying exposure to the USD.6
We consider two CS momentum strategies, which use the currency’s perfor-
mance as a signal. The CS-mom 1 strategy uses the performance in the most
recent month as a signal (Burnside et al. (2011), Menkhoff et al. (2012b)), and
the CS-mom 12 strategy uses the performance in the most recent year skipping
5 Verdelhan (2018) uses an equal-weighted average of currency depreciation rates, as opposed
to currency excess returns, when studying variation in exchange rates and refers to it as the dollar
factor. This means that the dollar factor does not include the effect of the forward discount (i.e.,
the interest rate differential under his assumption that the CIP holds), but he controls for this
in his analysis. He also considers excess returns on portfolios based on the exposure to his dollar
factor and refers to the difference between portfolios of high- and low-dollar exposures as the global
component of the dollar factor (or simply “dollar global”). He reports that the global component is
highly correlated with the dollar factor itself (0.85).
6 Hassan and Mano (2019) decompose violations of UIP into three portfolio-based components: a
cross-currency, a between-time-and-currency, and a cross-time component. Empirically, the cross-

currency component drives the correspondence to CS-carry, whereas the cross-time component
drives the correspondence to dollar carry. The between-time-and-currency component, theoreti-
cally shared by both CS-carry and dollar carry, is empirically less important. The decomposition
allows Hassan and Mano (2019) to link dollar carry to the regression-based forward premium
puzzle (see, e.g., Fama (1984), Bansal and Dahlquist (2000), Backus, Foresi, and Telmer (2001)).
the most recent month as a signal (Asness, Moskowitz, and Pedersen (2013)).
Specifically, weights are rank-based as described above.
We also consider two TS momentum strategies, which use the sign of the cur-
rency’s recent performance as a signal. The TS-mom 1 strategy uses the cur-
rency’s prior-month performance (Burnside, Eichenbaum, and Rebelo (2011))
and the TS-mom 12 strategy uses the currency’s performance over the prior
12 months as a signal (Moskowitz, Ooi, and Pedersen (2012)). They both go long
(short) currencies with positive (negative) performance. Similar to Moskowitz,
Ooi, and Pedersen (2012), the latter TS strategy also scales the weights in-
versely with the conditional return volatilities (same estimates as those used
in the construction of the UMVE portfolio).
The CS-value strategy uses the RER signal in equation (8), such that a rel-
atively low (high) RER today indicates that the foreign currency is cheap (ex-
pensive) (Asness, Moskowitz, and Pedersen (2013)). Specifically, weights are
again based on the rank weights as described above.
This set of trading strategies that are based on the USD perspective com-
prises our main set of test results. However, this set may be incomplete. The
first concern that arises relates to the choice of measurement currency, or nu-
meraire. For example, an investor funding herself in JPY, and receiving a pay-
off in JPY, faces a different excess return than a USD-funded investor in the
same currency i. The second concern relates to the choice of base currency. All
of the CS strategies that we discuss are not exposed to the USD by construc-
tion. With respect to the remaining strategies, one could contemplate a differ-
ent base, for example, exposure to the euro (EUR), even if the measurement
currency continues to be USD. Lastly, the Hansen and Richard (1987) result
requires unconditional testing of a model using all admissible strategies, and
we do not have a formal metric to assess how close our choice is to the full set.
We address these concerns in the Internet Appendix.
II. Evidence
A. Data
We construct a data set of daily spot and one-month forward exchange rates
expressed as USD per unit of the foreign currency for the G10 currencies (AUD,
CAD, EUR spliced with DEM, JPY, NZD, NOK, SEK, CHF, GBP, USD) from
January 1, 1976 to May 31, 2020. Daniel, Hodrick, and Lu (2017) offer com-
pelling arguments for exclusion of emerging currencies and the European cur-
rencies other than the EUR. In particular, emerging currencies reflect credit
risk of the respective sovereigns, and thus, the economic composition of the as-
sociated currency risk premiums is different (see, e.g., Na et al. (2018), Cher-
nov, Creal, and Hördahl (2020)).
We use data from several providers through Datastream. Our monthly data
set includes the last day of every month from the daily data set. Forward
exchange rates for the Australian dollar (AUD) and the New Zealand dollar
(NZD) are available from December 1984, and thus January 1985 is the com-
mon starting month for currency excess returns.
We also collect monthly CPIs from the Organisation for Economic Co-
operation and Development (OECD) for the period January 1976 to May 2020.
Only quarterly CPIs are available for the AUD and NZD, so we forward-fill the
quarterly values to get monthly observations. Given that the CPI is published
with a lag, and we want to ensure that all variables are observable at time t,
we construct the RER in equation (7) using CPIs lagged by three months.
The returns on the trading strategies are constructed using data going as
far back as possible. All strategies use all nine currencies versus the USD
from January 1985. Note that since both the value and the momentum signals
rely on spot exchange rates, for which we have data going back to 1976, we
have nine years of data to estimate the first conditional means and covariance
matrix for January 1985, when the currency excess return sample starts. We
next expand the sample each month by one month to update the estimates in
an OOS fashion. This strategy gets us to the target equation (9) in two steps.
First, we forecast percentage changes in spot rates via

i
St+1 /Sti − 1 = δ̄t · ziQt + φ̄t · Sti /St−12
i
− 1 + εt+1
i
.
The coefficients δ̄t and φ̄t are reestimated via a panel regression using histor-
ical data up to month t. Then, using the definition of the return on a forward
position given in equation (6), the time t expected excess currency return is
obtained via
μit = (Sti /Fti ) · Et (St+1

i
/Sti ) − 1

= (Sti /Fti − 1) + (Sti /Fti )δ̄t · ziQt + (Sti /Fti )φ̄t · Sti /St−12
i
−1 ,
with (Sti /Fti )δ̄t and (Sti /Fti )φ̄t corresponding to δti and φti in equation (9),
respectively.
In the case of the alternative conditional forecasts of excess returns, we run
a panel regression of next month’s realized currency returns on the condi-
tional betas as implied by the expected return specification in equation (10).
We use data only up to time t to estimate λtcarry and λtdollar , with the caveat that
since equation (10) calls for forecasting excess returns directly, we cannot use
the depreciation rates from 1976 to 1984 as a “burn-in” sample. Instead, we
use the sample from 1985 to 1994 to estimate constant loadings on the betas
for this period. We then update the loadings in real time, in exactly the same
fashion as in our main model. The factor risk premiums are therefore allowed
to change over time as more data become available.
B. Preliminary Evidence
Before we proceed with testing the model, we check whether the main
ei ei
objects that we use for constructing the UMVE, Et (Rt+1 ) and Vt (Rt+1 ), are
plausible. Specifically, we check whether they predict Rt+1 ei

and (Rt+1 ei
−
ei 2
Et (Rt+1 )) , respectively.
ei ei
Panel A of Table I gives the results from regressing Rt+1 on Et (Rt+1 ) and
(Rt+1 − Et (Rt+1 )) on Vt (Rt+1 ) for individual currencies in a panel setting. Un-
ei ei 2 ei
der the null hypothesis that we have identified the correct conditional means
and variances, the expected slope coefficients equal one. Indeed, slopes in both
regressions are found to be significantly different from zero and insignificantly
different from one. Even pointwise these coefficients are pretty close to one, at
0.84 and 0.93, respectively.
In contrast, when we use the variance-based estimate of the conditional ex-
ei
pectation, labeled Et (Rt+1 ) (var), the point estimate is much lower at 0.573,
ei
and insignificantly different from both zero and one. We also regress Et (Rt+1 )
from our advocated model on the alternative specification in equation (10). The
slope is 0.425 with an adjusted R2 of 0.066. Thus, the two approaches produce
substantively different versions of conditional means.
Panel B of Table I performs the same analysis for strategies. The conditional
expected return and variance of each strategy are computed using the expected
return vector and covariance matrix for the underlying individual currencies,
along with the conditional portfolio weights of each strategy. The conclusion
is the same as for individual returns. Because the strategies involve multiple
returns, the results suggest that the conditional covariances of returns are
estimated well.
To better understand these predictability results, we display the coefficients
δti and φti from our model of expected excess returns in equation (9). The
currency-specific values of these two coefficients are so similar that it is hard
to distinguish them visually. Thus, Figure 1 plots their cross-sectional aver-
ages. For ease of interpretation, we standardize the coefficients in the figure, so
they correspond to the annualized expected return response to a one-standard-
deviation increase in the RER or trend signal. As we use an expanding window
to estimate them, it is expected that they stabilize toward the end of the sam-
ple, which they do around −2.1% and 1.5%, respectively. The signs are stable
and as expected over the sample—a high relative RER implies low future re-
turns, consistent with mean reversion, while a high trend signal implies high
future returns, consistent with momentum. The figure also gives the implied
coefficient on the forward discount, 3.4%, standardized in the same manner as
the other coefficients. The magnitudes of the coefficients imply economically
large variation in conditional expected returns arising from all three signals,
with the forward discount (carry) having the largest impact followed by the
RER and trend signals.
ei
Panel C of Table I documents the magnitude of time variation in Et (Rt+1 )
1/2 ei
and Vt (Rt+1 ) for both individual currencies and strategies. In both cases, the
conditional expectations are as variable as the conditional standard deviation,
indicating substantial time variation in the UMVE portfolio weights and value
from timing currency strategy investments.
Table I
Predictive Ability of Conditional Expectations and Variance
Panel A reports panel regressions of currency excess returns Rt+1 ei on the conditional expected excess returns as estimated from our advocated
ei
t
ei ) (var). We also regress (Rei − E (Rei ))2 (using the advocated
model, Et (Rt+1 ), and from a variance-based alternative estimate of the mean, Et (Rt+1 t+1 t+1
ei ). Panel B presents the same for the strategy
estimate of the conditional expectation) on the conditional variance as estimated by our model, Vt (Rt+1
returns. Standard errors are clustered by month. Panel C reports summary statistics for estimated conditional means and standard deviations.
ei )] denotes the grand (across time and currencies) average risk premium, E[V 1/2 (Rei )] is the grand average conditional standard deviation,
E[Et (Rt+1 t t+1
ei )] is the standard deviation of conditional expected returns, and V 1/2 [V 1/2 (Rei )] is the standard deviation of conditional standard
V 1/2 [Et (Rt+1 t t+1
deviations. As in Panels A and B, “var” refers to the variance-based estimate of the conditional mean. Panel D reports the SR of the UMVE portfolio
as a function of the signals used in its construction, along with the “alpha” of the full-model UMVE portfolio regressed on alternative UMVE portfolios.
“Fwd disc” refers to forward discount. The numbers in Panels C and D are annualized (except for t-statistics and R2adj s), and all UMVE portfolios are
normalized to have the same volatility as that of the dollar strategy.
Panel A. Currency Returns

ei ei
t
Rt+1 Rt+1 ei − E (Rei ))2
(Rt+1 t+1
ei )
Et (Rt+1 0.838 ei ) (var)
Et (Rt+1 0.573 ei )
Vt (Rt+1 0.931
s.e. 0.253 s.e. 0.256 s.e. 0.068
R2adj 0.009 R2adj 0.004 R2adj 0.053
Panel B. Strategy Returns

Pricing Currency Risks
ei ei
t
Rt+1 Rt+1 ei − E (Rei ))2
(Rt+1 t+1
ei )
Et (Rt+1 0.791 ei ) (var)
Et (Rt+1 0.748 ei )
Vt (Rt+1 1.055
s.e. 0.237 s.e. 0.279 s.e. 0.087
R2adj 0.003 R2adj −0.003 R2adj 0.072
(Continued)
17
18
Table 1—Continued
Panel C. Summary Statistics of Conditional Means and Standard Deviations

ei )]
E[Et (Rt+1 ei )] (var)
E[Et (Rt+1 ei )]
E[Vt1/2 (Rt+1 ei )]
V 1/2 [Et (Rt+1 ei )] (var)
V 1/2 [Et (Rt+1 ei )]
V 1/2 [Vt1/2 (Rt+1
Currencies 1.09 4.51 10.71 4.36 2.37 4.77

Strategies 2.55 1.57 7.90 3.39 2.81 3.62
Panel D. UMVE Portfolios
Signals used SR alpha t-Stat R2adj
All signals 1.075 n/a n/a n/a

Fwd disc 0.871 3.59 2.54 0.515
The Journal of Finance®
Fwd disc/Trend 0.871 3.26 2.63 0.586

Fwd disc/RER 0.998 1.47 2.13 0.796
Var-based 0.441 7.51 5.53 0.094
Figure 1. Loadings on signals in expected excess currency returns. The loadings on the
real exchange rate (RER), δti , and on the trend, φti , are estimated in real time using an expanding
window. The figure plots standardized values of cross-sectional averages of δti and φti , as well as
the standardized (unit) coefficient on the forward discount, Sti /Fti − 1. The standardization makes
the coefficient values correspond to the annualized expected return response to a one-standard-
deviation increase in the respective signals. (Color figure can be viewed at wileyonlinelibrary.com)
Panel D explores the role of the three signals that we use to construct
conditional expectations in terms of their effect on the UMVE portfolio.
In the first row, we report the unconditional SR for the full model, which
equals 1.075. We emphasize that UMVE portfolio returns correspond to a
real-time implementable trading strategy. In the subsequent rows, we con-
sider UMVEs formed using the forward discount signal alone, the forward
discount and trend, the forward discount and RER, and the variance-based
estimate of the conditional expectation. All of these combinations result
in lower SRs. The variance-based SR is particularly low at 0.4 and lower
than the carry SRs reported in the literature. This evidence suggests that
conditional expected returns do not line up with conditional exposures to
the PCs of currency returns. We expand on this finding in the subsequent
analysis.
To assess the economic significance of using all of the signals, we implement
the Barillas and Shanken (2017) test and check whether the UMVE formed
using all three signals has an alpha with respect to any of these four alterna-
tive UMVEs. All of the alphas are significant. The corresponding adjusted R2 s
vary between 50% and 80% for signal-based strategies and the R2 is only 9.4%
for the variance-based strategy. We find no significant alphas when we reverse
the direction of the regressions. Thus, the UMVE portfolio from the first row
explains UMVE returns from other versions of our model, but not the other
way around.
Table II
Testing the UMVE
Panel A shows the annualized Sharpe ratio (SR), average excess return, and t-statistic of the
average excess return to each currency, along with its “alpha,” “beta,” and R2 with respect to the
UMVE portfolio. The t-statistics are heteroskedasticity-adjusted. Panel B reports the same for the
strategy returns. The sample is monthly from January 1985 to May 2020.
Panel A. Currency Returns
Currency SR E[Re ] t-Stat α t-Stat β t-Stat R2adj
AUD 0.241 2.83 1.44 0.73 0.34 0.24 2.46 0.025

CAD 0.099 0.74 0.59 −0.03 −0.02 0.09 1.47 0.007
CHF 0.149 1.68 0.89 3.22 1.56 −0.04 −0.43 0.014
EUR 0.145 1.52 0.86 1.83 0.95 −0.21 −2.53 −0.002
GBP 0.217 2.17 1.29 1.08 0.65 0.18 1.95 0.008
JPY 0.054 0.60 0.32 2.45 1.20 0.12 1.38 0.022
NOK 0.205 2.26 1.22 1.27 0.63 0.14 1.68 0.005
NZD 0.427 5.29 2.54 3.72 1.67 −0.18 −2.06 0.012
SEK 0.134 1.48 0.80 0.28 0.14 0.13 1.57 0.008
Panel B. Strategy Returns
Strategy SR E[Re ] t-Stat α t-Stat β t-Stat R2adj
Dollar 0.258 2.06 1.54 1.61 1.14 0.05 0.84 0.000

Dollar Carry 0.576 4.56 3.43 2.40 1.75 0.25 4.28 0.062
CS-Carry 0.469 4.14 2.79 0.43 0.28 0.43 5.85 0.151
TS-Carry 0.685 3.44 4.08 1.25 1.47 0.26 8.47 0.163
CS-Mom 1 0.179 1.46 1.06 0.61 0.36 0.10 1.19 0.007
CS-Mom 12 0.175 1.47 1.04 −1.22 −0.82 0.31 4.88 0.087
TS-Mom 1 0.440 2.73 2.62 1.45 1.30 0.41 4.79 0.034
TS-Mom 12 0.473 5.19 2.82 1.65 0.86 0.15 2.65 0.088
CS-Value 0.529 4.02 3.15 0.98 0.69 0.35 5.71 0.136
C. Testing the Model’s Implications

Table II presents our initial testing results. The Internet Appendix summa-
rizes the split-sample analysis. Panel A reports summary statistics for indi-
vidual currencies (average excess returns and sample SRs). The SR for the
UMVE reported in Table I, Panel D, exceeds the largest currency SR, NZD,
by 2.5 times. Panel A also shows alphas and betas from individual regressions
(11). The largest t-statistic for an alpha is 1.67 for NZD. The largest adjusted
R2 from regressing a return on the UMVE is 2.5% for AUD.
Panel B shows the same information for strategies. The UMVE SR exceeds
the largest strategy SR, TS-carry, by a factor of 1.6. None of the strategies
has a significant alpha relative to the UMVE, with the highest individual t-
statistic of 1.75 for the dollar carry strategy. On average, the alphas are about
70% smaller than the mean excess return on the strategies. The adjusted R2 s
in these regressions are low, which implies that only a small component of the
variation in these portfolios’ returns is being priced. The carry strategies and
Table III
Asset Pricing Tests
We contrast our main model of the UMVE, labeled “optimal,” with alternative methods of UMVE
construction. The row labeled “fwd disc” refers to the case in which the conditional expectations
of excess returns are constructed using the forward discount only. The row labeled “var-based”
corresponds to case in which the conditional expectation is estimated using the variance-based
approach. The GRS p-value is for the standard joint test of all “alpha” equal to zero. The MHR
p-value is for the conditional test of the model as implied by MHRs. The horizons are 1, 3, 6, 12,
24, and 48 months. See the text for details. The “OOS” p-value is for the joint out-of-sample test,
which evaluates whether all average dynamically hedged strategy returns are equal to zero. To
implement the OOS test, we compute the conditional beta on the UMVE portfolio OOS and thus
hedge out the model-implied priced component in real time. This test is also an unconditional test
of each model’s conditional implications. The columns under “All strategies” correspond to tests
using all nine currency trading strategies as tests assets, whereas the columns under “Dollar +
all carry” only uses the dollar, dollar carry, CS-carry, and TS-carry. The sample is monthly from
January 1985 to May 2020.
All Strategies Dollar + All Carry
Model GRS MHR OOS GRS MHR OOS
Optimal 0.470 0.277 0.399 0.467 0.568 0.384

Fwd disc 0.010 0.001 0.006 0.309 0.030 0.133
Var-based 0.000 0.040 0.025 0.010 0.287 0.199
CS-value exhibit the largest exposures to the UMVE with an R2 around 15%,
followed by momentum around 9%, and the rest with an R2 ranging between
0% and 6%. Notably, the dollar strategy returns do not reflect any priced risk,
consistent with the argument and evidence in Boudoukh et al. (2018).
Table III presents asset pricing tests of the alternative UMVE portfolios. The
first row displays the GRS and MHR tests applied to the nine strategies for our
advocated UMVE (labeled “optimal”) under the columns “all strategies.” The
MHR test uses strategy returns at the 1-, 3-, 6-, 12-, 24-, and 48-month hori-
zons, and also adds the UMVE portfolio itself as one of the test assets. The
p-values are 0.47 and 0.28, respectively. These results imply that the candi-
date UMVE prices single-horizon returns to the strategies both uncondition-
ally and conditionally.
This failure to reject is meaningful. Consider a simpler model in which con-
ditional expectations are computed on the basis of the forward discount alone
and test it using the first four strategy portfolios, which do not rely on mo-
mentum or value signals (see the second row of Table III, under the columns
“dollar + all carry”). GRS fails to reject with a p-value of 0.31, but MHR does
reject with a p-value of 0.03. In other words, the UMVE based only on the con-
ditional mean returns implied by the forward discount does not price long-run
returns to the dollar and carry-based strategies. We further reject the ver-
sion of the UMVE based on the variance-based specification of the conditional
mean with both tests when the full set of test assets is used. When we consider
the limited “dollar + all carry” test set, GRS rejects also, while MHR fails to
reject.
Since our advocated UMVE portfolio construction is the only method that is
not rejected in any of the tests, we conclude that it yields a good proxy for the
true UMVE portfolio. One might worry that the failure to reject is because of
low power. However, we reject a number of alternative models, which suggests
that this is not an issue. Moreover, in the Internet Appendix, we provide a
formal evaluation of the power of the GRS test in our specific case and conclude
that the test has good power properties.
D. Time-Series Implications
We analyze the factor structure implied by our model by performing PC anal-
ysis on the covariance matrix of currency returns. In this subsection, we con-
sider the unconditional covariance matrix of individual currencies. In the next
subsection, we consider conditional strategies.
Panel A of Figure 2 displays the contributions of each of the nine PCs to the
overall variation. The first PC explains 58%, the second 18%, and the third 9%
of the variation.
Panel B shows the loadings of the first three PCs on the individual currency
returns. We see that the first PC loads similarly on each currency, suggesting
that it is related to the dollar strategy. Indeed, the correlation between this first
PC and the dollar factor is 0.998. The second PC goes long the high-interest
currencies and short the low-interest currency and is thus related to carry—
their correlation is 0.718. The third PC is harder to associate with an extant
strategy, though a possible interpretation is geographical with largely positive
loadings for European countries and largely negative loadings for the other
countries. These results are consistent with Verdelhan (2018), who argues that
dollar and carry are the main factors driving currency returns.
Panel C examines how much of the UMVE portfolio’s variation can be ex-
plained by the PCs. The Internet Appendix presents split-sample analysis.
Rather than reporting the individual contribution of each PC as in Panel A,
we now report their cumulative contribution. Starting with the first PC and
gradually adding one after another, we also compute how much of the variance
of UMVE returns can be explained by a given number of PCs. The answer is
not much. It is close to zero if we use the first PC alone. Recall that this PC
explains almost 60% of the variation in individual currency returns. If we use
all nine PCs, we progress to about 10% of the variation. This is consistent with
low R2 reported in Panel B of Table II. The flipside of this result is that UMVE
has a significant alpha with respect to the PCs as can be seen in Panel D of
Figure 2.
We conclude that time-series variation in currency returns can be summa-
rized by three factors, two of which are close to the dollar and CS-carry strate-
gies. However, these factors are weakly related to the UMVE, which prices the
cross-section of both currencies and strategies. This suggests that timing based
on the conditional dynamics of currency returns is an important ingredient of
the pricing success.
(A) (B)
(C) (D)
Figure 2. Principal component analysis of currencies. Panel A shows the fraction of vari-
ance across the nine individual currencies that each principal component (PC) explains. The PCs
are obtained from the unconditional covariance matrix of individual currency returns. Panel B
shows the loadings of the three first PCs on each currency. Panel C shows in blue the cumulative
amount of currency variance explained as one goes from using one to all nine PCs. In red is the R2
of a regression of the UMVE returns on an increasing number of PCs. Panel D shows the annual-
ized “alpha” of a regression of the UMVE returns on an increasing number of PCs. The error bars
correspond to the 95% confidence interval computed using heteroskedasticity-adjusted standard
errors. The sample is monthly from January 1985 to May 2020. (Color figure can be viewed at
wileyonlinelibrary.com)
E. Factor Timing
The currency strategies involve a substantive amount of currency timing.
Thus, a natural starting point is to investigate how well these trading strate-
gies can explain UMVE portfolio returns. Figure 3 addresses this issue via
a PC analysis of strategies, which parallels that of individual currencies in
Figure 2.
Panel A shows individual contributions of PCs to the overall variation in
strategy returns. Here, the first three PCs explain 65% of the variation. Panel
(A) (B)
(C) (D)
Figure 3. Principal component analysis of strategies. Panel A shows the fraction of variance
across the nine currency trading strategies that each principal component (PC) explains. The PCs
are obtained from the unconditional covariance matrix of these currency strategies. Panel B shows
the loadings of the three first PCs on each strategy. Panel C shows in blue the cumulative amount
of strategy variance explained as one goes from using one to all nine PCs. In red is the R2 of
a regression of the UMVE returns on an increasing number of these PCs. Panel D shows the
annualized “alpha” of a regression of the UMVE returns on an increasing number of PCs. The
error bars correspond to the 95% confidence interval computed using heteroskedasticity-adjusted
standard errors. The sample is monthly from January 1985 to May 2020. (Color figure can be
viewed at wileyonlinelibrary.com)
B displays the loadings of these three PCs on each strategy, but no clear in-
terpretation emerges regarding what these “factors of strategies” represent.
Panel C shows how much of the variance of UMVE returns can be explained
by strategy PCs. The first three PCs explain only about 20% of the variation
in UMVE returns. Thus, again, the main drivers of the time-series variation
in asset returns are not the main drivers of priced risk, as summarized by the
UMVE. The first five PCs can explain about 50% of the variation, and this
number stays the same after adding the remaining PCs.
The improvement from 10% to 50% is a testament to the importance of cur-

rency timing implicit in the strategies. Yet, a lot left is on the table. UMVE
alphas with respect to these PCs continue to be significant as documented in
Panel D.
We next consider the timing of the nine currency strategies that is implicit in
the UMVE portfolio. If the nine strategies considered in this paper were con-
ditionally noncollinear, then, because strategy weights are known, one could
reexpress the UMVE as a portfolio of strategies. The weights in such a port-
folio would change every period, so one would still have to establish opti-
mal portfolio weights in terms of currency returns, or alternatively, directly
model conditional expected returns and the covariance matrix of the strategies
themselves.
In practice, some of the strategy returns are closely related to each other.
For instance, the dollar carry is conditionally perfectly correlated with the
dollar. The different carry and momentum strategies are likewise related to
each other. Thus, we cannot construct a portfolio of strategies that tracks the
UMVE perfectly.
Instead, to obtain further intuition about the nature of the UMVE portfo-
lio, we use four strategies and strive to get as close to the UMVE as possible.
The four strategies are: the dollar, the sum of the CS- and TS-carry, the sum
of the four momentum strategies, and CS-value. We then apply equation (1)
to strategy returns. We can implement this equation because our model pro-
duces real-time conditional expectations and variance for each currency and
the portfolio weights for each strategy are known.
Regressing the UMVE-tracking portfolio on our UMVE portfolio gives an
R2 of 84%. Its SR is 0.93 versus 1.07 for the UMVE portfolio. The Barillas
and Shanken (2017) test rejects the null that the UMVE-tracking portfolio
can explain the UMVE with a p-value of 0.2%. Thus, we cannot replicate the
optimal portfolio with extant trading strategies. We report deviations of these
portfolio weights from those of the UMVE in the Internet Appendix.
Despite its failure to fully explain the UMVE, it is still informative to study
the properties of the resulting strategies-based proxy. Figure 4 displays the
12-month moving averages of the weights for the four strategies in this track-
ing portfolio. As a reference point, Asness, Moskowitz, and Pedersen (2013)
famously suggest to consider 50% in momentum and 50% in value. Consistent
with the spirit of that suggestion, the sample averages of the strategy weights
are positive. In contrast to the Asness, Moskowitz, and Pedersen (2013) recom-
mendation, however, the optimal strategy weights have strong time variation.
Indeed, the ratio of the sample standard deviation of the weights divided by
their sample mean is 1.21. We note that time variation in portfolio weights
does not imply higher transaction costs than a constant-weight strategy as all
of the trading strategies are based on one-month forward contracts. Thus, all
strategies, simple and complicated, require reestablishment of a new portfolio
every period.
Panel A shows that the portfolio weight on the Dollar strategy is not related
to the business cycle and in general decreasing over the sample. The portfolio
(A) (B)
(C) (D)
Figure 4. The UMVE tracking with strategies. We deduce portfolio positions in the strate-
gies that allow us to track the UMVE portfolio. We consider the dollar strategy (we discard the
dollar carry because it is conditionally perfectly correlated with the dollar), the sum of CS- and
TS-carry (labeled as carry), the sum of CS- and TS-momentum (both 1 and 12 months; labeled as
momentum), and CS-value. We report the 12-month moving average of the UMVE tracking portfo-
lio weights, which are expressed in decimal form. They have been scaled so that the unconditional
standard deviation of UMVE returns matches the unconditional standard deviation of the dollar
factor. The sample is monthly from January 1985 to May 2020. (Color figure can be viewed at
wileyonlinelibrary.com)
weight on the Carry strategies (Panel B) tends to be high in the period prior to
the financial crisis and low thereafter. This is due to a strong decrease in the
cross-sectional standard deviation of forward discounts (interest differentials)
over the sample. The conditional variance of Carry strongly increases during
the financial crisis and then drops throughout the post-2009 period, which is
why the Carry portfolio weight increases after being close to zero at the end of
the financial crisis. The portfolio weight on momentum (Panel C) is overall de-
creasing after its peak in the early 1990s, with increasing weight in recessions.
The latter is due to the cross-sectional standard deviation of the momentum
signal increasing more than the conditional variance of returns in such peri-
ods. The value strategy (Panel D) tends to have low weight in recessions and
high weight in expansions, though there is substantial variation also not asso-
ciated with the business cycle. There is no trend in the weight on value over
the sample.
Figure 5 evaluates the economic significance of the documented time-varying
weights by comparing the UMVE-tracking portfolio to alternative portfolios of
the same four strategies. In addition to the 50% in momentum and 50% in
value portfolio, we consider equal weighting (25% in each of the four strate-
gies) and a volatility timing strategy that uses the inverse of the conditional
variance of each strategy (obtained from our conditional covariance matrix es-
timate) as the portfolio weights. We display TS of the difference between the
conditional SR of the UMVE-tracking portfolio and one of these three alter-
natives. The differences in annualized SR are large, ranging between 0.32 for
equal-weighted and 0.42 for momentum-value strategies, on average.
According to the Barillas and Shanken (2017) test reported in Panel D, the
fixed-weight strategies cannot explain the UMVE-tracking portfolio (i.e., the
latter has significant “alpha” with respect to the former). The latter can explain
the former despite being a suboptimal proxy for the UMVE. In sum, there is
substantial time variation in the conditional mean and variance to the strat-
egy returns that leads to significant gains for a mean-variance investor from
timing these portfolios.
F. Unpriced Components of Returns

The low R2 s of strategy returns in Table II suggest large unpriced compo-
nents in these returns. Daniel et al. (2020), in the context of equity strategies,
make a strong case for hedging out unpriced components to enhance strategy
performance. It is easy to construct a conditional hedge in our setting as we
have an explicit UMVE portfolio.
In equation (2), we use the conditional covariance matrix of the underlying
currencies to construct, in real time, the conditional beta of each strategy on
the UMVE,
wpt t wt∗
β pt = ,
wt∗ t wt∗
where w pt is a vector of strategy weights as in equation (12). A portfolio with

∗
systematic exposure, that is, the hedged portfolio, is simply β pt Rt+1 . We refer
to the residual as the hedging portfolio return,
∗
Rh,t+1 = R p,t+1 − β pt Rt+1 ,
(A) (B)
(C) (D)
Figure 5. Portfolios of strategies versus the UMVE-tracking portfolio. Panels A, B, and C

display the difference between the conditional Sharpe ratios (SRs) of the UMVE tracking portfolio
from Figure 4 and those of the following alternatives constructed from the same four strategies:
25% in each; weights inversely related to the conditional variance of each strategy (obtained from
our conditional covariance matrix for individual depreciation rates); and 50% in CS-mom-12 and
50% in CS-value. Panel D displays p-values for the Barillas-Shanken (2107) test of whether the
UMVE tracking portfolio can be explained by an alternative portfolio (labeled UMVE proxy on
LHS) or whether a fixed-weight portfolio can be explained by the UMVE tracking portfolio (labeled
UMVE proxy on RHS). The sample is monthly from January 1985 to May 2020. (Color figure can
be viewed at wileyonlinelibrary.com)
where the model implies that E(Rh,t+1 ) = 0. Since our time t estimates of both
the UMVE portfolio weights and the conditional covariance matrix only use
data available at time t, this hedging strategy is implementable in real time.
Thus, testing whether the sample average of Rh,t+1 equals zero amounts to an
OOS test of our model.
Figure 6 compares the SRs of the original strategies to those of the hedging
and hedged components. The SRs for the hedging components are close to zero
Figure 6. Sharpe ratios of original, hedging, and hedged strategies. The blue bars show
the sample annualized Sharpe ratios (SRs) of each strategy. The red bars show the SR for each
strategy’s hedging portfolio, defined as the part of the strategy returns that is unpriced according
to the model. The yellow bars show the SR of the factor returns when unpriced risks are hedged
out. Hedging uses real-time conditional betas, so all portfolios are tradeable in real time. The
sample is monthly from January 1985 to May 2020. (Color figure can be viewed at wileyonlineli-
brary.com)
as should be the case for the unpriced component if our model is well specified.
The SRs for the hedged components are much larger than those for the original
strategies as anticipated. For instance, for CS-carry, the SR goes from about
0.5 to slightly higher than one. The one exception is the dollar strategy, where
hedging does not seem to make much of a difference.
Table III reports GRS tests applied to the real-time hedging return of the
strategies in columns labeled “OOS” for out-of-sample. Specifically, we jointly
test whether the average hedged returns to the strategies are zero, as implied
by the model. The use of the conditional betas means that this is an uncondi-
tional test of the models’ conditional implications. The approach is similar in
spirit to that undertaken by Lewellen and Nagel (2006). Overall, this OOS test
yields results similar to the standard GRS test in the same table: only our full
model is not rejected when tested on all strategy returns.
The Internet Appendix discusses an affine model that interprets the ev-
idence in the tradition of Lustig, Roussanov, and Verdelhan (2011), among
many others.
G. Properties of the Projected SDF

The estimated UMVE corresponds to the linear projection of the SDF on the
set of currency excess returns. Figure 7 plots TS of both the implied condi-
Figure 7. Properties of the UMVE. The top panel shows the annualized conditional maximum
Sharpe ratio (MSR) as implied by the UMVE estimated in the paper. The blue solid line corre-
sponds to the full model, whereas the dashed red line shows the case in which only the forward
discount is used for estimating currency risk premiums. In addition, the green line depicts an ex
post MSR, computed over the past 36 months, that serves as a quantitative benchmark. The mid-
dle panel shows the realized returns to the UMVE portfolio implied from this SDF. The color code
is the same as for the top panel. Returns have been scaled to have the same standard deviation as
that of the dollar strategy returns. The blue solid line in the bottom panel is the same conditional
UMVE-based MSR as in the top panel. The dashed red line shows the conditional variance of quar-
terly real per capita nondurable+services log consumption growth, as estimated by a GARCH(1,1).
Both series are standardized to mean zero and unit variance to facilitate comparison. (Color figure
can be viewed at wileyonlinelibrary.com)
tional MSR and returns to the UMVE portfolio, which are perfectly negatively
correlated with this SDF. The MSR varies strongly over the sample, from a
high of about 4 to a low of about 0.5.
We see that the MSR experiences its largest value in a dramatic move dur-
ing the 1992 Exchange Rate Mechanism crisis. Apart from this event, the main
take away is a substantial decline over the sample, mainly caused by a trend
decrease in interest rate differentials across countries. In particular, the same
decline is shared by the MSR of the model that only uses the RWH when esti-
mating currency expected returns (red dashed line). Notably, recessions, indi-
cated by yellow bars in the figure, are associated with a decline in the MSR.
To get a sense of whether these magnitudes are reasonable, we also display
an ex post realized MSR. Specifically, at each date t, we use the past 36 months
of returns to compute the sample SR for each strategy. We then select and plot
the largest SR. These SRs range between 0.15 and 2.8 and feature a similar
declining pattern throughout the sample. These rough computations suggest
plausibility of our estimated price of risk.
The middle panel of Figure 7 shows the time series of returns to the UMVE
portfolio (the dashed red line displays the case of using the forward discount
only). As can be seen, the returns are relatively smooth, with little evidence of
a disaster event. For instance, the Global Financial Crisis is associated with
small negative returns and one big positive spike. The sample skewness and
excess kurtosis are −0.05 and 3.07, respectively, which suggest modest pricing
of disaster risks in currency returns.7 Moreover, we can see a slow decrease in
the volatility of UMVE returns over the sample, consistent with the pattern in
the MSR.
As a next step, we attempt to relate these two objects to observable variables
that have the potential to capture currency risk premiums. Following the liter-
ature, we focus on macro and financial variables. Among macro variables, we
have a particular interest in consumption, as it has a rich history and connec-
tion to equilibrium models. Among financial variables, we consider the Fama
and French (2015) five-factor model as the de-facto SDF projection on the space
of excess equity returns. Finally, we consider measures of intermediary lever-
age. The full list of candidate covariates and the motivations for considering
them are offered in the Internet Appendix.
Table IV reports the results. Panel A displays univariate regressions for the
significant covariates. We see that quarterly consumption growth and three-
year consumption growth, a proxy for the long-run consumption component,
are significant. This result is consistent with Lustig and Verdelhan (2007) and
Zviadadze (2017), who argue that these variables are important for explaining
the CS-carry premium. Our results generalize theirs because we document a
7 We emphasize that this conclusion is not a mechanical implication of how we construct the
UMVE, that is, portfolio weights are constructed using only the mean and variance of excess re-
turns. Our portfolio is represented by a linear combination of excess returns. If these returns
exhibit systematic exposure to crash risk, that will show up in the UMVE. Ex ante, or by construc-
tion, nothing prevents this from happening. If crash risk is important for risk pricing, assets that
load on crash risk will have high SR and high weights in the UMVE.
Table IV
Covariates
Panel A reports results from regressing the UMVE portfolio returns on various factors in uni-
variate regressions. Log real per-capita nondurable+services consumption growth is denoted by
c. HML is the high-minus-low book-to-market factor from the Fama-French five-factor model,
whereas CMA is the Conservative-minus-Aggressive investment factor from the same. Panel B
shows regressions of the CS-carry strategy, the priced component of CS-carry and the unpriced
component of CS-carry returns on the intermediary capital factor of He, Kelly, and Manela (2017),
HKM, as well as shocks to realized volatility in the equity and FX markets. The priced and un-
priced components are obtained in an OOS fashion, as described in the text. Panel C displays a
regression of the three-year change in the MSR on the three-year change in conditional consump-
tion variance (obtained from a GARCH(1,1)), equity variance, and FX variance. We use three-year
changes instead of levels due to a downward drift in the price of risk over the sample period. The
t-statistics are heteroskedasticity-adjusted. The sample is monthly from January 1985 to May
2020.
Panel A. UMVE Returns vs. Observable Factors
Quarterly c 3-Year c HML CMA
Beta 32.452 6.740 0.011 0.021

t-stat 2.516 3.167 2.039 2.724
R2adj 0.031 0.073 0.006 0.013
N 141 131 425 425
Panel B. UMVE, CS-Carry, Its Unpriced Component and Covariates
CS-Carry CS-Carry Priced CS-Carry Unpriced
HKM Equity RV FX RV HKM Equity RV FX RV HKM Equity RV FX RV
Beta 0.111 −2.066 −6.023 0.012 −0.252 −1.350 0.099 −1.814 −7.122
(t-stat) 3.482 −6.743 −6.628 1.236 −1.470 −1.924 3.572 −4.954 −4.980
R2adj 0.084 0.117 0.151 0.003 0.006 0.015 0.086 0.116 0.124
N 425 425 425 425 425 425 425 425 425
Panel C. MSR versus aggregate variance
Cons. Variance Equity RV FX RV
Beta −22508.189 −1.478 −12.372

t-stat −3.928 −0.717 −2.806
R2adj 0.140 0.000 0.032
N 129 399 399
significant relation to the UMVE, and not just carry returns. This generaliza-
tion also explains why we find strong statistical significance. Focusing on the
UMVE portfolio removes the unpriced components of returns and thus helps
recover a more precise relationship with consumption growth as compared to
a regression of strategy returns, like CS-carry, on consumption. The R2 s of the
regressions are only 3% and 7%, indicating that factors other than consump-
tion risk may be important for understanding the currency market conditional
risk-return trade-off.
Panel A also reports the results for the only Fama-French factors that have
a significant relation to the UMVE portfolio, namely, value (HML) and invest-
ment (CMA). Even though they are significant, the relation to the UMVE re-
turns is weak as the betas are small. Further, the R2 s are low, at 0.6% and
1.3%, respectively.
Panel B focuses on the properties of the CS-carry, a strategy that has re-
ceived a lot of attention both in the empirical and the equilibrium literatures.
The left columns of the panel, labeled “CS-carry,” relate returns to this strat-
egy to intermediary leverage (He, Kelly, and Manela (2017)), realized vari-
ance of equity (Lustig, Roussanov, and Verdelhan (2011)), and exchange rates
(Menkhoff et al. (2012a)). We construct the shocks to variance in the same man-
ner as the authors of the two latter papers while we get the intermediary factor
from the former authors’ webpage. Consistent with the literature, the returns
have a strongly statistically significant correlation with these three variables.
However, it is the unpriced component of CS-carry returns that is related
to these variables, not the priced component. The middle columns of Panel B
give the results from regressing the real-time hedged component of the CS-
carry return on these variables. Recall from Section 3.6 that this is the OOS-
priced return component. The slope coefficients are now insignificant. In the
right columns, we show instead the corresponding results using the hedging
component (the OOS-unpriced component) of CS-carry. In this case, the slope
coefficients are strongly significant and similar to those using the raw CS-
carry returns. Recall from Figure 6 that the SR of the unpriced component
is effectively zero, while the SR of the priced component is greater than one.
Thus, these variables cannot explain the average return to the CS-carry strat-
egy. This is not entirely surprising given the results in Table III, which shows
that only about 15% of the variation in CS-carry returns is priced. The differ-
ent conclusion relative to prior literature can also be understood as a case of
omitted factors (Giglio and Xiu (2021)), or more generally, misspecified factors
(Jagannathan and Wang (1998)).
Panel C relates the MSR to various measures of aggregate uncertainty: con-
ditional consumption variance, realized equity variance, and realized foreign
exchange (FX) variance. The conditional variance of consumption growth is es-
timated using a GARCH(1,1) model. Due to the downward trend in the price of
risk in the sample, we run these regressions in three-year differences. The con-
ditional MSR, that is, the conditional volatility of the SDF, is negatively related
to consumption variance and FX variance. Thus, periods of high uncertainty
are associated with a lower SR in the currency market. The sign is also neg-
ative for equity variance, although it is insignificant in this case. The bottom
panel of Figure 7 plots the MSR versus the conditional variance of consump-
tion growth. The common trend and the negatively correlated cyclical relation
are both clearly visible.
We investigate further the drivers behind the time variation in the compo-
nent of the MSR that is correlated negatively with consumption variance. We
go through a series of exercises where we relate the projection of the MSR
onto consumption variance to one of the three variables: (i) cross-sectional
dispersion in conditional expectations, (ii) the level (average) of conditional

expectations, and (iii) the conditional currency return variance, which is calcu-
lated as the sum of the eigenvalues of the conditional currency return covari-
ance matrix. Item (i) implies higher conditional expected returns to judiciously
chosen long-short strategies if the spread between expected returns widens
during times of low consumption variance. In its turn, that would translate
into a higher MSR. Item (ii) leads to a higher MSR if the cross-sectional av-
erage level of returns matters. Lastly, if currency returns become less volatile
when consumption variance is low as per item (iii), then, all else equal, the
MSR increases. We find that only the first variable is significantly related to
the consumption variance in three-year changes (we difference to avoid spuri-
ous regression bias due to highly persistent series), with a t-statistic of 3.05.
III. Conclusion
We construct an estimate of the UMVE portfolio of currencies. This is fea-
sible as most of the literature focuses on the G10 currencies and the corre-
sponding nine exchange rates versus the USD. A small cross-section makes
estimation of the conditional covariance matrix relatively precise. We then use
standard expanding panel regressions and a limited set of longstanding cur-
rency drivers to obtain conditional expected returns. These inputs are all that
is needed to construct the real-time portfolio weights.
The benefit of this approach is that it is a theoretically motivated empir-
ical characterization of the risk-return trade-off. The use of the conditional
covariance matrix means we also characterize the factor structure in realized
returns, which allow us to discuss priced versus unpriced sources of common
variation in currency returns. Our analysis leads to a number of new insights.
First, the UMVE portfolio has a sample SR in excess of one and is not ex-
plained by any existing factors. At the same time, the UMVE portfolio accounts
for known currency factors risk premiums and is not rejected as the pricing
factor in standard model tests. Thus, our proposed estimation methodology is
validated in the data and uncovers heretofore undiscovered sources of risk.
Second, we show that most of the common variation in currency returns is
due to unpriced risks, that is, factors that do not command a risk premium. We
show that existing currency strategies do indeed have large uncompensated
components that can be hedged away to obtain much higher SRs.
Third, we document large gains to timing many of the existing factors, owing
to the fact that the conditional dynamics of these factors are strongly time-
varying. In fact, the standard deviation of conditional risk premiums is as large
as the level of unconditional risk premiums.
Fourth, the priced currency risks are effectively unrelated to priced risks
in the equity market, shocks to equity and FX variances, and intermediary
capital. Consumption growth, however, is related to priced risks, though eco-
nomically the relation is not very strong. That said, our results support a role
for consumption risk in understanding currency risk and returns.
Finally, the conditional MSR of currency returns is strongly downward

trending over the sample, with a cyclical component that is negatively related
with measures of the conditional variance of macroeconomic aggregates and
market returns.
Initial submission: September 13, 2021; Accepted: July 20, 2022

Editors: Stefan Nagel, Philip Bond, Amit Seru, and Wei Xiong
REFERENCES
Ackermann, Fabian, Walt Pohl, and Karl Schmedders, 2017, Optimal and naive diversification in
currency markets, Management Science 63, 3147–3529.
Aloosh, Arash, and Geert Bekaert, 2022, Currency factors, Management Science 68, 3975–4753.
Asness, Clifford S., Tobias J. Moskowitz, and Lasse Heje Pedersen, 2013, Value and momentum
everywhere, Journal of Finance 68, 929–985.
Backus, David K., Silverio Foresi, and Chris I. Telmer, 2001, Affine term structure models and the
forward premium anomaly, Journal of Finance 56, 279–304.
Bansal, Ravi, and Magnus Dahlquist, 2000, The forward premium puzzle: Different tales from
developed and emerging economies, Journal of International Economics 51, 115–144.
Bansal, Ravi, and Ivan Shaliastovich, 2013, A long-run risks explanation of predictability puzzles
in bond and currency markets, Review of Financial Studies 26, 1–33.
Barillas, Fancisco, and Jay Shanken, 2017, Which alpha?, Review of Financial Studies 30, 1316–
1338.
Baz, Jamil, Francis Breedon, Vasant Naik, and Joel Peress, 2001, Optimal portfolios of foreign
currencies, Journal of Portfolio Management 28, 102–111.
Bekaert, Geert, and Jun Liu, 2004, Conditioning information and variance bounds on pricing ker-
nels, Review of Financial Studies 17, 339–378.
Bilson, John, 1984, Purchasing power parity as a trading strategy, Journal of Finance 39, 715–724.
Black, Fischer, Michael Jensen, and Myron Scholes, 1972, The capital asset pricing model: Some
empirical tests, in Michael Jensen, ed.: Studies in the Theory of Capital Markets (Praeger
Publishers Inc., New York).
Boudoukh, Jacob, Matthew Richardson, Ashwin Thapar, and Franklin Wang, 2018, Is the dollar a
global risk factor?, Working paper, NYU Stern.
Burnside, Craig, 2012, Carry trades and risk, in Jessica James, Ian Marsh, and Lucio Sarno, eds.:
Handbook of Exchange Rates (John Wiley and Sons, Ltd, Hoboken, NJ).
Burnside, Craig, Martin Eichenbaum, Isaac Kleshchelski, and Sergio Rebelo, 2011, Do peso prob-
lems explain the returns to the carry trade?, Review of Financial Studies 24, 853–891.
Burnside, Craig, Martin Eichenbaum, and Sergio Rebelo, 2011, Carry trade and momentum in
currency markets, Annual Review of Financial Economics 3, 511–535.
Cassel, Gustav, 1918, Abnormal deviations in international exchanges, Economic Journal 28, 413–
415.
Chernov, Mikhail, and Drew Creal, 2021, The PPP view of multihorizon currency risk premiums,
Review of Financial Studies 34, 2728–2772.
Chernov, Mikhail, Drew Creal, and Peter Hördahl, 2020, Sovereign credit and exchange rate risks:
Evidence from Asia-Pacific local currency bonds, Journal of International Economics, forth-
coming.
Chernov, Mikhail, Lars Lochstoer, and Stig Lundeby, 2022, Conditional dynamics and the multi-
horizon risk-return trade-off, Review of Financial Studies 3, 1310–1347.
Cochrane, John, 2005, Asset Pricing, revised edition (Princeton University Press, Princeton, NJ).
Colacito, Ric, Mariano M. Croce, Federico Gavazzoni, and Robert Ready, 2018, Currency risk fac-
tors in a recursive multicountry economy, Journal of Finance 73, 2719–2756.
Cornell, W.B., and J. K. Dietrich, 1978, The efficiency of the market for foreign exchange under
floating exchange rates, Review of Economics and Statistics 60, 111–120.
Dahlquist, Magnus, and Julien Pénasse, 2022, The missing risk premium in exchange rates,
Journal of Financial Economics 143, 697–715.
Daniel, Kent, Robert J. Hodrick, and Zhongjin Lu, 2017, The carry trade: Risks and drawdowns,
Critical Finance Review 6, 211–262.
Daniel, Kent, Lira Mota, Simon Rottke, and Tano Santos, 2020, The cross-section of risk and
return, Review of Financial Studies 33, 1927–1979.
Pasquale Della, Corte,, Lucio Sarno, and Ilias Tsiakas, 2009, An economic evaluation of empirical
exchange rate models, Review of Financial Studies 22, 3491–3530.
Eichenbaum, Martin, Benjamin K. Johannsen, and Sergio Rebelo, 2021, Monetary policy and the
predictability of nominal exchange rates, Review of Economic Studies 88, 192–228.
Fama, Eugene, 1984, Forward and spot exchange rates, Journal of Monetary Economics 14, 319–
338.
Fama, Eugene, and Kenneth French, 1993, Common risk factors in the returns on stocks and
bonds, Journal of Financial Economics 33, 3–56.
Fama, Eugene, and Kenneth French, 2015, A five-factor asset pricing model, Journal of Financial
Economics 116, 1–22.
Farhi, Emmanuel, and Xavier Gabaix, 2016, Rare disasters and exchange rates, Quarterly Journal
of Economics 131, 1–52.
Ferson, Wayne, and Andrew Siegel, 2001, The efficient use of conditioning information in portfo-
lios, Journal of Finance 56, 967–982.
Ferson, Wayne, and Andrew Siegel, 2003, Stochastic discount factor bounds with conditioning
information, Review of Financial Studies 16, 567–595.
Ferson, Wayne, and Andrew Siegel, 2009, Testing portfolio efficiency with conditioning informa-
tion, Review of Financial Studies 22, 2735–2758.
Ronald Gallant, A., Lars P. Hansen, and George Tauchen, 1990, Using conditional moments of
asset payoffs to infer the volatility of intertemporal marginal rates of substitution, Journal of
Econometrics 45, 141–179.
Ghosh, Anisha, Christian Julliard, and Alex P. Taylor, 2019, An information-theoretic asset pricing
model, Working paper, LSE.
Gibbons, Michael, Stephen A. Ross, and Jay Shanken, 1989, A test of the efficiency of a given
portfolio, Econometrica 59, 1121–1152.
Giglio, Stefano, and Dacheng Xiu, 2021, Asset pricing with omitted factors, Journal of Political
Economy 129, 1947–1990.
Greenaway-McGrevy, Ryan, Nelson C. Mark, Donggyu Sul, and Juh-Lin Wu, 2018, Identifying
exchange rate common factors, International Economic Review 59, 2193–2218.
Hansen, Lars, and Ravi Jagannathan, 1991, Implications of security market data for models of
dynamic economies, Journal of Political Economy 99, 225–262.
Hansen, Lars, and Scott Richard, 1987, The role of conditioning information in deducing testable
restrictions implied by dynamic asset pricing models, Econometrica 55, 587–613.
Hassan, Tarek K., and Rui C. Mano, 2019, Forward and spot exchange rates in a multi-currency
world, Quarterly Journal of Economics 134, 397–450.
He, Zhiguo, Bryan Kelly, and Asaf Manela, 2017, Intermediary asset pricing: New evidence from
many asset classes, Journal of Financial Economics 126, 1–35.
Jagannathan, Ravi, 1996, Relation between the slopes of the conditional and unconditional mean-
standard deviation frontiers of asset returns, in K. Sawaki S. Saito, and K. Kubota, eds.:
Modern Portfolio Theory and its Applications: Inquires into Asset Valuation Problems (Center
for Academic Societies, Osaka, Japan).
Jagannathan, Ravi, and Zhenyu Wang, 1996, The conditional CAPM and the cross-section of ex-
pected returns, Journal of Finance 51, 3–53.
Jagannathan, Ravi, and Zhenyu Wang, 1998, An asymptotic theory for estimating beta-pricing
models using cross-sectional regression, Journal of Finance 53, 1285–1309.
Jorda, Oscar, and Alan M. Taylor, 2012, The carry trade and fundamentals: Nothing to fear but
FEER itself, Journal of International Economics 88, 74–90.
Keynes, John Maynard, 1923, A Tract on Monetary Reform (MacMillan and Co., Ltd., London).
Kho, Boing-Chan,1996, Time-varying risk premia, volatility, and technical reading profits: Evi-
dence from foreign currency futures markets, Journal of Financial Economics 41, 249–290.
Koijen, Ralph S.J., Tobias J. Moskowitz, Lasse Heje Pedersen, and Evert B. Vrugt, 2018, Carry,
Korsaye, Sofonias Alemu, Fabio Trojani, and Andrea Vedolin, 2020, The global factor structure of
exchange rates, Journal of Financial Economics, forthcoming.
Ledoit, Olivier, and Michael Wolf, 2020, Analytical nonlinear shrinkage of large-dimensional co-
variance matrices, Annals of Statistics 48, 3043–3065.
Lettau, Martin, and Markus Pelger, 2020, Factors that fit the time series and cross-section of stock
returns, Review of Financial Studies 33, 2274–2325.
Lewellen, Jonathan, and Stefan Nagel, 2006, The conditional CAPM does not explain asset pricing
anomalies, Journal of Financial Economics 79, 289–314.
Lo, Andrew, and Craig A. MacKinlay, 1990, Data-snooping biases in tests of financial asset pricing
models, Review of Financial Studies 3, 431–467.
Lustig, Hanno, Nikolai Roussanov, and Adrien Verdelhan, 2011, Common risk factors in currency
markets, Review of Financial Studies 24, 3731–3777.
Lustig, Hanno, Nikolai Roussanov, and Adrien Verdelhan, 2014, Countercyclical currency risk
premia, Journal of Financial Economics 111, 527–553.
Lustig, Hanno, and Adrien Verdelhan, 2007, The cross section of foreign currency risk premia and
consumption growth risk, American Economic Review 97, 89–117.
Maurer, Thomas, Thuy-Duong To, and Ngoc-Khanh Tran, 2020, Market timing and predictability
in FX markets, Review of Finance, forthcoming.
Maurer, Thomas, Thuy-Duong To, and Ngoc-Khanh Tran, 2022, Pricing implications of covari-
ances and spreads in currency markets, Review of Asset Pricing Studies 12, 336–388.
Meese, Richard A., and Kenneth Rogoff, 1983, Empirical exchange rate models of the seventies:
Do they fit out of sample?, Journal of International Economics 14, 3–24.
Menkhoff, Lukas, Lucio Sarno, Maik Schmeling, and Andreas Schrimpf, 2012a, Carry trades and
global foreign exchange volatility, Journal of Finance 67, 681–718.
Menkhoff, Lukas, Lucio Sarno, Maik Schmeling, and Andreas Schrimpf, 2012b, Currency momen-
tum strategies, Journal of Financial Economics 106, 660–684.
Moskowitz, Tobias J., Yao Hua Ooi, and Lasse Heje Pedersen, 2012, Time series momentum,
Na, Seunghoon, Stephanie Schmitt-Grohe, Martin Uribe, and Vivian Yue, 2018, The twin Ds:
Optimal default and devaluation, American Economic Review 108, 1773–1819.
Nucera, Federico, Lucio Sarno, and Gabriele Zinna, 2022, Currency risk premia redux, Working
paper, Cambridge University.
Okunev, John, and Derek White, 2003, Do momentum-based strategies still work in foreign cur-
rency markets?, Journal of Financial and Quantitative Analysis 38, 425–447.
Orlowski, Piotr, Valeri Sokolovski, and Erik Sverdrup, 2021, Benchmark currency stochastic dis-
count factors, Working paper, HEC Montreal.
Porter, Michael, 1971, A theoretical and empirical framework for analyzing the term structure of
exchange rate expectations, Staff Papers (International Monetary Fund) 18, 613–645.
Sandulescu, Mirela, Fabio Trojani, and Andrea Vedolin, 2021, Model-free international stochastic
discount factors, Journal of Finance 76, 935–976.
Stutzer, Michael, 1996, A simple nonparametric approach to derivative security valuation, Journal
of Finance 60, 1633–1652.
Sweeney, Richard, 1986, Beating the foreign exchange market, Journal of Finance 41, 163–182.
Verdelhan, Adrien, 2010, A habit-based explanation of the exchange rate risk premium, Journal
of Finance 65, 123–146.
Verdelhan, Adrien, 2018, The share of systematic variation in bilateral exchange rates, Journal of
Finance 73, 375–418.
Zviadadze, Irina, 2017, Term-structure of consumption risk premia in the cross-section of currency
returns, Journal of Finance 72, 1529–1566.
Supporting Information
Additional Supporting Information may be found in the online version of this
article at the publisher’s website:
Appendix S1: Internet Appendix.
Replication Code.

Jofi 13190

Uploaded by

Copyright:

Available Formats

Jofi 13190

Uploaded by

Document Information

Original Title

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

Jofi 13190

Uploaded by

Copyright:

Available Formats

THE JOURNAL OF FINANCE • VOL. , NO.

Pricing Currency Risks

MIKHAIL CHERNOV, MAGNUS DAHLQUIST, and LARS LOCHSTOER*

IN THIS PAPER, WE ARGUE that research on cross-sectional (CS) currency

* Mikhail Chernov is at UCLA Anderson School. Magnus Dahlquist is at Stockholm School of

admissible dynamic trading strategies in these currencies unconditionally. As

Jagannathan (1991) bounds as a diagnostic in their models.4 A variation on

I. Linear Factor Models and Exchange Rates

A. The UMVE Portfolio

basis assets has an N × 1 vector of weights w pt that are determined only by

The UMVE portfolio is a dynamic trading strategy in these assets and

has the property that

C. Implementing the UMVE for Exchange Rates

where Dt is the number of trading days in the month. Given a potentially

with γ = 1. This is a particular violation of UIP, which posits γ = 0. We refer

Our mean-reversion signal is motivated by the literature on the role of RER

Qti = Sti · Pti /Pt , (7)

In a last step, we cross-sectionally demean the signal at each time t to create a

μit = λtcarry · βitcarry + λtdollar · βitdollar , (10)

where a factor (carry or dollar) is a portfolio of the currencies, RFp,t+1 =

D. Testing Linear Factor Models

for each test asset p (e.g., Cochrane (2005), Section 12.1).

resulting UMVE weights by performing standard Gibbons, Ross, and Shanken

D.2. Trading Strategies

5 Verdelhan (2018) uses an equal-weighted average of currency depreciation rates, as opposed

cross-currency, a between-time-and-currency, and a cross-time component. Empirically, the cross-

μit = (Sti /Fti ) · Et (St+1

plausible. Specifically, we check whether they predict Rt+1 ei

Panel A. Currency Returns

Panel B. Strategy Returns

Panel C. Summary Statistics of Conditional Means and Standard Deviations

Currencies 1.09 4.51 10.71 4.36 2.37 4.77

Panel D. UMVE Portfolios

Signals used SR alpha t-Stat R2adj

All signals 1.075 n/a n/a n/a

Fwd disc/Trend 0.871 3.26 2.63 0.586

Panel A. Currency Returns

Currency SR E[Re ] t-Stat α t-Stat β t-Stat R2adj

AUD 0.241 2.83 1.44 0.73 0.34 0.24 2.46 0.025

Panel B. Strategy Returns

Strategy SR E[Re ] t-Stat α t-Stat β t-Stat R2adj

Dollar 0.258 2.06 1.54 1.61 1.14 0.05 0.84 0.000

C. Testing the Model’s Implications

All Strategies Dollar + All Carry

Model GRS MHR OOS GRS MHR OOS

Optimal 0.470 0.277 0.399 0.467 0.568 0.384

The improvement from 10% to 50% is a testament to the importance of cur-

F. Unpriced Components of Returns

where w pt is a vector of strategy weights as in equation (12). A portfolio with

Figure 5. Portfolios of strategies versus the UMVE-tracking portfolio. Panels A, B, and C

G. Properties of the Projected SDF

Panel A. UMVE Returns vs. Observable Factors

Quarterly c 3-Year c HML CMA

Beta 32.452 6.740 0.011 0.021

Panel B. UMVE, CS-Carry, Its Unpriced Component and Covariates

CS-Carry CS-Carry Priced CS-Carry Unpriced

HKM Equity RV FX RV HKM Equity RV FX RV HKM Equity RV FX RV

Panel C. MSR versus aggregate variance

Cons. Variance Equity RV FX RV