Volatility Risk Premia and Exchange Rate Predictability

y
Pasquale DELLA CORTE Tarun RAMADORAI Lucio SARNO
December 2013
Abstract
We investigate the predictive capability of currency volatility risk premia for currency
returns. The volatility risk premium is the dierence between realized volatility and
(model-free) currency-option-implied volatility, and reects the costs of insuring against
currency volatility uctuations. A portfolio that sells high-insurance-cost currencies and
buys low-insurance-cost currencies generates sizeable out-of-sample returns and Sharpe
ratios. These returns are mainly generated by spot exchange rates rather than interest
rate dierentials, and carry a greater weight in the minimum variance currency strategy
portfolio than both carry and momentum. Canonical risk factors cannot price the returns
from this strategy, which appear more consistent with time-varying limits to arbitrage.
Keywords: Exchange Rates; Volatility Risk Premium; Predictability, Minimum-
Variance Currency Portfolio.
JEL Classication: F31; F37.

Acknowledgements: We are grateful to John Campbell, Kenneth Froot, Federico Gavazzoni, Philippos
Kassimatis, Lars Lochstoer, Stefan Nagel, Andrea Vedolin, Adrien Verdelhan and participants at the Oxford-
Man 2013 Conference on Currency Trading and Risk Premia, the 2013 NBER Summer Institute in Asset
Pricing, the 2013 Annual Conference on Advances in the Analysis of Hedge Fund Strategies at Imperial
College London and various seminar presentations for helpful conversations and suggestions. We also thank
JP Morgan and Aslan Uddin for some of the data used in this study. Sarno acknowledges nancial support
from the Economic and Social Research Council (No. RES-062-23-2340). All errors remain ours.
y
Pasquale Della Corte is at Imperial College Business School, Imperial College London; email:
[Link]@[Link]. Tarun Ramadorai is at the Sad Business School, Oxford-Man Institute,
University of Oxford and CEPR; email: [Link]@[Link]. Lucio Sarno is at Cass Business
School, City University London and CEPR; email: [Link]@[Link].
1 Introduction
What explains currency uctuations? Finance practitioners and academics have been both
fascinated with this question, and struggled in vain with it for decades.
1
Recently, there has
been signicant interest in a closely-related question, which is to better understand the high
returns to popular currency investment strategies. Adopting the cross-sectional asset pricing
approach of constructing portfolios sorted by currency characteristics, such as interest rate
dierentials or lagged returns, researchers have shown that there are large returns to carry
and momentum strategies in currencies.
2
In this paper, we investigate the predictive information content in the currency volatility
risk premium for exchange rate returns. Our key, novel result is that there is economically
valuable and statistically signicant predictive information in the currency volatility risk pre-
mium for future currency excess returns and spot exchange rate returns over the 1996 to 2011
period, in a cross-section of up to 20 currencies.
3
A useful summary statistic of the impor-
tance of our new currency strategy (which we dub \ 11) is that it has the highest weight
(33%) in the global minimum variance portfolio of ve currency strategies that we consider,
including carry and momentum. This is despite the fact that \ 11 does not have the highest
returns among the strategies considered. It does, however, have extremely desirable correla-
tion properties relative to the other well-studied currency strategies, which partly arise from
the excellent performance of \ 11 during crises, and primarily from the fact that the currency
excess returns of \ 11 are almost completely obtained through prediction of spot currency
returns rather than from interest rate dierentials.
4
1
The diculty to explain and forecast nominal exchange rates was rst recorded in the seminal study
of Meese and Rogo (1983), which documents that it is endishly dicult to nd theoretically motivated
variables able to beat a random walk forecasting model for currencies. The literature has found it dicult to
move far ahead of this result in the past three decades (e.g. see Engel, Mark and West, 2008).
2
See, for example, Lustig and Verdelhan (2007), Ang and Chen (2010), Burnside, Eichenbaum, Kleshchel-
ski, and Rebelo (2011), Lustig, Roussanov, and Verdelhan (2011), Barroso and Santa Clara (2013) and
Menkho, Sarno, Schmeling, and Schrimpf (2012a,b), who all build currency portfolios to study return pre-
dictability and/or currency risk exposure.
3
To be clear from the outset, our strategy does not trade volatility products. We simply use the expected
volatility risk premium as conditioning information to sort currencies, build currency portfolios, and uncover
predictability in currency excess returns and spot exchange rate returns.
4
We use interchangeably the terms spot currency returns and exchange rate returns to dene the change
in nominal exchange rates over time; similarly we use interchangeably the terms excess returns or portfolio
returns to refer to the returns from implementing a long-short currency trading strategy that buys and sells
currencies on the basis of some characteristic.
1
The currency volatility risk premium is the dierence between expected future realized
volatility, and a model-free measure of expected volatility derived from currency options. A
growing literature studies the variance or the volatility risk premium in dierent asset classes,
including equity, bond, and foreign exchange (FX) markets.
5
In general, this literature has
shown that the volatility risk premium is on average negative: expected volatility is higher
than historical realized volatility, and since volatility is persistent, expected volatility is also
generally higher than future realized volatility. In other words, the volatility risk premium
represents compensation for providing volatility insurance. Therefore, akin to the interpreta-
tion in Garleanu, Pedersen, and Poteshman (2009), the currency volatility risk premium that
we construct can be interpreted as the cost of insurance against volatility uctuations in the
underlying currency when it is high (realized volatility is higher than the option-implied
volatility), insurance is relatively cheap, and vice versa.
We use the currency volatility risk premium to rank currencies and to build currency
portfolios, sorting currencies into quintile portfolios by this variable at the beginning of each
month. Our trading strategy is to buy currencies with relatively cheap volatility insurance,
i.e., the highest volatility risk premium quintile, and short currencies with relatively expensive
volatility insurance, i.e., the lowest volatility risk premium quintile. We track returns on this
trading strategy (\ 11) over the subsequent period, meaning that these returns are purely out-
of-sample, conditioning only on information available at the time of portfolio construction.
The performance of \ 11 stems virtually entirely from the predictability of spot exchange
rates rather than from interest rate dierentials. That is, currencies with relatively cheap
volatility insurance tend to appreciate over the subsequent month, while those with relatively
more expensive volatility insurance tend to depreciate over the next month. The observed
predictability of spot exchange rates associated with \ 11 is far stronger than that arising from
carry (which generates returns that are almost entirely driven by interest rate dierentials,
and not by any predictive ability for spot rate changes) and currency momentum, as well as
other currency trading strategies that we consider. As mentioned earlier, this is part of the
reason for the excellent diversication benets that the \ 11 strategy oers in a currency
5
See, for example, Carr and Wu (2009), Eraker (2008), Bollerslev, Tauchen, and Zhou (2009), Todorov
(2010), Drechsler and Yaron (2010), Han and Zhou (2010), Mueller, Vedolin and Yen (2011), Londono and
Zhou (2012) and Buraschi, Trojani and Vedolin (2013).
2
portfolio.
There are several possible interpretations of our results, of which we consider two to be
most likely. One possibility is that \ 11 captures uctuations in aversion to volatility risk,
so that currencies with high volatility insurance have low expected returns and vice versa.
Note that our result is cross-sectional, since we are long and short currencies simultaneously.
As a result, if this explanation were true, it would rely either on dierent currencies loading
dierently on a global volatility shock, or indeed on market segmentation causing expected
returns on dierent currencies to be determined independently. We test this explanation
both using cross-sectional asset pricing tests of volatility risk premium-sorted portfolios on
a global FX volatility risk factor, as well as by estimating the loadings of currency returns
on various proxies for global volatility risk and building portfolios sorted on these estimated
loadings. Neither of these tests produces evidence consistent with the proposed explanation,
with the long-short strategy generated from estimated loadings on the global volatility risk
factor producing far inferior returns to \ 11, which are also virtually uncorrelated with \ 11
returns. In sum, the data appear to reject an explanation based on uctuations in aversion
to global volatility risk.
The second explanation that we consider for our results relies on the presence of limits to
arbitrage, and its eects on the interaction between hedgers and speculators in the currency
market. There is a growing theoretical and empirical literature suggesting that such inter-
actions are important in asset return determination (see, for example, Acharya, Lochstoer,
and Ramadorai, 2013; Adrian, Etula, and Muir, 2013; and Gromb and Vayanos, 2010 for an
excellent survey of the literature). In the currency markets, this explanation comprises two
components. First, it requires time-variation in the amount of arbitrage capital available to
natural providers of currency volatility insurance (speculators), such as nancial institu-
tions or hedge funds. Second, risk-averse natural hedgers of currencies such as multina-
tional rms, or nancial institutions that inherit currency positions from their clients, should
be more willing to hedge and be more comfortable with holding (or entering into contracts
denominated in) currencies with relatively inexpensive volatility insurance. Such institutions
will also be more likely to avoid positions in currencies with relatively expensive volatility
protection. The combination of these two ingredients would be sucient to generate the
patterns that we see in the data.
3
A simple example may be helpful: assume that speculators face a shock to their available
arbitrage capital. This limits their ability to provide cheap volatility insurance, especially in
currencies in which they have large positions for example, they may reduce their outstanding
short put option positions in the currencies in which they trade.
6
These limits on speculators
ability to satisfy demand for volatility insurance increases net demand in the options market
for the specic currencies in which they are most active, increasing current option prices and
making hedging more expensive. As in Garleanu, Pedersen, and Poteshman (2009), this net
demand imbalance would show up in a lower volatility risk premium for the currencies thus
aected. Given the high cost of volatility insurance, natural hedgers scale back on the amount
of spot currency they are willing to hold, or are reluctant to get into new expensive hedges.
This net demand will predictably depress spot prices, leading to relatively low returns on the
spot currency position. When capital constraints loosen, we should see the opposite behavior,
i.e., a reversal in both the volatility risk premium and the spot currency position.
In the cross-section of currencies, this mechanism implies that, in a world with limited and
time-varying arbitrage capital, an institution wishing to hedge against risk (or deleveraging) in
one currency position rather than another will generate excess demand for volatility insurance
for the currency to which it is more exposed, in turn increasing the spread in volatility risk
premia across currencies.
This explanation for our baseline result has additional testable implications. Most obvi-
ously, the explanation implies that the returns from the \ 11 strategy, post-formation, should
be temporary, i.e., there should be reversion in currency returns once arbitrage capital returns
to the market. Conrming this prediction, we nd that currency volatility risk-premium
sorted portfolio returns reverse over a holding period of a few months. Moreover, at times
when funding liquidity is lower (i.e., times of high capital constraints on speculators), and
demand for volatility protection is higher (i.e., times of increased risk aversion of natural
hedgers), we should nd that the spread in the cost of volatility insurance across currencies,
and the spread in spot exchange rate returns across portfolios should both increase. In our
empirical analysis, we nd that when the TED spread a commonly used proxy for funding
6
Short put options is a favoured strategy of many hedge funds; see Agarwal and Naik (2004), for example.
Also see Fung and Hsieh (1997) for how lookback options can be used to capture the returns to momentum
trading strategies implemented by hedge funds.
4
liquidity (see, for example, Garleanu and Pedersen, 2011) increases, the returns on \ 11
are substantially higher. Fluctuations in risk aversion, as proxied by changes in the VIX,
are also useful in explaining our returns, and add signicant additional explanatory power
when interacted with the TED spread. We also measure capital ows to currency and global
macro hedge funds, and nd that when hedge fund ows are high, signifying increased funding
and thus lower hedge fund capital constraints, the returns to \ 11 are lower and vice versa,
providing useful evidence in support of the limits to arbitrage explanation. We also obtain
evidence in favour of the limits to arbitrage explanation by investigating the behavior of cur-
rency trading in our insurance-cost sorted portfolios and documenting that \ 11 is connected
to measures of order ow in a way that is consistent with the proposed explanation.
7
Specif-
ically we nd that non-nancial traders tend to buy currencies which are cheaper to insure,
and sell currencies which are more expensive to insure; by contrast, nancial traders appear
to trade in a way that is exactly opposite to that of non-nancial traders. This pattern of
trading behavior serves to corroborate our other evidence suggesting that \ 11 returns are
driven by the interaction of natural hedgers and speculators in currency markets.
The paper is structured as follows. Section 2 denes the volatility risk premium and
its measurement in currency markets. Section 3 describes our data and some descriptive
statistics. Section 4 presents our main empirical results on the volatility risk premium-sorted
strategy, Section 5 reports formal asset pricing tests, while Section 6 investigates two alterna-
tive mechanisms that could explain our ndings. Section 7 concludes. A separate Internet
Appendix provides robustness tests and additional supporting analyses.
2 Foreign Exchange Volatility Risk Premia
Volatility Swap. A volatility swap is a forward contract on the volatility realized on the
underlying asset over the life of the contract. The buyer of a volatility swap written at time
t, and maturing at time t +t, receives the payo (per unit of notional amount):
\ 1
t;
= (1\
t;
o\
t;
) (1)
7
This evidence links our work to another important stream of the exchange rate literature on forecasting
currency returns using currency order ow. For example, Froot and Ramadorai (2005), Evans and Lyons
(2005) and Rime, Sarno and Sojli (2010) show that order ow has substantial predictive power for exchange
rate movements.
5
where 1\
t;
is the realized volatility of the underlying, o\
t;
is the volatility swap rate, and
both 1\
t;
and o\
t;
are dened over the life of the contract from time t to time t + t, and
quoted in annual terms. However, while the realized volatility is determined at the maturity
date t +t, the swap rate is agreed at the start date t.
The value of a volatility swap contract is obtained as the expected present value of the
future payo in a risk-neutral world. This implies, because \ 1
t;
is expected to be 0 under
the risk-neutral measure, that the volatility swap rate equals the risk-neutral expectation of
the realized volatility over the life of the contract:
o\
t;
= 1
Q
t
[1\
t;
] (2)
where 1
Q
t
[] is the expectation under the risk-neutral measure Q, 1\
t;
=
q
t
1
R
t+
t
o
2
s
d:,
and o
2
s
denotes the (stochastic) volatility of the underlying asset.
Volatility Swap Rate. We synthesize the volatility swap rate using the model-free
approach derived by Britten-Jones and Neuberger (2000), and further rened by Demeter,
Derman, Kamal and Zou (1999), Jiang and Tian (2005), and Carr and Wu (2009).
Building on the pioneering work of Breeden and Litzenberger (1978), Britten-Jones and
Neuberger (2000) derive the model-free implied volatility entirely from no-arbitrage conditions
and without using any specic option pricing model. Specically, they show that the risk-
neutral expected integrated return variance between the current date and a future date is fully
specied by the set of prices of call options expiring on the future date, provided that the
price of the underlying evolves continuously with constant or stochastic volatility but without
jumps.
Demeter, Derman, Kamal, and Zou (1999) show that the Britten-Jones and Neuberger
(2000) solution is equivalent to a portfolio that combines a dynamically rebalanced long po-
sition in the underlying, and a static short position in a portfolio of options and a forward
that together replicate the payo of a log contract.
8
The replicating portfolio strategy cap-
tures variance exactly, provided that the portfolio of options contains all strikes with the
appropriate weights to match the log payo. Jiang and Tian (2005) further demonstrate that
8
The log contract is an option whose payo is proportional to the log of the underlying at expiration
(Neuberger, 1994).
6
the model-free implied variance is valid even when the underlying price exhibits jumps, thus
relaxing the diusion assumptions of Britten-Jones and Neuberger (2000).
The risk-neutral expectation of the return variance between two dates t and t + t can be
formally computed by integrating option prices expiring on these dates over an innite range
of strike prices:
1
Q
t

1\
2
t;

= i

Z
F
t;
0
1
1
2
1
t;
(1)d1 +
Z
1
F
t;
1
1
2
C
t;
(1)d1
!
(3)
where 1
t;
(1) and C
t;
(1) are the put and call prices at t with strike price 1 and maturity
date t +t, 1
t;
is the forward price matching the maturity date of the options, o
t
is the price
of the underlying, i = (2,t) exp (i
t;
t), and i
t;
is the t-period domestic riskless rate.
The risk-neutral expectation of the return variance in Equation (3) delivers the strike price
of a variance swap 1
Q
t

1\
2
t;

, and is referred to as the model-free implied variance. Even

though variance emerges naturally from a portfolio of options, it is volatility that participants
prefer to quote. Our empirical analysis focuses on volatility swaps, and we synthetically
construct the strike price of this contract as
1
Q
t
[1\
t;
] =
q
1
Q
t

1\
2
t;

(4)
and refer to it as model-free implied volatility.
While straightforward, this approach is subject to a convexity bias. The main complication
in valuing volatility swaps arises from the fact that the strike of a volatility swap is not
equal to the square root of the strike of a variance swap due to Jensens inequality, i.e.,
1
Q
t
[1\
t;
]
q
1
Q
t

1\
2
t;

. The convexity bias that arises from the above inequality leads
to imperfect replication when a volatility swap is replicated using a buy-and-hold strategy of
variance swaps (e.g., Broadie and Jain, 2008). Simply put, the payo of variance swaps is
quadratic with respect to volatility, whereas the payo of volatility swaps is linear.
We deal with this bias in approximation in two ways. First, we measure the convexity
bias using a second-order Taylor expansion as in Brockhaus and Long (2000) and nd that
it is empirically small.
9
More importantly, when we re-do our empirical exercise with model-
9
Brockhaus and Long (2000) show that 1
Q
|
[1\
|,r
] =
q
1
Q
|

1\
2
|,r

\
2
8n
3=2
where : and \
2
denote the
mean and variance of the future realized variance, respectively, under the risk-neutral measure Q. 1
Q
|
[1\
|,r
]
is certainly less than or equal to
q
1
Q
|

1\
2
|,r

due to the Jensens inequality, and \

2
,8:
3/2
measures the
convexity error.
7
free implied variances, we nd virtually identical results. Hence the convexity bias has no
discernible eect on our results and the approximation in Equation (4) works well in our
framework, which explains why it is widely used by practitioners (e.g., Knauf, 2003).
Computing model-free implied volatility requires the existence of a continuum in the cross-
section of option prices at time t with maturity date t. In the FX market, over-the-counter
options are generally quoted in terms of Garman and Kohlhagen (1983) implied volatilities
at xed deltas. Liquidity is generally spread across ve levels of deltas. From these quotes,
we extract ve strike prices corresponding to ve plain vanilla options, and follow Jiang and
Tian (2005), who present a simple method to implement the model-free approach when option
prices are only available on a nite number of strikes.
Specically, we use a cubic spline around these ve implied volatility points. This inter-
polation method is standard in the literature (e.g., Bates, 1991; Campa, Chang, and Reider,
1998; Jiang and Tian, 2005; Della Corte, Sarno, and Tsiakas, 2011) and has the advantage
that the implied volatility smile is smooth between the maximum and minimum available
strikes. We then compute the option values using the Garman and Kohlhagen (1983) valu-
ation formula,
10
and use trapezoidal integration to solve the integral in Equation (3). This
method introduces two types of approximation errors: (i) the truncation errors arising from
observing a nite number, rather than an innite set of strike prices, and (ii) a discretization
error resulting from numerical integration. Jiang and Tian (2005), however, show that both
errors are small, if not negligible, in most empirical settings.
Volatility Risk Premium. In this paper we study the predictive information content
in volatility risk premia for future exchange rate returns. To this end, we work with the
ex-ante payo or expected volatility premium to a volatility swap contract. The volatility
risk premium can be thought of as the dierence between the physical and the risk-neutral
expectations of the future realized volatility.
11
Formally, the t-period volatility risk premium
at time t is dened as
\ 11
t;
= 1
P
t
[1\
t;
] 1
Q
t
[1\
t;
] (5)
10
This valuation formula can be thought of as the Black and Scholes (1973) formula adjusted for having
both domestic and foreign currency paying a continuous interest rate.
11
A number of papers dene the volatility risk premium as dierence between the risk-neutral and the
physical expectation. Here we follow Carr and Wu (2009) and take the opposite denition as it naturally
arises from the long-position in a volatility swap contract.
8
where 1
P
t
[] is the conditional expectation operator at time t under the physical measure
P. Following Bollerslev, Tauchen, and Zhou (2009), we proxy 1
P
t
[1\
t;
] by simply using the
lagged realized volatility, i.e., 1
P
t
[1\
t;
] = 1\
t;
=
q
252

i=0
:
2
ti
, where :
t
is the daily
log return on the underlying security. This approach is widely used for forecasting exercises
it makes \ 11
t;
directly observable at time t, requires no modeling assumptions, and is
consistent with the stylized fact that realized volatility is a highly persistent process. Thus, at
time t, we measure the volatility risk premium over the [t. t +t] time interval as the ex-post
realized volatility over the [t t. t] interval and the ex-ante risk-neutral expectation of the
future realized volatility over the [t. t +t] interval, i.e., \ 11
t;
= 1\
t;
1
Q
t
[1\
t;
].
For our purposes, we viewcurrencies with high \ 11
t;
as those which are relatively cheap
to insure at each point in time t, as their expected realized volatility under the physical measure
(i.e., the variable against which agents hedge) is lower than the cost of purchasing option-based
insurance which is primarily driven by expected volatility under the risk-neutral measure.
Conversely, those currencies with relatively low \ 11
t;
are more expensive to insure at
time t. Our adoption of this terminology closely follows the logic in Garleanu, Pedersen, and
Poteshman (2009), who provide theory and empirical evidence to support the conjecture that
end-user demand for options has eects on their prices when dealers cannot perfectly hedge.
3 Data and Currency Portfolios
We now describe the data and the construction of currency portfolios that we employ in our
analysis.
Exchange Rate Data. We collect daily spot and one-month forward exchange rates vis-
-vis the US dollar (USD) from Barclays and Reuters via Datastream. The empirical analysis
uses monthly data obtained by sampling end-of-month rates from January 1996 to August
2011. Our sample consists of the following 20 countries: Australia, Brazil, Canada, Czech
Republic, Denmark, Euro Area, Hungary, Japan, Mexico, New Zealand, Norway, Poland,
Singapore, South Africa, South Korea, Sweden, Switzerland, Taiwan, Turkey, and United
Kingdom. We refer to this cross-section as Developed and Emerging Countries. A number
of currencies in this sample may not be traded in large amounts, even though quotes on forward
9
contracts (deliverable or non-deliverable) are available.
12
Hence, we also consider a subset of
the most liquid currencies, which we refer to as Developed Countries. This sample includes:
Australia, Canada, Denmark, Euro Area, Japan, New Zealand, Norway, Sweden, Switzerland,
and the United Kingdom.
Currency Option Data. We employ daily data from January 1996 to August 2011 on
over-the-counter (OTC) currency options, obtained from JP Morgan.
The OTC currency option market is characterized by specic trading conventions. While
exchange traded options are quoted at xed strike prices and have xed calendar expiration
dates, currency options are quoted at xed deltas and have constant maturities. More impor-
tantly, while the former are quoted in terms of option premia, the latter are quoted in terms
of Garman and Kohlhagen (1983) implied volatilities on baskets of plain vanilla options.
For a given maturity, quotes are typically available for ve dierent combinations of plain-
vanilla options: at-the-money delta-neutral straddles, 10-delta and 25-delta risk-reversals,
and 10-delta and 25-delta buttery spreads. The delta-neutral straddle combines a call and a
put option with the same delta but opposite sign this is the at-the-money (ATM) implied
volatility quoted in the FX market. In a risk reversal, the trader buys an out-of-the money
(OTM) call and sells an OTM put with symmetric deltas. The buttery spread is built up
by buying a strangle and selling a straddle, and is equivalent to the dierence between the
average implied volatility of an OTM call and an OTM put, and the implied volatility of a
straddle. From these data, one can recover the implied volatility smile ranging from a 10-delta
put to a 10-delta call.
13
To convert deltas into strike prices, and implied volatilities into option
prices, we employ domestic and foreign interest rates, obtained from JP Morgan, which are
equivalent to those obtained using Datastream and Bloomberg.
This recovery exercise yields data on plain-vanilla European call and put options on 20
currency pairs vis--vis the US dollar, with maturity of one year. Practitioner accounts suggest
that natural hedgers such as corporates prefer hedging using intermediate-horizon derivative
contracts to the more transactions-costs intensive strategy of rolling over short term positions
12
According to the Triennial Survey of the Bank for International Settlements (2013), the top 10 currencies
account for about 90 percent of the average daily turnover in FX markets.
13
In market jargon, a 25-delta call is a call whose delta is 0.25 whereas a 25-delta put is a put with a delta
equal to 0.25.
10
in currency options, and hence the one-year volatility swap is a logical contract maturity to
detect interactions between hedgers and speculators. It is also the most liquid maturity.
14
Hedge Fund Flows. To construct a measure of new arbitrage capital available to
hedge funds, we use data from a large cross-section of hedge funds and funds-of-funds from
January 1996 to December 2011, which is consolidated from data in the HFR, CISDM, TASS,
Morningstar, and Barclay-Hedge databases, and comprises of roughly US$ 1.5 trillion worth
of assets under management (AUM) towards the end of the sample period. Patton and
Ramadorai (2013) provide a detailed description of the process followed to consolidate these
data.
We select the subset of 634 funds from these data, those self-reporting as currency funds
or global macro funds, and construct the net ow of new assets to each fund as the change
in the funds AUM across successive months, adjusted for the returns accrued by the fund
over the month this is tantamount to an assumption that ows arrive at the end of the
month, following return accrual. We then normalize the gures by dividing them by the
lagged AUM, and then value-weight them across funds to create a single aggregate time-series
index of capital ows to currency and global macro funds.
Positions on Currency Futures. We also employ weekly data from the Commitments
of Traders, a report issued by the Commodity Futures Trading Commission (CFTC). The
report aggregates the holdings of participants in the US futures markets (primarily based in
Chicago and New York). It is typically released every Friday and reects the commitments of
traders for the prior Tuesday. The CFTC provides a breakdown of aggregate positions held
by commercial traders and nancial (or non-commercial) traders. The former are merchants,
foreign brokers, clearing members or banks using the futures market primarily to hedge their
business activities. The latter are hedge funds, large nancial institutions and individual
investors using the futures market for speculative purposes. We collect weekly data from
January 1996 to August 2011 on the Australian dollar, Brazilian real, Canadian dollar, Euro,
Japanese yen, Mexican peso, New Zealand dollar, South African rand, Swiss Franc, and British
pound relative to the USD dollar.
14
This is dierent from currency options per se, which tend to be most liquid at shorter maturities of one
and three months.
11
In our empirical analysis, we use positions on currency futures for two exercises. Firstly, we
construct an aggregate hedging measure of FX risk as in Acharya, Lochstoer, and Ramadorai
(2013), and report a detailed description in the Appendix. Secondly, we examine whether the
buying and selling actions of dierent players in futures market follow the pattern implied by
the \ 11 strategy.
Currency Excess Returns. We dene spot and forward exchange rates at time t as
o
t
and 1
t
, respectively. Exchange rates are dened as units of US dollars per unit of foreign
currency such that an increase in o
t
indicates an appreciation of the foreign currency. The
excess return on buying a foreign currency in the forward market at time t and then selling it
in the spot market at time t + 1 is computed as 1A
t+1
= (o
t+1
1
t
) ,o
t
, which is equivalent
to the spot exchange rate return minus the forward premium 1A
t+1
= ((o
t+1
o
t
) ,o
t
)
((1
t
o
t
) ,o
t
). According to the CIP condition, the forward premium approximately equals
the interest rate dierential (1
t
o
t
) ,o
t
' i
t
i

t
, where i
t
and i

t
represent the domestic
and foreign riskless rates respectively, over the maturity of the forward contract. Since CIP
holds closely in the data at daily and lower frequency (e.g., Akram, Rime and Sarno, 2008),
the currency excess return is approximately equal to an exchange rate component (i.e., the
exchange rate change) minus an interest rate component (i.e., the interest rate dierential):
1A
t+1
' ((o
t+1
o
t
) ,o
t
) (i
t
i

t
).
Carry Trade Portfolios. At the end of each period t, we allocate currencies to ve
portfolios on the basis of their interest rate dierential relative to the US, (i

t
i
t
) or forward
premia since (1
t
o
t
) ,o
t
= (i

t
i
t
) via CIP. This exercise implies that Portfolio 1 com-
prises 20% of all currencies with the highest interest rate dierential (lowest forward premia)
and Portfolio 5 comprises 20% of all currencies with the lowest interest rate dierential (high-
est forward premia), and we refer to the long-short portfolio formed by going long Portfolio 1
and short Portfolio 5 as C1. We compute the excess return for each portfolio as an equally
weighted average of the currency excess returns within that portfolio, and individually track
both the interest rate dierential and the spot exchange rate component that make up these
excess returns.
Lustig, Roussanov, and Verdelhan (2011) study these currency portfolio returns using their
rst two principal components. The rst principal component implies an equally weighted
12
strategy across all long portfolios, i.e., borrowing in the US money market and investing in
foreign money markets. We refer to this zero-cost strategy as 1C1. The second principal
component is equivalent to a long position in Portfolio 1 (investment currencies) and a short
position in Portfolio 5 (funding currencies), and corresponds to borrowing in the money mar-
kets of low yielding currencies and investing in the money markets of high yielding currencies.
We refer to this long/short strategy as C1 in our tables and we use both 1C1 and C1
in risk-adjustment below.
Momentum Portfolios. At the end of each period t, we form ve portfolios based on
exchange rate returns over the previous 3-months. We assign the 20% of all currencies with
the highest lagged exchange rate returns to Portfolio 1, and the 20% of all currencies with
the lowest lagged exchange rate returns to Portfolio 5. We then compute the excess return
for each portfolio as an equally weighted average of the currency excess returns within that
portfolio. A strategy that is long in Portfolio 1 (winner currencies) and short in Portfolio 5
(loser currencies) is then denoted as `C`.
15
Value Portfolios. At the end of each period t, we form ve portfolios based on the
level of the real exchange rate.
16
We assign the 20% of all currencies with the lowest real
exchange rate to Portfolio 1, and the 20% of all currencies with the highest real exchange
rate to Portfolio 5. We then compute the excess return for each portfolio as an equally
weighted average of the currency excess returns within that portfolio. A strategy that is long
in Portfolio 1 (undervalued currencies) and short in Portfolio 5 (overvalued currencies) is then
denoted as \ 1.
Risk Reversal Portfolios. At the end of each period t, we form ve portfolios based
on out-of-the-money options. We compute for each currency in each time period the risk
reversal, which is the implied volatility of the 10-delta call less the implied volatility of the 10-
15
Consistent with the results in Menkho, Sarno, Schmeling and Schrimpf (2012b), sorting on lagged
exchange rate returns or lagged currency excess returns to form momentum portfolios makes no qualitative
dierence to our results below. The same is true if we sort on returns with other formation periods in the
range from 1 to 12 months.
16
We compute the real exchange rate at the end of each month as 111
|
= o
|
,111
|
, where o
|
is the
nominal exchange rate and 111
|
is the purchasing power parity rate. We collect the PPP data published
annually every March by the OECD, and retrieve monthly data by forward lling, i.e., we use the last available
PPP rate until the next February. For Singapore and Taiwan, we use data from the PENN World Table.
13
delta put, and assign the 20% of all currencies with the lowest risk reversal to Portfolio 1, and
the 20% of all currencies with the highest risk reversal to Portfolio 5. We then compute the
excess return for each portfolio as an equally weighted average of the currency excess returns
within that portfolio. A strategy that is long in Portfolio 1 (high-skewness currencies) and
short in Portfolio 5 (low-skewness currencies) is then denoted as 11.
Volatility Risk Premia Portfolios. At the end of each period t, we group currencies
into ve portfolios using the 1-year volatility risk premium constructed as described earlier.
We allocate 20% of all currencies with the highest expected volatility premia, i.e., those which
are cheapest to insure, to Portfolio 1, and 20% of all currencies with the lowest expected
volatility premia, i.e., those which are expensive to insure, to Portfolio 5. We then compute
the average excess return within each portfolio, and nally calculate the portfolio return from
a strategy that is long in Portfolio 1 (cheap volatility insurance) and short in Portfolio 5
(expensive volatility insurance), and denote it \ 11.
4 The \ 11 Strategy: Empirical Evidence
4.1 Summary Statistics and the Returns to \ 11
Table 1 presents summary statistics for the annualized average realized volatility 1\
t;
, syn-
thetic volatility swap rate o\
t;
= 1
Q
t
[1\
t;
], and volatility risk premium \ 11
t;
= 1\
t;

o\
t;
for the 1-year maturity (t = 1) (in what follows, we drop the t subscript, as it is always
1 year).
The table shows that, on average across developed currencies, 1\
t
equals 10.68 percent,
with a standard deviation of 2.88 percent, and o\
t
equals 11.31 percent, with a standard
deviation of 2.75 percent. The average volatility risk premium \ 11
t
across these currencies,
which is the dierence of these two variables, is equal to 0.62 percent, with a standard
deviation of 1.58 percent. For the full sample of developed and emerging countries, 1\
t
and o\
t
are slightly larger than for the sample of only developed currencies, and so is the
volatility risk premium, \ 11
t
, which equals 0.92 on average. We might expect to see this
the average price that natural hedgers have to pay to satisfy their demand for volatility
insurance is larger when including emerging market currencies.
14
Table 2 describes the returns generated by our short expensive-to-insure, long cheap-to-
insure currency strategy, reporting summary statistics for the ve portfolios that are obtained
when sorting on the volatility risk premium. In this table, 1
L
is the long portfolio that buys
the top 20% of all currencies with the cheapest volatility insurance, 1
2
buys the next 20% of
all currencies ranked by expected volatility premia, and so on till the fth portfolio, 1
S
which
is the portfolio that buys the top 20% of all currencies which are the most expensive to insure.
\ 11 essentially buys 1
L
and sells 1
S
, with equal weights, so that \ 11 = 1
L
1
S
.
The table reveals several facts about \ 11. First, there is a strong general tendency of
portfolio returns to decrease as we move from 1
L
towards 1
S
; the decrease is not monotonic
for developed countries, but it is monotonic for the full sample for the FX returns component.
Second, the \ 11 return stems mainly from the long portfolio, 1
L
. Third, the return from 1
L
can be almost completely attributed to spot rate changes. Finally, the bottom panel of Table
2 shows the transition matrix between portfolios. This shows that there is currency rota-
tion across quintile portfolios such that the steady-state transition probabilities are identical.
Thus the performance of the strategy cannot simply be attributed to long-lived positions in
particular currencies.
The returns to \ 11 are very robust. We describe a few robustness checks before pro-
ceeding further. First, we compute volatility risk premia using simple at-the-money implied
volatility rather than the more complicated model-free implied volatility. We also implement
the simple variance swap formula of Martin (2012). In both cases, results are virtually iden-
tical for developed countries and improve for developed and emerging countries. We report
these results in the Internet Appendix. Second, in our empirical work we also experiment
with an AR(1) process for 1\ to form expectations of 1\ rather than using lagged 1\
over the previous 12 months. Again, we nd that the results are virtually identical to those
reported in Table 2. Third, we report the net of transactions costs returns to \ 11 and
other currency strategies in the Internet Appendix, and show that these are similar to those
reported in Table 2, especially for the more liquid Developed sample of countries. Fourth,
in the Internet Appendix, we check whether a simple strategy based on sorting currencies
by the dierence between longer-term and short-term realized volatility eectively captures
the returns from \ 11. Using denitions of long-term ranging from six to 24 months and
short-term from one to six months, we nd that while there are a number of high-return
15
portfolios in the set, there is substantial variation in these returns across portfolios, leading
to concerns of potential data-mining. Finally, we show in the Internet Appendix that the
identities of the currencies most often found in the corner \ 11 portfolios are not easily
recognizable from other currency strategies such as carry. We formalize this last exercise
by explicitly comparing the returns of \ 11 to the conventional set of currency strategies
considered in the literature thus far, which we present in the next section.
4.2 Comparing \ 11 and Other Currency Strategies
In Table 3, we present the returns to a number of long-short currency strategies computed using
only time t 1 information, to compare the predictability generated by strategies previously
proposed in the literature with the new \ 11 strategy that we propose. We compare C1,
`C`, \ 1, and 11 with our \ 11 based strategy. We report results for both subsamples
(Developed, and Developed and Emerging) in our data.
Panel A of the table shows the results for the portfolio excess returns (including interest-
rate dierentials) generated by these trading strategies. Consistent with a vast empirical
literature (e.g., Lustig, Roussanov, and Verdelhan, 2011, Burnside, Eichenbaum, Kleshchelski,
and Rebelo, 2011, and Menkho, Sarno, Schmeling, and Schrimpf, 2012a), C1 delivers a
very high average excess return indeed, the highest of all strategies considered. The Sharpe
ratio of the carry trade is 0.61 for the sample of developed countries, and 0.74 for the full
sample. `C` also generates positive excess returns, albeit less striking than carry, which is
consistent with the recent evidence in Menkho, Sarno, Schmeling, and Schrimpf (2012b) that
the performance of currency momentum has weakened substantially during the last decade;
the Sharpe ratio is 0.27 for both samples of countries. Both \ 1 and 11 do very well, with
Sharpe ratios of 0.62 and 0.48 respectively.
In contrast, the \ 11 strategy that we introduce generates a Sharpe ratio of 0.48 and 0.29
for the two samples of countries considered, signifying that it outperforms the momentum
strategy. The \ 11 strategy works better for the developed countries in our sample than for
the whole sample of developed and emerging countries. One plausible explanation for this
is that there is a greater prevalence of hedging using more sophisticated instruments such as
currency options in developed markets than in emerging markets.
Panel A of the table suggests that the returns to the \ 11 strategy are somewhat modest
16
in comparison with those of the other strategies that we provide as comparison. However,
Panel B of the table introduces the main benet of the \ 11 strategy, namely that the
lions share of its returns accrue as a result of spot rate predictability. This predictability is
virtually twice as large as the best competitor strategy over the sample period, generating an
annualized mean spot exchange rate return of 4.4% for the developed countries, and 3.72%
for the full cross-section of all 20 countries in our sample. In contrast, the exchange rate
return from C1 is close to zero for both samples, and while other strategies, notably \ 1,
have relatively better performance in predicting movements in the spot rate than C1, the
preponderance of their returns are derived from interest rate dierentials.
Several of the other moments presented in Panel B of Table 3 are also worth highlighting.
First, the returns from \ 11 display desirable skewness properties, as its unconditional skew-
ness is positive (albeit small for the full sample), and the maximum drawdown is comparable
to that of `C` and far better (i.e., higher) than that of C1. Another way to see this,
of course, is to compare the (very dierent) returns to 11 and \ 11, as 11 is constructed
to replicate a long high skewness-short low skewness portfolio. Finally, the table shows that
the portfolio turnover of the \ 11 strategy (measured in terms of changes in the composition
of the short and long legs of the \ 11 strategy, 1:c
S
and 1:c
L
in Table 3) is reasonable
lying in between the very low turnover of C1 and the high turnover of `C`. This means
that the \ 11 strategy is likely to perform well also for lower rebalancing periods and that
transaction costs which are known to be relatively small in currency markets are unlikely
to impact signicantly on the performance of \ 11.
4.3 Combining \ 11 with Other Currency Strategies
Panel C of Table 3 documents the correlation of the \ 11 strategy with the other strategies,
and nds that the strategy tends to be negatively correlated with C1 (with correlations of
-0.18 and -0.21 for the two samples) and only mildly positively correlated with `C` (with
correlations of 0.09 and 0.10 for the two samples). The correlation with \ 1 for Developed
countries is higher, but at 0.23 there is substantial orthogonal information in the strategy
indeed several of the other strategies are much more highly correlated with one another.
Apart from showing that the strategy is distinct from those already studied in the literature,
this also implies that combining \ 11 with C1, `C`, \ 1, and 11 may well yield
17
sizable diversication benets to an investor. It is also useful to note that the correlations for
the excess returns from the strategies, presented in the table, are very close in magnitude to
the correlations acquired from the exchange rate component of these returns in other words,
it is the currency component of the returns to this strategy that is the proximate source of
the diversication benets.
Figure 1 provides a graphical illustration of the dierences in the performance of the
strategies highlighted in Table 2, and restricts the plot to the sample of Developed Countries
to conserve space. The gure plots the one-year rolling Sharpe ratio for these strategies, and
makes visually clear the marked dierence in the evolution of risk-adjusted returns of \ 11
relative to the others. While there is a substantial improvement in the Sharpe ratio of \ 11
during the recent crisis, the strategy is not driven entirely by this episode the Sharpe ratio
has been relatively stable over the sample period, and appears to be no more volatile than
the Sharpe ratio of C1 and `C`.
Table 4 shows the subsample performance of the currency component of these strategies
as a complement to Figure 1. It is clear that the performance of \ 11 is greater in crisis
and NBER recession periods. However it is important to highlight that, outside of these
recession periods, the return to \ 11 is still large and positive, and higher than that of all
the competitor strategies. Even if \ 11 were to be used primarily as a hedge, it has very
desirable properties, delivering positive returns outside of crisis periods, and very high returns
within crisis periods.
Figure 2 plots the cumulative returns of the strategies over the sample period (again, only
for the Developed Countries), decomposing the cumulative excess return into its two con-
stituents: the exchange rate component (FX) and the interest rate gain component (yield).
Both C1 and `C` have a positive yield component, although in the case of the carry
trade the yield component is the sole positive driver of the carry return because the cumu-
lative FX return component is negative. For `C`, most of the excess return is driven by
spot predictability, so the cumulative yield component has a positive but relatively minor
contribution to momentum returns. \ 11 returns are dierent in that they are made up of a
mildly negative yield component (for both sample of countries considered), and therefore the
component due to spot return predictability is in fact larger than the full portfolio return.
The performance of \ 11 is similar to \ 1, except that \ 1 also has positive yield, but far
18
lower currency returns.
Taken together, the results from this section suggest that, while the carry trade strategy is
taken in isolation the best performing strategy in terms of excess returns and delivers the
highest Sharpe ratio, the \ 11 strategy has creditable excess returns overall, an important
tendency to deliver returns during crisis periods that are far higher than the crashes commonly
experienced with the carry trade, and far stronger predictive power for exchange rate returns,
which is a unique feature relative to alternative currency trading strategies. The importance
of these last two features of the \ 11 strategy is twofold. First, a currency investor would
likely gain a great deal of diversication benet from adding \ 11 to a currency portfolio
to enhance risk-adjusted returns. Second, a spot currency trader interested in forecasting
exchange rate uctuations (as opposed to currency excess returns) would greatly value the
signals provided by \ 11.
To shed light on the added value of the \ 11 strategy for a currency investor, we compute
the optimal currency portfolio for an investor who uses all of the ve strategies considered
here: C1, \ 1, 11, `C`, and \ 11. Specically, consider a portfolio of ` assets with
covariance matrix . The global minimum variance portfolio is the portfolio with the lowest
return volatility and represents the solution to the following optimization problem: min n
0
n
subject to the constraint that the weights sum to unity, n
0
t = 1,where n is the `1 vector of
portfolio weights on the risky assets, t is a `1 vector of ones, and is the `` covariance
matrix of the asset returns. The weights of the global minimum variance portfolios are given
by n =

1

. We compute the optimal weights for both the Developed and Developed &
Emerging samples, and report the results graphically in Figure 3.
The results show that the optimal weight assigned to the \ 11 strategy is the highest
across all ve currency strategies, equalling 33 percent or a full third of the portfolio, and the
same in both samples of currencies. The Sharpe ratio of the minimum volatility portfolio
for the Developed sample, for instance, is quite impressive and equal to 0.92. However, it
would drop substantially to 0.79 if the investor had no access to the \ 11 strategy (i.e., only
employs the other four currency strategies). Similarly for the Developed & Emerging sample.
These ndings conrm the value of \ 11 in a currency portfolio despite its return not being
the highest among the strategies considered. It has extremely desirable correlation properties
which cannot be replicated using information from any of the other well-studied currency
19
strategies.
17
5 Pricing \ 11 Returns
In this section we carry out both cross-sectional and time-series asset pricing tests to determine
whether \ 11 returns can be understood as compensation for systematic risk.
5.1 Time-Series Regressions
As a rst step, Table 5 simply regresses the time-series of \ 11 returns on a number of risk
factors proposed in the literature. First, Panel A conrms the results found in Tables 2 and 3,
by using 1C1, C1, `C`, \ 1, and 11 as right-hand side variables, and shows that for
both Developed and Developed and Emerging samples, there is substantial alpha relative to
these factors. Panel B uses the three Fama-French factors and adds equity market momentum,
denoted `C`1. Again, \ 11 has alpha relative to these factors which is virtually identical
to that in the prior panel. Finally, Panel C of Table 5 employs the Fung-Hsieh (2004) factor
model, which has been used in numerous previous studies; see for example, Bollen and Whaley
(2009), Ramadorai (2013), and Patton and Ramadorai (2013). The set of factors comprises
the excess return on the S&P 500 index; a small minus big factor constructed as the dierence
between the Wilshire small and large capitalization stock indexes; excess returns on portfolios
of lookback straddle options on currencies, commodities, and bonds, which are constructed
to replicate the maximum possible return to trend-following strategies on their respective
underlying assets; the yield spread of the US 10-year Treasury bond over the 3-month T-bill,
adjusted for the duration of the 10-year bond; and the change in the credit spread of Moodys
BAA bond over the 10-year Treasury bond, also appropriately adjusted for duration. Yet
again, the table shows that the alpha of \ 11 is virtually unaected by the inclusion of these
factors.
5.2 Cross-Sectional Tests
Our cross-sectional tests rely on a standard stochastic discount factor (SDF) approach (Cochrane,
2005), and we focus on a set of risk factors in our investigation that are motivated by the ex-
17
Results are qualitatively identical for the larger sample of Developed and Emerging Countries.
20
isting currency asset pricing literature. We begin by briey reviewing the methods employed,
and denote excess returns of portfolio i in period t + 1 by 1A
i
t+1
. The usual no-arbitrage
relation applies, so risk-adjusted currency excess returns have a zero price and satisfy the
basic Euler equation:
E[`
t+1
1A
i
t+1
] = 0. (6)
with a linear SDF `
t
= 1 /
0
(,
t
j), where ,
t
denotes a vector of risk factors, / is the vector
of SDF parameters, and j denotes factor means.
This specication implies a beta pricing model in which expected excess returns depend
on factor risk prices `, and risk quantities ,
i
, which are the regression betas of portfolio excess
returns on the risk factors:
E

1A
i

= `
0
,
i
(7)
for each portfolio i (see e.g., Cochrane, 2005).
The relationship between the factor risk prices in equation (7) and the SDF parameters in
equation (6) is simply given by ` =
f
/, where
f
is the covariance matrix of the risk factors.
Thus, factor risk prices can be easily obtained via the SDF approach, which we implement
by estimating the parameters of equation (6) via the generalized method of moments (GMM)
of Hansen (1982).
18
We also present results from the more traditional two-stage procedure of
Fama and MacBeth (1973) in our empirical implementation.
In our asset pricing tests we consider a two-factor linear model that comprises 1C1 and one
additional risk factor, which is one of C1 and \ C1
FX
. 1C1 denotes the average return
from borrowing in the US money market and equally investing in foreign money markets.
C1 is the carry portfolio described earlier. \ C1
FX
is a global FX volatility risk factor
constructed as the innovations to global FX volatility, i.e., the residuals from an autoregressive
model applied to the average realized volatility of all currencies in our sample, as in Menkho,
Sarno, Schmeling, and Schrimpf (2012a).
19
18
Estimation is based on a pre-specied weighting matrix and we focus on unconditional moments (i.e., we
do not use instruments other than a constant vector of ones) since our interest lies in the performance of the
model to explain the cross-section of expected currency excess returns (see Cochrane, 2005; Burnside, 2011).
19
In the Internet Appendix, we also consider the innovations to global average precentage bid-ask spreads
in the spot market (1o
J
) and the option market (1o
1\
). 1o
J
is constructed by averaging over
a month the daily average bid-ask spread of the spot exchange rates. 1o
1\
is constructed by averaging
21
In assessing our results, we are aware of the statistical problems plaguing standard asset
pricing tests, recently emphasized by Lewellen, Nagel, and Shanken (2010). Asset pricing
tests can often be highly misleading, in the sense that they can indicate strong but illusory
explanatory power through high cross-sectional 1
2
statistics, and small pricing errors, when
in fact a risk factor has weak or even non-existent pricing power. Given the relatively small
cross-section of currencies in our data, as well as the relatively short time span of our sample,
these problems can be severe in our tests. As a result, when interpreting our results, we only
consider the cross-sectional 1
2
and Hansen-Jagannathan (HJ) tests on the pricing errors,
if we can condently detect a statistically signicant risk factor, i.e., if the GMM estimates
clearly point to a statistically signicant market price of risk ` on a factor.
Table 6 reports GMM estimates of /, portfolio-specic ,s, and implied `s, as well as
cross-sectional 1
2
statistics and the HJ distance measure (Hansen and Jagannathan, 1997).
In the table, standard errors are constructed as in Newey and West (1987) with optimal lag
length selection according to Andrews (1991). Besides the GMM tests, we employ traditional
Fama-MacBeth (FMB) two-pass OLS regressions to estimate portfolio betas and factor risk
prices. Note that we do not include a constant in the second stage of the FMB regressions, i.e.
we do not allow a common over- or under-pricing in the cross-section of returns - however our
results are virtually identical when we replace the 1C1 factor with a constant in the second
stage regressions.
20
Since 1C1 has virtually no cross-sectional relation to portfolio returns,
it serves the same purpose as a constant that allows for common mispricing.
Panels A and B of Table 6 show clearly how neither of the risk factors considered enters the
SDF with a statistically signicant risk price `, and that this is the case for both the developed
countries and the full sample. As expected, the FMB results in the table are qualitatively, and
in most cases also quantitatively identical to the one-step GMM results. The bottom part of
the panels show that there is little cross-sectional variation across the 5 portfolios sorted by
the cost of currency insurance, which is what we conrm more formally in the asset pricing
tests.
over a month the daily average bid-ask spread of the 1-year at-the-money implied volatilities. Innovations
are computed as the residuals to a rst-order autoregressive process. Higher bid-ask spreads indicate lower
liquidity, so that our aggregate measures can be seen as global proxies for the FX spot market and the FX
option market illiquidity, respectively.
20
Also see Lustig and Verdelhan (2007) and Burnside (2011) on the issue of whether or not to include a
constant in these regressions.
22
The best performing SDF in these tests includes 1C1 and \ C1
FX
, and generates a
respectable cross-sectional 1
2
(0.27), but the market price of risk is insignicantly dierent
from zero. The HJ test delivers large p-values for the null of zero pricing errors in all cases
but we attach no information to this result given the lack of clear statistical signicance of the
market price of risk. We also carried out asset pricing tests using the same methods and risk
factors in which we attempt to price only the exchange rate component of the returns from
\ 11, and the Internet Appendix attempts using an illiquidity factor constructed as the
average bid-ask spread across all currencies. In both cases, the results are equally disappointing
in that all risk factors included in the various SDF specications are statistically insignicant.
Overall, the asset pricing tests reveal that it is not possible to understand the returns
from the \ 11 strategy as compensation for global risk, using the carry risk factor, global
volatility risk, or illiquidity in the FX market of the kind used in the literature. These results
are consistent with our earlier results that indicate that \ 11 returns are very dierent from
the returns of conventional currency strategies, and hence their source is likely to stem from
a dierent mechanism than compensation for canonical sources of systematic risk. Therefore,
we now turn to examining potential explanations.
6 Understanding the Drivers of \ 11
We consider two possible alternative explanations for our results. The rst is Aversion
to Volatility Risk. It might be the case that the currency-specic volatility risk premium
captures uctuations in aversion to volatility risk. As a result, currencies with relatively
expensive volatility insurance would have low expected returns and vice versa.
Our \ 11 strategy is cross-sectional, since we are long and short currencies simultaneously.
As a result, if this explanation were correct, it would rely either on dierent currencies loading
dierently on a global volatility shock, or indeed on market segmentation causing expected
returns on dierent currencies to be determined independently. This latter possibility is very
dicult to evaluate, and if our strategy did indeed provide evidence of this, it would have
far-reaching consequences.
To evaluate the rst of these possibilities, i.e., currencies loading dierently on a global
volatility shock, we describe above the ineectiveness of using the global FX volatility risk
23
factor of Menkho, Sarno, Schmeling, and Schrimpf (2012a) to price the returns from our
portfolio. However, it could be the case that this proxy is not the best suited to capture the
returns from our strategy, and we try other possibilities. We do so by estimating the loadings
of currency returns on various proxies for global volatility risk, and building portfolios sorted
on these estimated loadings. Specically, we estimate the following regression:
1A
it
= c
i
+,
i
G\ C1
t
+
it
.
for each currency i. Here G\ C1 is a proxy for global volatility risk premia and we employ
various measures, including the average volatility risk premium across our currencies (with
equal weights); the rst principal component of the currencies volatility risk premia; and
the equity volatility risk premium computed as the dierence between the time-t one-month
realized volatility on the S&P500 and the VIX index.
We estimate these regressions using rolling windows of 36 months. After obtaining esti-
mates of the ,
i
coecients, we sort currencies into ve portfolios on the basis of these ,
i
estimates. Finally, we construct a long-short strategy which buys currencies with low betas
and sells currencies with high betas. In essence, this strategy exploits dierences in exposure
of individual currencies to global measures of volatility risk premia, which is a direct test of
the above hypothesis.
The results using our three measures for G\ C1 are qualitatively identical and we report
in Table 7 the results for G\ C1 set equal to the average volatility risk-premium across the
currencies in our sample. The Internet Appendix contains results for the other two measures.
The table shows that the performance of this strategy is strictly inferior to the performance
of the \ 11 strategy, and the correlation between the returns from the two strategies is tiny.
On the basis of this evidence, we conclude that there is little support for \ 11 returns being
driven by aversion to global volatility risk in the data.
The second possible explanation that we consider is Limits to Arbitrage, in the spirit
of Acharya, Lochstoer, and Ramadorai (2013). According to this explanation, the returns
to \ 11 arise from the interaction between natural hedgers of FX risk, and currency market
speculators. When the risk-bearing capacity of currency-market speculators is aected by
shocks to the availability of arbitrage capital, this will make currency options across the
board more expensive, with particular impacts on those currencies to which speculators have
24
high exposure. This will result in selling pressure on expensive-to-insure currencies as natural
hedgers such as corporations sell pre-existing currency holdings, abandon expensive currency
hedges, and become more reluctant to denominate contracts in these currencies. Conversely,
this mechanism results in relatively less pressure on cheap-to-insure currencies, for which
natural hedgers are happy to hold higher inventories or take on more real currency exposure.
This yields the positive long-short returns in the \ 11 portfolio.
This explanation has implications which we test in Table 8. The table presents coecients
from predictive time-series regressions of the exchange rate component of \ 11 on a number
of conditioning factors implied by this mechanism. We report results from the exchange rate
component of \ 11 since we are primarily interested in understanding the predictive power
for spot exchange rates, but the results for excess returns are, not surprisingly, qualitatively
identical and quantitatively very similar.
21
The rst column in both panels shows the univariate regression of the exchange rate
component of \ 11 on the 12-month rolling average of the lagged TED spread. When
funding liquidity is lower (i.e., times of high capital constraints on speculators), we should
nd that the expected (exchange rate) return from \ 11 should increase, and Table 8 provides
strong conrmation for this for developed countries. While the sign of the coecient on TED
is positive for the full sample of countries, it is not statistically signicant. This could be
because the TED spread is possibly less useful as a proxy for funding liquidity constraints in
emerging markets.
The second column shows that when the 12-month rolling average of changes in VIX (our
proxy for increases in the risk aversion of market participants, yielding both greater limits to
arbitrage and an increased desire to hedge) is positive, \ 11 returns increase (signicantly for
the sample of developed countries), again consistent with the limits to arbitrage explanation.
This is similar to the results in Nagel (2012), who shows that a strategy of liquidity provision in
equity markets has returns which are highly correlated with VIX. Similarly, the third column
shows that a general nancial distress indicator (FSI, constructed by the Federal Reserve Bank
of St. Louis) that captures the principal component of a variety of liquidity and volatility
indicators is statistically signicant.
The fourth column of the table interacts TED with changes in VIX, and nds strong
21
See the Internet Appendix for a detailed description of the conditioning factors used for this exercise.
25
statistically signicant predictive power of this interaction for the FX returns on our strategy in
both developed and emerging countries, suggesting that when funding liquidity is constrained
and risk aversion is high, \ 11 returns increase.
The next three columns check the predictive ability for \ 11 of market participants
positioning information. The rst two of these columns use the (normalized) net short futures
position of (both commercial and non-commercial) traders on the Australian dollar (l1)
and the Japanese yen (J11 ) relative to the USD dollar, respectively.
22
For Developed as well
as Developed and Emerging samples, at times when there is greater futures-related hedging
of the l1 by FX traders, the returns to the \ 11 strategy increase. However, we nd
no real impact for the net short position on the J11 . The nal column of the table adds in
measures of capital ows into hedge funds. When aggregate capital ows into hedge funds are
high, signifying that they experience fewer constraints on their ability to engage in arbitrage
transactions, we nd that returns for our \ 11 strategy are lower and vice versa.
The nal three rows of Table 8 consider several of the variables described above simul-
taneously to test their joint and separate explanatory power. We include TED, changes in
VIX and the interaction separately to avoid potential collinearity in the regressions as these
variables are highly correlated with one another. More generally, it is clear that the vari-
ables used in the univariate regressions are likely to contain a substantial common component.
Nonetheless, we nd that all these variables retain their signs and are generally statistically
signicant in these multivariate predictive regressions, oering support to the limits to arbi-
trage explanation of our results.
Next, we examine post-formation portfolio returns. If the limits to arbitrage explanation
is correct, the predictability of volatility insurance costs cannot be long-lived. According to
this explanation, either speculators face a shock that reduces their available arbitrage capital
and limits their ability to provide cheap volatility insurance, or there is an increase in hedger
risk aversion causing their demand for hedging to increase. As a result, net demand for
volatility insurance increases, making hedging more expensive, which will be reected in a
lower volatility risk premium, i.e., more expensive currency options. In the face of high
volatility insurance costs, natural hedgers scale back on the amount of spot currency they
22
l1 is taken as representative of a typical high-interest currency bought by carry traders, whereas J11
is a traditional safe haven currency.
26
are willing to hold, predictably depressing spot prices and leading to relatively low returns
on the spot currency position. When capital constraints loosen, however, we should see
the opposite behavior, i.e., the volatility risk premium reverts to the mean, and reversals in
currency returns.
This yields an additional testable implication, namely, reversal in post-formation cumula-
tive returns on the \ 11 strategy, which is exactly what we nd in Figure 4. The gure plots
cumulative post-formation risk-adjusted excess returns (left panel) and risk-adjusted currency
returns (right panel) over periods of 1. 2. . . . . 20 months for the VRP-sorted portfolios, for
both samples of countries examined.
23
Returns in the post-formation period are overlapping,
as we form new portfolios each month, but track these portfolios for 20 months.
In the gure, the excess returns increase and peak after 3 months for the Developed
Countries sample and 4 months for the full sample, and subsequently decline. Looking at spot
exchange rate returns, the peak in cumulative post-formation exchange rate return occurs
around 4 months for the developed sample and 5-6 months for the full sample. This evidence
of a reversal appears consistent with the prediction of the limits to arbitrage explanation of
the economic source of VRP predictive power. Moreover the relatively high frequency of
the reversal suggests that an explanation based on risk aversion to volatility combined with
market segmentation, an explanation described earlier, is somewhat less likely.
Finally, we examine whether the observed buying and selling actions of dierent players
in the currency market follow the pattern implied by the limits to arbitrage explanation,
i.e., that currencies in the high volatility-insurance portfolio are sold and those in the low
volatility insurance portfolio are bought. We do so using the CFTC data on the position
of commercial and nancial traders in FX markets, essentially taking the currencies ranked
by their volatility insurance costs, and documenting the traders positions (cumulative net
positions), rather than returns.
24
We view the CFTC position data as a proxy for cumulative
23
Specically, we plot returns net of the exposure to carry trade risk, i.e., we use the residuals from a
regression of \ 11 returns on C1, so that the returns can be considered as alphas over and above carry
trade returns. Using raw portfolio returns or their exchange rate component produces a very similar pattern
for the full sample, and a virtually identical pattern for the developed sample, as expected given that we know
already from previous analyses that C1 has little pricing power for \ 11 portfolios.
24
To allow for meaningful cross-currency comparisons, we need to ensure that net positions are comparable
across currencies, as their absolute size diers across currencies. We therefore divide net positions by their
standard deviation computed over a rolling window of 3 month.
27
order ow across dierent segments of FX market participants and a large proportion of the
total FX market, given that there is evidence that the CFTC position data and currency order
ow capture very similar information (e.g., Klitgaard and Weir, 2004).
The results of this exercise are reported in Figure 5, which plots the position in the \ 11
portfolio for nancial and non-nancial traders. We nd that the position of non-nancial
traders follows exactly the pattern implied by the limits to arbitrage explanation such
traders of currencies sell expensive-insurance (1
S
) currencies and buy cheaper-insurance (1
L
)
currencies. Financial traders display exactly the opposite behavior, with a strongly negative
position in the \ 11 portfolio, acting as market-makers that provide liquidity to satisfy the
buying (selling) demand for low (high)-insurance currencies.
25
Taken together, the results in this section lend support to a limits to arbitrage explanation
for the predictability of spot exchange rates associated with \ 11. There is a growing
theoretical and empirical literature that highlights the role of limits to arbitrage and the
interaction between hedgers and speculators in asset markets, and we view our results as
suggestive that currency markets may be another venue in which such mechanisms are at
work.
7 Conclusions
We show that the currency volatility risk premium has substantial predictive power for the
cross-section of currency returns. Sorting currencies by their volatility risk premia gener-
ates economically signicant returns in a standard multi-currency portfolio setting. This
predictive power is specically related to spot exchange rate returns, and not to interest
rate dierentials, and the spot rate predictability is much stronger than that observed from
carry, currency momentum, currency value, or risk-reversal strategies. Moreover, the returns
from the volatility risk premium strategy are largely uncorrelated with these other currency
strategies, thus providing a substantial diversication gain to investors.
25
We also replicate this exercise using a data set on customer order ow in the FX market from a major bank
over the sample period from January 2001 to December 2010. The data cover all currencies in our Developed
sample with the exception of the Danish Krone, and order ow is measured as net buying pressure against the
US dollar (i.e., buyer-initiated minus seller-initiated trades). The ow data are categorized into two groups:
commercial and nancial customers. The results, reported in Figure A.5 in the Internet Appendix, suggest a
very similar and qualitatively identical pattern to the one obtained for the CFTC data.
28
We nd that currencies for which volatility insurance is relatively cheap predictably appre-
ciate, while currencies for which volatility hedging using options is relatively more expensive
predictably depreciate. Standard risk factors cannot price the returns from the long-short
portfolio that we construct from these components. We consider two candidate explanations
for these ndings, and provide suggestive evidence that they can be rationalized in terms of the
time-variation of limits to arbitrage capital and the incentives of hedgers and speculators in
currency markets. Overall, the results in our paper provide new insights into the predictabil-
ity of exchange rate returns, an area in which evidence has been dicult to obtain. We also
introduce a new currency strategy with useful diversication properties into the expanding
and important research area on this topic.
29
References
Acharya, V., L.A. Lochstoer, and T. Ramadorai (2013). Limits to Arbitrage and Hedging:
Evidence from Commodity Markets, Journal of Financial Economics, forthcoming.
Adrian, T., E. Etula, and T. Muir (2013). Financial Intermediaries and the Cross-Section of
Asset Returns, Journal of Finance, forthcoming.
Agarwal, V., and N.Y. Naik (2004). Risks and Portfolio Decisions Involving Hedge Funds,
Review of Financial Studies 17, 6398.
Akram, Q.F., D. Rime, and L. Sarno (2008). Arbitrage in the Foreign Exchange Market:
Turning on the Microscope, Journal of International Economics 76, 237253.
Andrews, D.W.K. (1991). Heteroskedasticity and Autocorrelation Consistent Covariance Ma-
trix Estimation, Econometrica 59, 817858.
Ang, A., and J. Chen (2010). Yield Curve Predictors of Foreign Exchange Returns, Working
Paper, Columbia Business School.
Bank for International Settlements (2013). Triennial Central Bank Survey of Foreign Exchange
and Derivatives Market in 2013. Basel: Bank for International Settlements Press.
Barroso, P., and P. Santa-Clara (2013). Beyond the Carry Trade: Optimal Currency Portfo-
lios, Journal of Financial and Quantitative Analysis, forthcoming.
Bates, D.S. (1991). The Crash of 87: Was it Expected? The Evidence from Options Mar-
kets, Journal of Finance 46, 10091044.
Black, F., and M. Scholes (1973). The Pricing of Options and Corporate Liabilities, Journal
of Political Economy 81 637654.
Bollen, N.P., and Whaley, R.E. (2009). Hedge Fund Risk Dynamics: Implications for Perfor-
mance Appraisal, Journal of Finance 64, 9871037.
Bollerslev, T., G. Tauchen, and H. Zhou (2009). Expected Stock Returns and Variance Risk
Premia, Review of Financial Studies 22, 44634492.
30
Breeden, D., and R. Litzenberger (1978). Prices of State-Contingent Claims Implicit in Op-
tion Prices, Journal of Business 51, 621651.
Britten-Jones, M., and A. Neuberger (2000). Option Prices, Implied Price Processes, and
Stochastic Volatility, Journal of Finance 55, 839866.
Broadie, M., and A. Jain (2008). Pricing and Hedging Volatility Derivatives, Journal of
Derivatives 15, 724.
Brockhaus, O., and D. Long (2000). Volatility Swaps Made Simple, Risk 19, 9295.
Buraschi, A., F. Trojani, and A. Vedolin (2013). When Uncertainty Blows in the Orchard:
Comovement and Equilibrium Volatility Risk Premia, Journal of Finance, forthcoming.
Burnside, C. (2011). The Cross Section of Foreign Currency Risk Premia and Consumption
Growth Risk: Comment, American Economic Review 101, 345676.
Burnside, C., M. Eichenbaum, I. Kleshchelski, and S. Rebelo (2011). Do Peso Problems
Explain the Returns to the Carry Trade? Review of Financial Studies 24, 853891.
Campa, J.M., P.H.K. Chang, and R.L. Reider (1998). Implied Exchange Rate Distributions:
Evidence from OTC Option Markets, Journal of International Money and Finance 17, 117
160.
Carr, P., and L. Wu (2009). Variance Risk Premiums, Review of Financial Studies 22,
13111341.
Cochrane, J.H. (2005). Asset Pricing. Princeton: Princeton University Press.
Della Corte, P., L. Sarno, and I. Tsiakas (2011). Spot and Forward Volatility in Foreign
Exchange, Journal of Financial Economics 100, 496513
Demeter, K., E. Derman, M. Kamal, and J. Zou (1999). A Guide to Volatility and Variance
Swaps, Journal of Derivatives 6, 932.
Drechsler, I. and A. Yaron (2011). Whats Vol Got to Do with It, Review of Financial
Studies 24, 145.
31
Engel, C., N.C. Mark, and K.D. West (2008). Exchange Rate Models are not as Bad as you
Think, NBER Macroeconomics Annual 2007, 381441.
Eraker, B. (2010). The Volatility Premium, Working Paper, Wisconsin School of Business.
Evans, M.D.D., and R.K. Lyons (2005). Meese-Rogo Redux: Micro-Based Exchange Rate
Forecasting, American Economic Review Papers and Proceedings 95, 405414.
Fama, E.F., and J.D. MacBeth (1973). Risk, Return, and Equilibrium: Empirical Tests,
Journal of Political Economy 81, 60736.
Froot, K.A., and T. Ramadorai (2005). Currency Returns, Intrinsic Value, and Institutional
Investor Flows, Journal of Finance 60, 15351566.
Fung, W., and D.A. Hsieh (1997). Empirical Characteristics of Dynamic Trading Strategies:
the Case of Hedge Funds, Review of Financial Studies 10, 275302.
Grleanu, N. B., and L.H. Pedersen (2011). Margin-Based Asset Pricing and the Law of One
Price, Review of Financial Studes 24, 19802022.
Grleanu, N. B., L.H. Pedersen, and A. Poteshman (2009). Demand-Based Option Pricing,
Review of Financial Studies 22, 42594299.
Garman, M.B., and S.W. Kohlhagen (1983). Foreign Currency Option Values, Journal of
International Money and Finance 2, 231237.
Gromb, D., and D. Vayanos (2010). Limits of Arbitrage: The State of the Theory, Annual
Review of Financial Economics 2, 251275.
Han B., and Zhou Y. (2011). Variance Risk Premium and Cross-Section of Stock Returns,
Working Paper, University of Texas at Austin.
Hansen L.P. (1982). Large Sample Properties of Generalized Method of Moments Estima-
tors, Econometrica 50, 102954.
Hansen, L.P., and R. Jagannathan (1997). Assessing Specication Errors in Stochastic Dis-
count Factor Model, Journal of Finance 52, 557590.
32
Jiang, G.J., and Y.S. Tian (2005). The Model-Free Implied Volatility and Its Information
Content, Review of Financial Studies 18, 13051342.
Klitgaard, T., and L. Weir (2004), Exchange Rate Changes and Net Positions of Speculators
in the Futures Market, Federal Reserve Bank of New York Economic Policy Review 10, 1728.
Knauf, S. (2003). Making Money from FX Volatility, Quantitative Finance 3, 4851.
Lewellen, J., S. Nagel, and J. Shanken (2010). A Skeptical Appraisal of Asset Pricing Tests,
Journal of Financial Economics 96, 175194.
Londono, J.M., and H. Zhou (2012). Variance Risk Premiums and the Forward Premium
Puzzle, Working Paper, Federal Reserve Board.
Lustig, H., N. Roussanov, and A. Verdelhan (2011). Common Risk Factors in Currency
Markets, Review of Financial Studies 24, 37313777.
Lustig, H., and A. Verdelhan (2007). The Cross Section of Foreign Currency Risk Premia
and US Consumption Growth Risk, American Economic Review 97, 89117.
Martin, I. (2012). Simple Variance Swaps, Working Paper, Stanford University.
Meese, R.A., and K. Rogo (1983). Empirical Exchange Rate Models of the Seventies: Do
They Fit Out of Sample? Journal of International Economics 14, 324.
Menkho, L., L. Sarno, M. Schmeling, and A. Schrimpf (2012a). Carry Trades and Global
FX Volatility, Journal of Finance 67, 681718.
Menkho, L., L. Sarno, M. Schmeling, and A. Schrimpf (2012b). Currency Momentum Strate-
gies, Journal of Financial Economics 106, 620684.
Mueller, P., Y.M. Yen, and A. Vedolin (2011). Bond Variance Risk Premiums, Working
Paper, London School of Economics.
Nagel, S. (2012), Evaporating Liquidity, Review of Financial Studies 25, 20052039.
Neuberger, A. (1994). The Log Contract: A New Instrument to Hedge Volatility, Journal
of Portfolio Management 20, 7480.
33
Newey, W.K., and K.D. West (1987). A Simple, Positive Semi-Denite, Heteroskedasticity
and Autocorrelation Consistent Covariance Matrix, Econometrica 55, 703708.
Patton, A., and T. Ramadorai (2013). On the High-frequency Dynamics of Hedge Fund Risk
Exposures, Journal of Finance 68, 597635.
Ramadorai, T. (2013). Capacity Constraints, Investor Information, and Hedge Fund Re-
turns, Journal of Financial Economics 107, 401416.
Rime, D., L. Sarno and E. Sojli (2010). Exchange Rates, Order Flow and Macroeconomic
Information, Journal of International Economics 80, 7288.
Todorov, V. (2010). Variance Risk Premium Dynamics: The Role of Jumps, Review of
Financial Studies 23, 345383.
34
Table 1. Volatility Risk Premia
This table presents summary statistics for the 1-year volatility risk premia (\ 11) dened as dierence
between the realized volatility (1\ ) and the synthetic volatility swap rate (o\). 1\ is computed at time
t using daily exchange rate returns over the previous year. o\ is constructed at time t using the implied
volatilities across 5 dierent deltas from 1-year currency options. Q

refers to the ,
||
percentile. C indicates
the 1-year autocorrelation coecient. \ 11, 1\ , and o\ are expressed in percentage per annum. The sample
period comprises daily data from January 1996 to August 2011. Exchange rates are from Datastream whereas
implied volatility quotes are proprietary data from JP Morgan.
\ 11 1\ o\ \ 11 1\ o\
Developed Developed & Emerging
`cc: 0.62 10.68 11.31 0.92 10.82 11.74
odc 1.58 2.88 2.75 1.78 3.10 3.22
o/cn 0.54 1.85 1.42 0.31 2.12 2.07
1n:t 5.97 6.86 5.29 7.88 7.85 8.06
Q
5
3.06 7.15 7.77 3.67 7.23 8.36
Q
95
1.65 18.40 16.76 1.57 19.43 17.86
C 0.19 0.33 0.53 0.17 0.27 0.46
35
Table 2. Volatility Risk Premia Portfolios
This table presents descriptive statistics of ve currency portfolios sorted on the 1-year volatility risk
premia at time t 1. The long (short) portfolio 1
J
(1
S
) contains the top 20% of all currencies with the
highest (lowest) volatility risk premia. H,1 denotes a long-short strategy that buys 1
J
and sells 1
S
. The
table also reports the rst order autocorrelation coecient (C
1
), the annualized Sharpe ratio (o1), and the
frequency of portfolio switches (1rc). Panel A displays the overall excess return, whereas Panel B reports
the exchange rate component only. Panel C presents the transition probability from portfolio i to portfolio ,
between time t and time t + 1. : indicates the steady state probability. Returns are expressed in percentage
per annum. The strategies are rebalanced monthly from January 1996 to August 2011. Exchange rates are
from Datastream whereas implied volatility quotes are proprietary data from JP Morgan.
Panel A: Excess Returns
1
J
1
2
1
3
1
4
1
S
H,1 1
J
1
2
1
3
1
4
1
S
H,1
Developed Developed & Emerging
'ca: 4.70 2.24 1.04 1.78 0.67 4.03 3.59 1.93 1.34 1.40 1.26 2.34
odc 9.08 9.27 9.76 10.07 9.72 8.33 9.32 8.68 8.89 10.44 8.81 8.18
o/cn 0.05 0.19 0.09 0.17 0.26 0.28 0.09 0.05 0.21 0.29 0.39 0.12
1nrt 3.13 5.14 5.80 3.85 3.82 3.47 3.09 4.79 3.85 4.16 3.73 3.26
o1 0.52 0.24 0.11 0.18 0.07 0.48 0.39 0.22 0.15 0.13 0.14 0.29
C
1
0.10 0.04 0.13 0.15 0.01 0.04 0.10 0.14 0.15 0.13 0.11 0.05
1rc 0.24 0.44 0.52 0.48 0.32 0.32 0.26 0.43 0.53 0.48 0.27 0.27
Panel B: FX Returns
'ca: 4.93 2.06 1.26 1.60 0.52 4.40 3.51 1.62 1.37 0.82 0.21 3.72
odc 9.05 9.24 9.63 9.96 9.64 8.35 9.26 8.62 8.74 10.31 8.75 8.17
o/cn 0.12 0.15 0.06 0.18 0.26 0.28 0.18 0.00 0.26 0.31 0.47 0.12
1nrt 3.17 5.24 5.88 4.06 3.83 3.61 3.07 4.80 4.02 4.36 3.94 3.50
o1 0.54 0.22 0.13 0.16 0.05 0.53 0.38 0.19 0.16 0.08 0.02 0.46
C
1
0.10 0.03 0.11 0.13 0.01 0.04 0.10 0.13 0.13 0.10 0.10 0.04
1rc 0.24 0.44 0.52 0.48 0.32 0.32 0.26 0.43 0.53 0.48 0.27 0.27
Panel C: Transition Matrix
1
J
0.77 0.18 0.03 0.01 0.01 0.75 0.20 0.03 0.01 0.01
1
2
0.17 0.56 0.20 0.06 0.02 0.16 0.57 0.20 0.05 0.02
1
3
0.03 0.20 0.49 0.20 0.08 0.03 0.22 0.48 0.22 0.05
1
4
0.01 0.05 0.21 0.52 0.21 0.01 0.08 0.23 0.52 0.16
1
S
0.00 0.02 0.08 0.21 0.69 0.01 0.02 0.05 0.19 0.73
0.19 0.20 0.20 0.20 0.20 0.19 0.23 0.20 0.19 0.18
36
Table 3. Currency Strategies
This table presents descriptive statistics of currency strategies formed using time t 1 information. C1
is the carry trade strategy that buys (sells) the top 20% of all currencies with the highest (lowest) interest
rate dierential relative to the US dollar. Similarly, 'O' is the momentum strategy that buys (sells)
currencies with the highest (lowest) past 3-month exchange rate return, \ 1 is the value strategy that buys
(sells) currencies with lowest (highest) real exchange rate, 11 is the risk reversal strategy that buys (sells)
currencies with the lowest (highest) 1-year 10-delta risk reversal, and \ 11 is the volatility risk premium
strategy that buys (sells) currencies with the highest (lowest) 1-year volatility risk premium. The table also
reports the rst order autocorrelation coecient (C
1
), the annualized Sharpe ratio (o1), the Sortino ratio
(oO), the maximum drawdown ('11), and the frequency of portfolio switches for the long (1rc
J
) and
the short (1rc
S
) position. Panel A displays the overall currency excess return whereas Panel B reports the
exchange rate return component only. Panel C presents the sample correlations of the currency excess returns.
Returns are expressed in percentage per annum. The strategies are rebalanced monthly from January 1996
to August 2011. Exchange rates are from Datastream whereas implied volatility quotes are proprietary data
from JP Morgan.
Panel A: Excess Returns
C1 'O' \ 1 11 \ 11 C1 'O' \ 1 11 \ 11
Developed Developed & Emerging
'ca: 6.49 2.58 5.78 5.30 4.03 7.42 2.22 3.55 5.38 2.34
odc 10.66 9.55 9.38 11.40 8.33 9.97 8.30 8.90 10.60 8.18
o/cn 0.92 0.35 0.26 0.72 0.28 0.92 0.03 0.15 0.14 0.12
1nrt 5.65 3.86 3.50 6.58 3.47 4.53 2.95 3.17 4.43 3.26
o1 0.61 0.27 0.62 0.46 0.48 0.74 0.27 0.40 0.51 0.29
oO 0.72 0.50 0.94 0.58 0.87 0.94 0.47 0.62 0.75 0.49
'11 0.37 0.16 0.14 0.37 0.18 0.21 0.13 0.14 0.24 0.18
C
1
0.09 0.00 0.03 0.07 0.04 0.01 0.09 0.01 0.08 0.05
1rc
J
0.13 0.48 0.09 0.17 0.24 0.15 0.49 0.07 0.22 0.26
1rc
S
0.07 0.43 0.07 0.27 0.32 0.16 0.46 0.06 0.26 0.27
Panel B: FX Returns
'ca: 0.34 2.03 2.95 1.42 4.40 0.65 1.45 0.06 0.22 3.72
odc 10.66 9.57 9.44 11.48 8.35 9.99 8.16 8.89 10.60 8.17
o/cn 0.93 0.42 0.29 0.75 0.28 1.05 0.02 0.16 0.21 0.12
1nrt 5.82 4.17 3.51 6.83 3.61 4.84 3.13 3.19 4.74 3.50
o1 0.03 0.21 0.31 0.12 0.53 0.07 0.18 0.01 0.02 0.46
oO 0.04 0.40 0.47 0.15 0.93 0.08 0.30 0.01 0.03 0.75
'11 0.43 0.20 0.24 0.40 0.19 0.35 0.15 0.27 0.29 0.18
C
1
0.11 0.00 0.02 0.08 0.04 0.03 0.12 0.01 0.08 0.04
1rc
J
0.13 0.48 0.09 0.17 0.24 0.15 0.49 0.07 0.22 0.26
1rc
S
0.07 0.43 0.07 0.27 0.32 0.16 0.46 0.06 0.26 0.27
Panel C: Correlations
C1 1.00 0.17 0.44 0.68 0.18 1.00 0.03 0.54 0.57 0.21
'O' 0.17 1.00 0.17 0.17 0.09 0.03 1.00 0.14 0.15 0.10
\ 1 0.44 0.17 1.00 0.49 0.23 0.54 0.14 1.00 0.64 0.10
11 0.68 0.17 0.49 1.00 0.01 0.57 0.15 0.64 1.00 0.12
\ 11 0.18 0.09 0.23 0.01 1.00 0.21 0.10 0.10 0.12 1.00
37
Table 4. Currency Strategies: Sub-Samples
This table presents descriptive statistics of foreign exchange (FX) returns to currency strategies formed
using time t 1 information. C1 is the carry trade strategy that buys (sells) the top 20% of all currencies
with the highest (lowest) interest rate dierential relative to the US dollar. Similarly, 'O' is the momentum
strategy that buys (sells) currencies with the highest (lowest) past 3-month exchange rate return, \ 1 is the
value strategy that buys (sells) currencies with lowest (highest) real exchange rate, 11 is the risk reversal
strategy that buys (sells) currencies with the lowest (highest) 1-year 10-delta risk reversal, and \ 11 is the
volatility risk premium strategy that buys (sells) currencies with the highest (lowest) 1-year volatility risk
premium. Returns are expressed in percentage per annum. The strategies are rebalanced monthly from
March 2001 to November 2001, and from December 2007 to June 2009 (Panel A), from January 1996 to
December 2006 (Panel C), and from January 2007 to August 2011 (Panel D). January 1996 to August 2011.
Exchange rates are from Datastream whereas implied volatility quotes are proprietary data from JP Morgan.
Panel A: NBER Recession Periods
C1 'O' \ 1 11 \ 11 C1 'O' \ 1 11 \ 11
Developed Developed & Emerging
'ca: 9.59 11.32 4.62 7.96 11.54 7.97 7.07 0.10 4.80 6.50
odc 17.11 15.40 12.03 19.07 10.11 14.69 10.49 9.92 15.20 9.38
o/cn 0.44 0.28 0.63 0.90 0.12 0.80 0.17 0.15 0.08 0.45
1nrt 3.71 2.87 3.43 4.13 2.26 2.84 2.77 2.95 2.54 2.88
o1 0.56 0.74 0.38 0.42 1.14 0.54 0.67 0.01 0.32 0.69
'11 0.40 0.16 0.12 0.41 0.09 0.32 0.07 0.18 0.29 0.09
C
1
0.35 0.12 0.09 0.23 0.27 0.17 0.04 0.09 0.31 0.22
Panel B: non-NBER Recession Periods
'ca: 2.09 0.40 2.65 3.08 3.14 0.64 0.46 0.05 1.11 3.23
odc 9.06 8.11 8.95 9.57 7.99 8.92 7.68 8.73 9.61 7.96
o/cn 0.87 0.04 0.16 0.11 0.26 0.92 0.19 0.16 0.17 0.26
1nrt 4.50 2.48 3.30 4.16 4.02 4.90 2.90 3.22 5.55 3.70
o1 0.23 0.05 0.30 0.32 0.39 0.07 0.06 0.01 0.12 0.41
'11 0.31 0.21 0.22 0.15 0.16 0.31 0.20 0.22 0.20 0.16
C
1
0.07 0.09 0.02 0.04 0.03 0.06 0.15 0.00 0.02 0.02
Panel C: Pre-Crisis Period
'ca: 1.91 0.81 3.00 2.94 2.18 1.09 0.71 0.58 1.28 3.04
odc 8.33 7.90 9.78 9.43 7.99 9.16 7.68 9.25 10.12 8.53
o/cn 0.91 0.02 0.31 0.32 0.07 1.06 0.01 0.25 0.24 0.19
1nrt 4.92 2.46 3.26 4.14 3.46 5.20 2.59 3.10 5.41 3.47
o1 0.23 0.10 0.31 0.31 0.27 0.12 0.09 0.06 0.13 0.36
'11 0.31 0.16 0.24 0.15 0.19 0.31 0.14 0.23 0.18 0.18
C
1
0.05 0.11 0.03 0.02 0.01 0.08 0.14 0.02 0.03 0.02
Panel D: Crisis Period
'ca: 3.34 4.88 2.81 2.13 9.61 4.73 3.17 1.15 2.25 5.30
odc 14.80 12.69 8.67 15.31 9.05 11.70 9.23 8.05 11.72 7.29
o/cn 0.66 0.50 0.22 1.13 0.54 0.89 0.10 0.12 0.12 0.07
1nrt 4.02 3.57 4.27 5.63 3.42 3.90 3.56 3.41 3.67 3.39
o1 0.23 0.38 0.32 0.14 1.06 0.40 0.34 0.14 0.19 0.73
'11 0.43 0.16 0.12 0.40 0.08 0.31 0.13 0.15 0.29 0.10
C
1
0.22 0.09 0.01 0.18 0.09 0.17 0.10 0.12 0.25 0.10
38
Table 5. Exchange Rate Returns and Risk Factors
This table presents time-series regression estimates. The dependent variable is the volatility risk premium
strategy (\ 11) that buys (sells) currencies with the highest (lowest) 1-year volatility risk premium. As
explanatory variables, we use the currency strategies decribed in Table 3 in Panel A, the Fama and French
(1992) and the equity momentum factors in Panel B, and the Fung and Hsieh (2001) factors in Panel C. Newey
and West (1987) with Andrews (1991) optimal lag selection are reported in parenthesis. The superscripts a, /,
and c indicate statistical signicance at 10%, 5%, and 1%, respectively. Returns are annualized. The strategies
are rebalanced monthly from January 1996 to August 2011. Exchange rates are from Datastream whereas
implied volatility quotes are proprietary data from JP Morgan. Fama and French (1992) factors are from
Frenchs website whereas the Fung and Hsieh (2001) are from Hsiehs website.
Panel A: Currency Factors
c 1O1 C1 'O' \ 1 11 1
2
Developed
0.05
b
0.14 0.22
b
0.11 0.10 0.04 0.05
(0.02) (0.09) (0.09) (0.08) (0.13) (0.12)
Developed & Emerging
0.04
o
0.04 0.31
c
0.09 0.32
b
0.08 0.15
(0.02) (0.07) (0.09) (0.08) (0.11) (0.09)
Panel B: Equity Factors
c 1
t
n
o'1 H'1 'O'1 1
2
Developed
0.05
b
0.07 0.05 0.09
o
0.05 0.01
(0.02) (0.06) (0.05) (0.05) (0.03)
Developed & Emerging
0.05
b
0.07
o
0.10
o
0.10
b
0.05
o
0.03
(0.02) (0.04) (0.05) (0.05) (0.03)
Panel C: Hedge Fund Factors
c 1o:d Cnrr Co:: 1nitj oi.c 1o:d Crcdit
Trc:d Trc:d Trc:d 'ar/ct ojrcad 'ar/ct ojrcad 1
2
Developed
0.05
b
0.14 0.17 0.09 0.04 0.05 0.09 0.07 0.01
(0.02) (0.12) (0.11) (0.17) (0.05) (0.05) (0.11) (0.21)
Developed & Emerging
0.04
b
0.35 0.03 0.08 0.02 0.10
b
0.16
b
0.07 0.06
(0.02) (0.1) (0.13) (0.16) (0.04) (0.05) (0.07) (0.10)
39
Table 6. Asset Pricing Tests
This table reports asset pricing results. In Panel A the linear factor model includes the dollar (1O1) and the carry trade (C1) factors. In Panel B
the linear factor model includes the dollar (1O1) and the innovations to the global FX volatility (\ O1
J
) factors. C1 is a long-short strategy that
buys (sells) the top 20% of all currencies currencies with the highest (lowest) interest rate dierential relative to the US dollar. 1O1 is equivalent to a
strategy that borrows in the US money market and equally invests in foreign currenies, and serves as a constant in the cross-section. The test assets are
excess returns to ve portfolios sorted on the 1-year volatility risk premia (\ 11) available at time t 1. Factor Prices reports GMM and Fama-MacBeth
(FMB) estimates of the factor loadings /, the market price of risk `. The
2
and the Hansen-Jagannathan distance are test statistics for the null hypothesis
that all pricing errors are jointly zero. Factor Betas reports least-squares estimates of time series regressions. The
2
(c) test statistic tests the null that
all intercepts are jointly zero. Newey and West (1987) with Andrews (1991) optimal lag selection are reported in parenthesis. sh denotes Shanken (1992)
standard errors. The p-values are reported in brackets. Returns are annualized. The portfolios are rebalanced monthly from January 1996 to August 2011.
Exchange rates are from Datastream whereas implied volatility quotes are proprietary data from JP Morgan.
Panel A: Carry Trade Factor
Factor Prices
/
DOL
/
CAR
`
DOL
`
CAR
1
2
1Ao1
2
1J /
DOL
/
CAR
`
DOL
`
CAR
1
2
1Ao1
2
1J
Developed Developed & Emerging
GAA
1
0.42 0.47 0.02 0.05 0.07 3.29 4.35 0.16 0.24 0.01 0.02 0.01 0.14 2.09 1.93 0.11
(0.36) (0.55) (0.02) (0.07) [0.23] [0.20] (0.35) (0.51) (0.02) (0.06) [0.59] [0.59]
GAA
2
0.35 0.37 0.02 0.03 0.13 3.32 4.30 0.24 0.09 0.02 0.02 0.13 2.10 1.90
(0.36) (0.54) (0.02) (0.07) [0.23] (0.35) (0.50) (0.02) (0.06) [0.59]
1A1 0.42 0.47 0.02 0.05 0.07 3.29 4.35 0.24 0.01 0.02 0.01 0.14 2.09 1.93
(0.37) (0.58) (0.02) (0.07) [0.23] (0.34) (0.52) (0.02) (0.06) [0.59]
(cI) (0.34) (0.61) (0.02) (0.08) [0.18] (0.30) (0.53) (0.02) (0.06) [0.56]
Factor Betas
c ,
DOL
,
CAR
1
2

2
(c) c ,
DOL
,
CAR
1
2

2
(c)
1
L
0.03 0.89 0.04 0.62 8.75 0.02 0.96 0.02 0.69 3.10
(0.02) (0.06) (0.07) [0.12] (0.01) (0.04) (0.05) [0.68]
1
2
0.01 0.94 0.04 0.71 0.01 0.96 0.02 0.79
(0.01) (0.08) (0.05) (0.01) (0.04) (0.04)
1
3
0.01 1.00 0.05 0.72 0.01 1.00 0.08 0.80
(0.01) (0.05) (0.05) (0.01) (0.04) (0.04)
1
4
0.01 1.15 0.15 0.81 0.01 1.20 0.09 0.84
(0.01) (0.05) (0.05) (0.01) (0.05) (0.06)
1
S
0.02 1.03 0.08 0.79 0.02 0.87 0.17 0.75
(0.01) (0.04) (0.05) (0.01) (0.05) (0.05)
(continued)
4
0
Table 6. Asset Pricing Tests (continued)
Panel B: Global Volatility Factor
Factor Prices
/
1OJ
/
\ OJ
FX
`
1OJ
`
\ OJ
FX
1
2
1'o1
2
HJ /
1OJ
/
\ OJ
FX
`
1OJ
`
\ OJ
FX
1
2
1'o1
2
HJ
Developed Developed & Emerging
G''
1
0.52 1.25 0.02 0.16 0.26 2.74 3.02 0.13 0.48 0.55 0.02 0.08 0.07 2.04 2.31 0.11
(0.32) (0.81) (0.02) (0.11) [0.39] [0.44] (0.72) (1.49) (0.02) (0.22) [0.51] [0.55]
G''
2
0.44 1.01 0.02 0.15 0.27 2.75 2.91 0.30 0.13 0.02 0.02 0.14 2.06 2.22
(0.32) (0.78) (0.02) (0.11) [0.41] (0.7) (1.43) (0.02) (0.22) [0.53]
1'1 0.52 1.24 0.02 0.16 0.26 2.74 3.02 0.48 0.55 0.02 0.08 0.07 2.04 2.31
(0.36) (0.81) (0.02) (0.11) [0.39] (0.66) (1.34) (0.02) (0.22) [0.51]
(:/) (0.33) (0.92) (0.02) (0.13) [0.37] (0.71) (1.49) (0.02) (0.25) [0.54]
Factor Betas
c ,
1OJ
,
\ OJ
FX
1
2

2
(c) c ,
1OJ
,
\ OJ
FX
1
2

2
(c)
1
J
0.03 0.90 0.08 0.62 10.01 0.02 0.97 0.02 0.69 2.94
(0.01) (0.06) (0.06) [0.07] (0.01) (0.05) (0.03) [0.71]
1
2
0.00 0.94 0.07 0.71 0.00 0.95 0.02 0.79
(0.01) (0.08) (0.05) (0.01) (0.05) (0.03)
1
3
0.01 1.02 0.03 0.71 0.01 0.99 0.03 0.79
(0.01) (0.06) (0.07) (0.01) (0.04) (0.02)
1
4
0.01 1.10 0.01 0.78 0.01 1.19 0.02 0.83
(0.01) (0.06) (0.04) (0.01) (0.05) (0.04)
1
S
0.02 1.04 0.06 0.78 0.01 0.91 0.03 0.71
(0.01) (0.04) (0.04) (0.01) (0.05) (0.03)
4
1
Table 7. ,-Sorted Portfolios: Average Volatility Risk Premia
This table presents descriptive statistics of ,-sorted currency portfolios. Each , is obtained by regressing
individual currency excess returns on the average volatility risk premia using a 36-month moving window. The
long (short) portfolio 1
J
(1
S
) contains the top 20% of all currencies with the lowest (highest) ,. H,1 denotes a
long-short strategy that buys 1
J
and sells 1
S
. The table also reports the rst order autocorrelation coecient
(C
1
), the annualized Sharpe ratio (o1), and the frequency of portfolio switches (1rc). Panel A displays the
overall excess return, whereas Panel B reports the exchange rate component only. Panel C presents the pre-
and post-formation ,s, and the pre- and post-formation interest rate dierential (if) relative to the US dollar.
Standard deviations are reported in brackets whereas standard errors are reported in parentheses. Returns
are expressed in percentage per annum. The strategies are rebalanced monthly from January 1996 to August
2001. Exchange rates are from Datastream whereas implied volatility quotes are proprietary data from JP
Morgan.
Panel A: Excess Returns
1
L
1
2
1
3
1
4
1
S
11 1
L
1
2
1
3
1
4
1
S
11
Developed Developed & Emerging
Acon 5.54 1.70 3.46 2.06 6.76 1.23 4.16 2.22 3.33 3.34 5.43 1.27
odc 9.50 10.48 9.09 10.17 11.90 10.91 8.61 10.00 9.49 9.97 11.38 10.67
oIcn 0.27 0.05 0.52 0.04 0.36 0.80 0.04 0.38 0.29 0.25 0.67 1.14
1&vt 3.04 4.55 5.03 4.53 4.93 6.78 2.35 4.83 4.79 3.99 5.45 8.15
o1 0.58 0.16 0.38 0.20 0.57 0.11 0.48 0.22 0.35 0.33 0.48 0.12
oO 1.11 0.25 0.52 0.30 0.81 0.19 0.88 0.37 0.50 0.49 0.66 0.22
A11 0.19 0.27 0.31 0.30 0.27 0.35 0.19 0.27 0.32 0.27 0.27 0.35
C
1
0.03 0.01 0.19 0.12 0.10 0.03 0.04 0.05 0.18 0.11 0.12 0.01
1vcq 0.18 0.25 0.32 0.29 0.09 0.09 0.16 0.18 0.28 0.26 0.10 0.10
Panel B: FX Returns
Acon 6.39 1.91 3.07 1.18 4.69 1.70 5.10 2.35 2.78 1.56 3.21 1.90
odc 9.41 10.41 9.06 10.08 11.88 10.97 8.52 9.94 9.44 9.82 11.33 10.72
oIcn 0.30 0.04 0.56 0.07 0.38 0.87 0.06 0.37 0.33 0.31 0.76 1.30
1&vt 3.11 4.54 5.15 4.43 4.98 7.02 2.34 4.88 4.84 3.99 5.65 8.75
o1 0.68 0.18 0.34 0.12 0.39 0.15 0.60 0.24 0.29 0.16 0.28 0.18
oO 1.33 0.28 0.46 0.17 0.56 0.28 1.13 0.39 0.42 0.23 0.38 0.35
A11 0.16 0.25 0.32 0.32 0.29 0.32 0.16 0.25 0.33 0.29 0.29 0.22
C
1
0.02 0.01 0.19 0.12 0.10 0.05 0.04 0.04 0.18 0.09 0.11 0.02
1vcq 0.18 0.25 0.32 0.29 0.09 0.09 0.16 0.18 0.28 0.26 0.10 0.10
Panel C: Portfolio Formation
jvc-i) 0.85 0.21 0.39 0.88 2.08 0.94 0.13 0.55 1.78 2.22
joct-i) 0.85 0.19 0.41 0.90 2.09 0.97 0.10 0.56 1.79 2.24
jvc-, 0.35 0.14 0.13 0.35 0.60 0.42 0.17 0.12 0.40 0.81
[0.46] [0.50] [0.46] [0.32] [0.32] [0.71] [0.73] [0.61] [0.51] [0.56]
joct-, 0.26 0.29 0.15 0.06 0.11 0.26 0.22 0.09 0.14 0.08
(0.11) (0.10) (0.08) (0.08) (0.06) (0.09) (0.11) (0.08) (0.06) (0.07)
42
Table 8. Risk Factors: Liquidity and Hedging
This table presents predictive regressions estimates. The dependent variable is the exchange rate return component of the \ 11 strategy at time t.
This strategy is a long/short portfolio that buys (sells) the top 20% of all currencies with the highest (lowest) 1-year expected volatility premia at time
t 1. The predictors are measured at time t 1, and include the T11 spread, the change in the \ 1A index, the change in the [Link] Fed Financial
Stress Index 1o1, the net short futures position (H11) of commercial and non-commercial traders on the Australian dollar (AUD) and the Japanese
yen (JPY) vis-a-vis the US dollar (USD), respectively, and the 1n:d 1|on: constructed as the AUM-weighted net ows into hedge funds (currency and
global macro funds) scaled by the lagged AUM. T11, \ 1A, and 1o1 are averaged on a 12-month rolling window. H11 is winsorized at 99%. Newey
and West (1987) with Andrews (1991) optimal lag selection are reported in parenthesis. The superscripts a, /, and c indicate statistical signicance at
10%, 5%, and 1%, respectively. Exchange rate returns are annualized. Exchange rates are from Datastream, implied volatility quotes are from JP Morgan,
futures positions are from the US Commodity Futures Trading Commission (CFTC), hedge fund ows are from Patton and Ramadorai (2013), 1o1 is
from [Link] Feds website, whereas all other data are from Bloomberg.
TJ1 1J1 JunJ TJ1 1J1 JunJ
o TJ1 \ 1 JS1 \ 1 .I1IS1 1Y IS1 Jlous 1
2
o TJ1 \ 1 JS1 \ 1 .I1IS1 1Y IS1 Jlous 1
2
Developed Developed & Emerging
0.04 0.16
b
0.03 0.01 0.06 0.00
(0.03) (0.07) (0.03) (0.06)
0.04
b
0.05
a
0.01 0.04
a
0.04 0.01
(0.02) (0.03) (0.02) (0.03)
0.04
b
0.38
b
0.02 0.04
a
0.32
a
0.01
(0.02) (0.18) (0.02) (0.16)
0.03 0.09
c
0.05 0.03 0.06
c
0.02
(0.02) (0.02) (0.02) (0.02)
0.05
b
0.03
c
0.01 0.04
a
0.02
b
0.01
(0.02) (0.01) (0.02) (0.01)
0.05
b
0.02 0.00 0.04
a
0.01 0.01
(0.02) (0.06) (0.02) (0.06)
0.05
b
1.50
b
0.02 0.05
b
1.14
a
0.01
(0.02) (0.72) (0.02) (0.74)
0.01 0.12 0.02
b
0.93 0.04 0.03 0.03 0.02
a
0.93 0.01
(0.04) (0.07) (0.01) (0.73) (0.04) (0.07) (0.01) (0.75)
0.05
b
0.04 0.02
b
1.15
a
0.03 0.04
b
0.03 0.02
b
0.86 0.02
(0.02) (0.03) (0.01) (0.68) (0.02) (0.03) (0.01) (0.67)
0.05
b
0.31
a
0.02
b
1.15
a
0.04 0.04
b
0.26 0.02
b
0.83 0.02
(0.02) (0.18) (0.01) (0.68) (0.02) (0.17) (0.01) (0.69)
0.04
a
0.08
c
0.02
b
0.93 0.06 0.04 0.05
c
0.02
b
0.72 0.03
(0.02) (0.02) (0.01) (0.65) (0.02) (0.02) (0.01) (0.66)
4
3
Figure 1. Rolling Sharpe Ratios
The gure presents for developed countries the 1-year rolling Sharpe ratios of currency strategies formed using t 1 information. CAR is the carry strategy that buys
(sells) the top 20% of all currencies with the highest (lowest) interest rate dierential relative to the US dollar. Similarly, MOM is the momentum strategy that buys (sells)
currencies with the highest (lowest) past 3-month exchange rate return, V AL is the value strategy that buys (sells) currencies with lowest (highest) real exchange rate,
RR is the risk reversal strategy that buys (sells) currencies with the lowest (highest) 1-year 10-delta risk reversal, and V RP is the volatility risk premium strategy that
buys (sells) currencies with the highest (lowest) 1-year volatility risk premium. The strategies are rebalanced monthly from January 1996 to August 2011. Exchange rates
are from Datastream whereas implied volatility quotes are proprietary data from JP Morgan. Appendix A reports a similar gure for developed & emerging countries.
4
4
Figure 2. Currency Strategies and Payos
The gure presents for developed countries the cumulative wealth to currency strategies formed using t 1 information. The strategies are rebalanced monthly from
January 1996 to August 2011, and described in Figure 1. Exchange rates are from Datastream whereas implied volatility quotes are proprietary data from JP Morgan.
Appendix A reports a similar gure for developed & emerging countries.
4
5
Figure 3. Global Minimum Volatility Portfolios
The gure presents the global minimum volatility portfolio (MVP) and the ecient frontier (solid line) built using the ve currency strategies described in Figure 1.
The dashed line denotes the ecient frontier that excludes the volatility risk premium (VRP) strategy. The portfolio weights (N 1) are reported in parentheses and
computed as w = (
1
)/(

1
) where is the N N covariance matrix of the strategies returns, and is a N 1 vector of ones.
4
6
Figure 4. Reversal in the Volatility Risk Premium Strategy
This gure presents cumulative average returns to the volatility risk premium (V RP) strategy after portfolio formation. V RP buys (sells) the top 20% of all currencies
with the highest (lowest) 1-year volatility risk premia. Post-formation returns are constructed for 1, 2, . . . , 20 months following the formation period. This is equivalent to
building new portfolios every month and recording them for the subsequent 1, 2, . . . , 20 months (using overlapping horizons). We cumulate risk-adjusted (with respect to
the carry trade strategy) excess returns and exchange rate returns. The strategies are rebalanced monthly from January 1996 to August 2011. Exchange rates are from
Datastream whereas implied volatility quotes are proprietary data from JP Morgan.
4
7
Figure 5. Futures Positions and Volatility Risk Premium Strategy
The gure presents the net position on currency futures into the volatility risk premium (VRP) strategy. We rank currencies by volatility risk premia into four baskets at
time t, and then compute the average net position on futures at time t. Finally, we take the dierence between the rst (currencies with the cheapest volatility insurance)
and the last (currencies with the most expensive volatility insurance) portfolio. The net (long minus short) position on futures is standardized over a 3-month rolling
window. The commercial traders are dened as merchants, foreign brokers, clearing members or investment banks using the futures market primarily to hedge their
business activities whereas the nancial (or non-commercial) traders are individual investors, hedge funds, and some large nancial institutions using the futures market
for speculative purposes. The data runs from January 1996 to August 2011 at weekly frequency (collected on Tuesday). Exchange rates are from Datastream, implied
volatility quotes are proprietary data from JP Morgan, whereas futures positions are from the Commodity Futures Trading Commission (CFTC).
4
8