Forward and Spot Exchange Rates in a Multi-Currency World

Transcription

1 Forward and Spot Exchange Rates in a Multi-Currency World Tarek A. Hassan Rui C. Mano Very Preliminary and Incomplete Abstract We decompose the covariance of currency returns with forward premia into a crosscurrency, a between time and currency, and a cross-time component. The surprising result of our decomposition is that the cross-currency and cross-time-components account for almost all systematic variation in expected currency returns, while the between time and currency component is statistically and economically insignificant. This finding has three surprising implications for models of currency risk premia. First, it shows that the two most famous anomalies in international currency markets, the carry trade and the Forward Premium Puzzle FPP), are separate phenomena that may require separate explanations. The carry trade is driven by persistent differences in currency risk premia across countries, while the FPP appears to be driven primarily by time-series variation in all currency risk premia against the US dollar. Second, it shows that both the carry trade and the FPP are puzzles about asymmetries in the risk characteristics of countries. The carry trade results from persistent differences in the risk characteristics of individual countries; the FPP is best explained by time variation in the average return of all currencies against the US dollar. As a result, existing models in which two symmetric countries interact in financial markets cannot explain either of the two anomalies. JEL Classication: F31 Keywords: Risk Premia in Foreign Exchange Markets, Forward Premium Puzzle, Carry Trade We are grateful to Craig Burnside, John Cochrane, Jeremy Graveline, Ralph Koijen, and Adrien Verdelhan. We also thank seminar participants at the University of Chicago, CITE Chicago, the Chicago Junior Finance Conference, KU Leuven, University of Sydney, New York Federal Reserve and at the SED annual meetings for useful comments. All mistakes remain our own. University of Chicago, Booth School of Business, NBER and CEPR, 5807 South Woodlawn Avenue, Chicago IL USA; Tarek.Hassan@ChicagoBooth.edu. University of Chicago, Department of Economics 1

2 1 Introduction A large body of theoretical work focuses on the Forward Premium Puzzle FPP), a set of stylized facts about the covariance of bilateral currency risk premia and forward premia. It states that a bilateral regression of currency returns on forward premia often yields a slope coefficient β i ) larger than one: rx i,t+1 = α i + β i f it s it ) + ε i,t+1, 1) where f it is the log one-period forward rate of currency i, s it is the log spot rate and rx i,t+1 = f it s i,t+1. 1 The FPP has drawn a lot of interest by theorists because it has bewildering implications for the joint dynamics of currency risk premia, interest rates, and exchange rates. Fama 1984) shows that, without loss of generality, β i > 1 implies cov i π it, f it s it ) > var i f it s it ) cov i π it, E it Δs i,t+1 ) < var i E it Δs i,t+1 ), 2) where π it = E it rx i,t+1 ) is the conditional expected return risk-premium ) on currency i at time t and E it Δs i,t+1 is its expected rate of depreciation. 2 The FPP thus implies that bilateral currency risk premia must be highly volatile and negatively correlated with expected depreciations. This finding is usually interpreted to mean that i) high interest rate currencies appreciate; ii) the carry trade, a trading strategy which is long high interest rate currencies and short low interest rate currencies, is profitable due to the FPP; and iii) volatile risk premia must play a role in determining bilateral exchange rates. i)-iii) are natural interpretations because they are the only way for the inequalities in 2) to hold in a world in which currencies are ex-ante symmetric. In such a symmetric world, the α i in 1) are unimportant and the conditional covariances in 2) equal unconditional covariances. The main finding of our paper is that most of the systematic variation in currency returns appears to be asymmetric across currencies. As a result, the interpretations i)-iii) are misleading and ex-ante symmetric models of currency returns e.g. models that feature two identical countries) can explain only a small, statistically insignificant, part of the systematic variation in currency returns. We generalize the regression-based approach in 1) to study the covariance of currency risk premia with forward premia without conditioning on a specific currency pair i. We decompose the unconditional covariance between currency risk premia and forward premia into a cross-currency, 1 The same relationship is often estimated using the change in the spot exchange rate as the dependent variable, in which case the coefficient estimate is 1 β i. An equivalent way of stating the FPP is thus that 1 β i < 0. 2 Throughout the paper we follow the convention in the literature and refer to conditional expected returns as risk premia. However, this terminology need not be taken literally. Our analysis is silent on whether currency returns are driven by risk premia, institutional frictions or other limits to arbitrage. See Burnside et al. 2011) and Lustig et al. 2011) for a discussion. 2

3 a cross-time, and a between time and currency component. If currencies are ex-ante symmetric, the cross-currency component should be unimportant and all of the interesting variation should be in the between time-and currency dimension. The data show the opposite. Most of the systematic variation in currency returns is in the cross-section the α i in 1)). Currencies that have permanently higher forward premia consistently pay higher expected returns returns than currencies with permanently lower forward premia. This static, cross-currency, variation in currency risk-premia explains the majority of the returns to the carry trade. Some of our specifications also show statistically significant variation in the cross-time dimension. The expected return on the US dollar appears to fluctuate with its average forward premium against all other currencies in the sample. This time-series variation is particular to the US dollar and, potentially, a small number of other currencies. It explains the vast majority of the variation that generates the forward premium puzzle. In contrast, we cannot reject the null that currency risk premia do not fluctuate between time and currency. Once the average forward premium of all currencies against the US dollar is controlled for, there is little evidence of additional time-series variation in currency risk premia. These results imply that the traditional interpretation of the FPP is misleading. First, currencies which have high interest rates relative to other currencies tend to depreciate rather than appreciate. Second, the carry trade and the FPP are not significantly related in the data and may thus require distinct theoretical explanations. Explaining the carry trade primarily requires explaining permanent differences in interest rates across countries that are partially, but not fully, reversed by predictable movements in exchange rates. High interest rate currencies depreciate, but not enough to reverse the higher returns resulting from the interest rate differential.) In contrast, explaining the FPP requires explaining the time series variation in the risk-premium of the US dollar against all other currencies. The US dollar may be one of a small number of currencies that pays higher expected returns when its interest rate is high relative to its own currency-specific average and to the world average interest rate at the time. However, this relationship is only marginally statistically significant in the data. Third, we find no robust evidence of a positive covariance between currency risk premia and expected appreciations in any of the three dimensions. We view these results as both good and bad news. The good news is that currency risk premia may be much simpler than previously thought. First, the majority of the variation in currency risk premia is static across currencies. Second, we find no statistically reliable evidence that currency risk premia vary over time, except potentially for a covariance of the risk-premium on the US dollar with the average forward premium on all other currencies. Third, this covariance is not so large that it requires a systematic association between the risk premium on the US dollar and the average exchange rate of the US dollar against all other currencies. In fact, we 3

4 can never reject the hypothesis that currency risk premia are uncorrelated with expected changes in exchange rates, neither for the US dollar nor for any of the other currencies in our sample. The bad news is that the asymmetries driving the carry trade and potentially) the FPP are not well understood. Much of the existing theoretical literature focuses on ex-ante symmetric countries and and may thus have to be re-interpreted in the light of our evidence. Taken at face value, these models explain the small, statistically insignificant, variation in currency risk premia in the between time and currency dimension. 3 The relatively small number of papers offering theoretical explanations of asymmetries in currency risk premia include Hassan 2013), Martin 2012) and Govillot, Rey, and Gourinchas 2010) who focus on differences in country size and Maggiori 2013) and Caballero, Farhi, and Gourinchas 2008) who focus on differences in financial development. Our work builds heavily on a series of papers that apply factor analysis to study the crosssection of multi-lateral currency returns. Most closely related are Lustig, Roussanov, and Verdelhan 2010, 2011) who identify a risk-factor that explains the cross-section of currency returns and a dollar factor that explains the time series variation in the returns on the US dollar. Our main contribution is to re-cast these findings in terms of regression coefficients, relate them to established puzzles in the literature, and to translate them into restrictions on linear models of currency risk premia. Many authors have described and theorized about the carry trade and the FPP. 4 We contribute to this literature in three ways. First, we show that the carry trade and the FPP are distinct, quantitatively unrelated, anomalies in the data. Second, we generalize the empirical approach that has framed the debate on the FPP to a multi-currency framework. Third, we use this framework to derive restrictions on linear models of multilateral currency risk premia. The remainder of this paper is structured as follows. Section 2 describes the data. Section 3 establishes the FPP and the carry trade as separate anomalies. Section 4 discusses the restrictions that our empirical results pose on linear models of currency risk premia. Section 5 discusses implications for models of exchange rate determination. Section 6 concludes. 3 Examples include Farhi and Gabaix 2008), Verdelhan 2010), Burnside et al. 2009), Heyerdahl-Larsen 2012), Yu 2011), Bacchetta et al. 2010), and, Ilut 2012). 4 See for example Hansen and Hodrick 1980), Bilson 1981), Meese and Rogoff 1983), Fama 1984), Backus et al. 1993), Evans and Lewis 1995), Bekaert 1996), Bansal 1997), Bansal and Dahlquist 2000), Backus et al. 2001), Evans and Lyons 2006), Graveline 2006), Burnside et al. 2006), Lustig and Verdelhan 2007), Brunnermeier et al. 2009), Alvarez et al. 2008), Jurek 2009), Bansal and Shaliastovich 2010), Burnside et al. 2011), Colacito and Croce 2011), and, Sarno et al. 2012) Menkhoff et al. 2012). Engel 1996) and Lewis 2011) provide excellent surveys. 4

5 2 Data Throughout the main text we use monthly observations of US dollar-based spot and forward exchange rates at the one, six, and twelve month horizon. All rates are from Thomson Reuters Financial Datastream. The data range from October 1983 to June For robustness checks we also use all UK pound based data from the same source as well as forward premia calculated using covered interest parity from interest rate data which is available for longer time horizons for some currencies). To the best of our knowledge our dataset nests the data used in recent studies on the cross-section of currency returns, including Lustig et al. 2011) and Burnside et al. 2011). Many of the decompositions we perform below require balanced samples. However, currencies enter and exit the sample frequently, the most important example of which is the Euro and the currencies it replaced. We deal with this issue in two ways. In our baseline sample 1 Rebalance ) we use the largest fully balanced sample we can construct from our data by selecting the 15 currencies with the longest coverage the currencies of Australia, Canada, Denmark, Hong Kong, Japan, Kuwait, Malaysia, New Zealand, Norway, Saudi Arabia, Singapore, South Africa, Sweden, Switzerland and the UK from December 1994 to June 2010). In addition, we construct three alternative samples which allow for entry of currencies at 3, 6, and 12 dates during the sample period, where we chose the entry dates to maximize coverage. The 3 Rebalance sample allows entry in December of 1989, 1997, and 2004 and covers 30 currencies. The 6 Relabance sample allows entry in December of 1989, 1993, 1997, 2001, 2004 and 2007 and covers 36 currencies. Our largest, 12 Rebalance, sample allows entry in June 1986, and in June of every second year thereafter through June 2008 and covers 39 currencies. In between each of these dates all samples are balanced except for a small number of observations removed by our data cleaning procedure see appendix for details). Currencies enter each of the samples if their forward and spot exchange rate data is available for at least four years prior to the re-balancing date the reason for this prior data requirement will become apparent below). 5 Throughout the main text we take the perspective of a US investor and calculate all returns in US dollars. In section 4.4 we discuss how our results change when we use different base currencies. Appendix B lists the coverage of individual currencies and describes our data selection and cleaning process in detail. 5 The only exception we make to this rule is for the first set of currencies entering the 12 Rebalance sample which become available in October

6 3 FPP & Carry Trade as Separate Anomalies Consider a version of the carry trade in which, at the beginning of each month, t, we form a portfolio of all available currencies weighted by the difference of their forward premia fp it = f it s it ) to the average forward premium of all currencies at the time fp t ). This portfolio is long currencies that have a higher forward premium than the average of all currencies at time t and short currencies that have a lower than average forward premium. Without loss of generality, we can write the expected return on this portfolio as E [rx i,t+1 fp it fp t )], 3) where E is the unconditional expectations operator across currencies, i, and time periods, t. We use these linear portfolio weights, fp it fp t ), because they allow us to relate portfolio returns directly to coefficients in linear regressions. Note, however, that our results would be very similar if we sorted currencies into bins and then analyzed the returns on a long-short strategy as in Lustig et al. 2011) or if we analyzed the returns on an equally weighted strategy as in Burnside et al. 2011). As with these alternative formulations, the return to on the carry trade portfolio is neutral with respect to the dollar, i.e. it is independent of the average exchange rate of the US dollar against all other currencies. Table 1 shows the annualized mean return on the carry trade portfolio in our 1 Rebalance sample. As expected, the carry trade is highly profitable. It yields a mean annualized net return of 4.96% with a Sharpe Ratio of.54. Somewhat surprising, however, is that the currencies that the carry trade is long i.e. currencies with high interest rates) on average depreciate relative to currencies with low interest rates. Our carry trade portfolio loses 2.16 percentage points of annualized returns due to this depreciation. As we show below, this is a general feature of the carry trade that holds across a wide range of plausible variations. Table 1: Expected Returns to the Carry Trade E [rx i,t+1 fp it fp t )] 4.96 Forward Premium 7.12 Appreciation Sharpe Ratio 0.54 Note: Annualized returns to the carry trade calculated by dividing the expression in 3) with the unconditional mean forward premium in the sample, fp. 1 month forward and spot exchange rates from the 1 Rebalance sample ranging from12/1994 to 6/

7 Currencies with high interest rates thus generally tend to depreciate. An obvious question is then why the FPP appears to suggest the opposite. The answer is in the country-specific intercepts in 1). We tend to find that β i > 1 in bilateral currency-by-currency) regressions, in which the intercepts, α i, absorb the currency-specific mean forward premium fp i ). If we wanted to trade on the correlation in the data that drives the FPP, we would thus have to buy currencies that have a higher forward premium than they usually do Cochrane, 2001). Such a strategy, we might call it the Forward Premium Trade FPT), weights each currency with the deviation of its current forward premium from its currency-specific average. We can write the expected return on the FPT as E [rx i,t+1 fp it fp i )]. Figure 1: Carry Trade vs. Forward Premium Trade Left Panel: Carry Trade, based on fp it fp t, Right Panel: Forward Premium Trade, based on fp it fp i The carry trade 3) thus exploits a correlation between currency returns and forward premia conditional on time, while the FPP describes a correlation conditional on currency. Figure 1 illustrates the difference between the carry trade and the FPT for the case in which a US investor considers investing in two foreign currencies. The left panel plots the forward premium of the New Zealand dollar and the Japanese yen over time. Throughout the sample period the forward premium of the former is always higher than the forward premium of the latter, reflecting the fact that New Zealand has consistently higher interest rates than Japan. The carry trade is 7

8 always long New Zealand dollars and always short Japanese yen. In contrast, the FPT evaluates the forward premium of each currency in isolation and goes long if the forward premium is higher than its sample mean. As a result, the FPT is not dollar neutral in the sense that it may be long or short both foreign currencies at any given point in time. It is immediately apparent that the FPT may be more difficult to implement in practice than the carry trade as it requires an estimate of the true sample mean forward premium of each country fp i ), which may not be known at time t. In what follows we denote the unconditional ex-ante) expectation of the country-specific and the unconditional mean forward premium as ˆ fp i = E 0 [fp i ], ˆ fp = E0 [fp]. The ex-ante implementable version of the FPT which we show below is the version that is relevant for estimating covariances of risk premia and forward premia) is thus where ˆ fp i fp i and ˆ fp fp in a finite sample. E [rx i,t+1 fp it fp ˆ )] i, 4) How do the carry trade and the FPT relate to each other? The expected returns on both portfolios load on different components of the unconditional covariance between currency returns and forward premia. To see this we can decompose the unconditional covariance into the sum of the expected returns on three trading strategies plus a constant term. Re-writing the unconditional covariance in expectation form, adding and subtracting fp ˆ i and f ˆ p and re-arranging yields [ cov rx i,t+1, fp it ) = E [rx i,t+1 rx) fp it fp)] = E rx i,t+1 fpi ˆ fp ˆ )] + E [rx i,t+1 fp it fp t fpi ˆ ˆ ))] fp + E [rx i,t+1 fp t ˆ )] fp } {{ } } {{ } } {{ } Static Trade Dynamic Trade Dollar Trade +E [rx i,t+1 fp ˆ fp )], } {{ } Constant 5) where rx refers to the unconditional sample mean currency return across currencies and time periods. The Static Trade trades on the cross-currency variation of forward premia. It is long currencies that have an unconditionally high forward premium and short currencies that have an unconditionally low forward premium. We may think of it as a version of the carry trade in which we never update our portfolio. We weight currencies once, based on our expectation of the currencies future average level of interest rates and never change the portfolio thereafter. The Dynamic Trade trades on the between time and currency variation in forward premia. It 8

9 is long currencies that have high forward premia relative to the time average forward premium of all currencies and relative to their currency-specific mean forward premium. We may think of the expected return on the Dynamic Trade as the incremental benefit of re-weighing the carry trade portfolio every period. Finally, the Dollar Trade trades on the cross-time variation in the average forward premium of all currencies against the US dollar. It goes long all foreign currencies when the average forward premium of all currencies against the US dollar is high relative to its unconditional mean and goes short all foreign currencies when it is low. 6 Upon inspection, the carry trade 3) is simply the sum of the Static and Dynamic trades, [ E rx i,t+1 [fp it fp t ]) = E rx } {{ } i,t+1 fpi ˆ ˆ ]) [ fp + E rx i,t+1 fp it fp t fpi ˆ ˆ )]) fp } {{ } } {{ } Carry Trade Static Trade Dynamic Trade while the FPT 4) is the sum of the Dynamic and the Dollar Trades. [ E rx i,t+1 fp it fp ˆ ]) [ i = E rx i,t+1 fp it fp t fpi ˆ ˆ )]) [ fp + E rx i,t+1 fp t ˆ ]) fp } {{ } } {{ } } {{ } FP Trade Dynamic Trade Dollar Trade The common element between the Carry Trade and the FPP is the Dynamic Trade, i.e. the between time and currency part of the unconditional covariance between currency returns and forward premia. In contrast, the cross-currency component is unique to the carry trade and the cross-time component is unique to the FPP. The question of whether the two anomalies, the carry trade and the FPP, are related in the data thus reduces to estimating the relative contribution of the Dynamic Trade. Table 2 lists the mean returns and Sharpe ratios of the three strategies, as well as the mean returns and Sharpe ratios of the carry trade and the FPT. All returns are again annualized and normalized by dividing with f p to facilitate comparison. Columns 1-4 of Panel A give the results for our 1 Rebalance sample, where we use all available data prior to December 1994 to estimate fp ˆ i and f ˆ p. Column 1 shows the results for one-month forwards, without taking into account bid-ask spreads. The mean annualized return on the static trade is 3.46% with a Sharpe ratio of.39. It thus contributes 70% of carry trade returns. In contrast, the Dynamic Trade contributes 30% of carry trade returns, with an annualized return of 1.50% and a Sharpe ratio of.24. Although the FPT is not commonly known as a trading strategy in foreign exchange markets it yields similar returns to the carry trade, with an expected annualized return of 4.05% and a Sharpe ratio of.28. The Dollar Trade contributes 63% to this overall return and has a Sharpe ratio of.25, with the Dynamic Trade contributing the remaining 37%. Columns 2-4 replicate the same decomposition but take into account bid-ask spreads in for- 6 The Dollar Trade was first described by Lustig et al. 2010). We follow their naming convention here. 9

10 Table 2: Currency Portfolios 1) 2) 3) 4) 5) 6) 7) 8) Panel A 1 Rebalance 3 Rebalances Horizon months) Static T E[rx i,t fp i fp)] Sharpe Ratio Dynamic T E[rx i,t fp i,t fp t fp i fp))] Sharpe Ratio Dollar T E[rx i,t fp t fp)] Sharpe Ratio Carry Trade E[rx i,t fp i,t fp t )] Sharpe Ratio % E[rx] Static T 70% 120% 92% 76% 69%. 105% 85% Forward Premium Trade E[rx i,t fp i,t fp i )] Sharpe Ratio % E[rx] Dollar T 63% 123% 88% 73% 57%. 106% 84% Panel B 6 Rebalances 12 Rebalances Static T E[rx i,t fp i fp)] Sharpe Ratio Dynamic T E[rx i,t fp i,t fp t fp i fp))] Sharpe Ratio Dollar T E[rx i,t fp t fp)] Sharpe Ratio Carry Trade E[rx i,t fp i,t fp t )] Sharpe Ratio % E[rx] Static T 57%. 86% 104% 69%. 85% 93% FP Trade E[rx i,t fp i,t fp i )] Sharpe Ratio % E[rx] Dollar T 54%. 88% 102% 51%. 68% 91% Bid-Ask Spreads No Yes Yes Yes No Yes Yes Yes Note: Mean returns and Sharpe ratios on the Static, Dynamic, Dollar, Carry, and Forward Premium Trades defined in equations 3), 4), and 5) calculated using 1, 6 and 12 month currency forward contracts against the US dollar. All returns are annualized and normalized by dividing with f p to facilitate comparison. The table also reports the percentage contribution of Static Dollar) Trade to the mean returns on the carry Forward Premium) trade, calculated by dividing its mean return with the maximum of zero and the sum of the mean returns on the Static Dollar) and Dynamic Trades. See Data Appendix for details. 10

11 ward and spot exchange markets. 7 Column 2 again uses one-month forward contracts, column 3 uses 6-month contracts, and column 4 uses 12-month contracts. Once we take into account bid-ask spreads, the mean returns on all trading strategies fall. In the case of the Dynamic Trade the mean return in column 2 actually turns negative. 8 However, the same basic pattern persists across all columns: the Static Trade accounts for % of the mean returns on the carry trade returns and the Dollar Trade accounts for % of the mean returns on the FPT. 9 The only potentially sensitive assumption we make in performing this decomposition is that investors use data prior to 1995 to estimate fp ˆ i and fp. ˆ To show that there is nothing particular about this cutoff date and the resulting selection of currencies in our 1 Rebalance sample), the remaining panels and columns repeat the same exercise using the 3, 6, and 12 Rebalance samples. In each case we use all available data before each cutoff date to update the estimates of fpi ˆ and fp. ˆ In the 3 Rebalance sample, investors thus update their expectation of the two means at 3 dates, and so forth. The results remain broadly the same across the different samples, where the Static Trade on average contributes 88.02% of the mean returns to the carry trade and the Dollar Trade on average contributes 66.10% of the mean returns on the FPT. In addition, the Sharpe ratio on the Dynamic Trade appears economically small or even negative in all calculations that take into account the bid-ask spread they range from to 0.19). While the carry trade delivers an economically significant Sharpe ratio in all samples ranging from 0.19 to 0.44 net of transaction costs), the FPT tends to deliver somewhat lower Sharpe ratios ranging from 0.00 to 0.27), particularly in the samples that allow more rebalances. Our main conclusion from Table 2 is that the Dynamic Trade, the common element between the carry trade and the FPT, contributes an economically small share to the expected returns on the two strategies. The majority of the returns on the carry trade are driven by static differences in expected returns across currencies and the majority of the returns on the FPT are driven by time series variation in the expected returns on the US dollar relative to all other currencies in the sample. 7 We calculate returns net of transaction costs as rx net i,t+1 = I[w it 0]fit bid s ask i,t+1 ) + 1 I[w it > 0])fit ask s bid i,t+1 ), where w it is the portfolio weight of currency i at time t. and I is an indicator function that is one if w it 0 and zero otherwise. 8 Transaction costs in currency markets are thus of the same order of magnitude as the mean returns on the Dynamic Trade. See Burnside et al. 2006) for a discussion. 9 Note that the mean returns on the three underlying trades no longer add up to the mean returns on the carry trade and the FPT when we take into account bid-ask spreads. We thus calculate the percentage contribution of Static Dollar) Trade by dividing its mean return with the maximum of zero and the sum of the mean returns on the Static Dollar) and Dynamic Trades. 11

12 4 Restrictions on Models of Currency Risk Premia In a world with more than two currencies, currency risk premia may vary across currencies, between time and currency, and across time, where each of these dimensions corresponds to one of the three basic trading strategies outlined above. In order to attach some notion of statistical significance to the variation of risk premia in each of these dimensions it is useful to re-write 5) in terms of regression coefficients β stat var fpi ˆ fp ˆ ) } {{ } Static Trade cov rx i,t+1, fp it ) = + β dyn var fp i,t fp t fpi ˆ fp ˆ )) + α dyn } {{ } Dynamic Trade + β dol var fp t fp ˆ ) + α dol α dol. } {{ } Dollar Trade The parameters {β stat, β dyn, β dol } are slope coefficients from regressions of currency returns on the variation in forward premia in the relevant dimension. rx i rx t+1 = β stat fpi ˆ fp ˆ ) + ɛ stat i,t+1 7) rx i,t+1 rx t+1 rx i rx) = β [fp dyn it fp t ) fpi ˆ fp ˆ )] + ɛ dyn 6) i,t+1 8) rx t+1 rx = γ + β fp dol t fp ˆ ) + ɛ dol i,t+1 9) where rx t+1 is the mean return across all currencies at time t + 1. The two constants α dyn = E [rx i fp i fp fp ˆ i fp) ˆ )] and α dol = E [ rxfp fp) ˆ ] are finite-sample estimation errors that result from the fact that investors do not have an infinitely large sample to estimate the means fp i and fp. In contrast, the three slope coefficients determine the systematic part of the mean returns calculated in Table 2. parameters in any linear model of currency risk premia. In this sense, they are deep Proposition 1 The slope coefficients β stat, β dyn, and β dol measure the covariance of risk premia with forward premia in the cross-currency, the between time and currency, and the cross-time dimension, respectively. Proof. By the properties of linear regression, we can write β stat as [ β stat = E rx i,t+1 rx t+1 ) f ˆp i fp ˆ )] [ = E E it {rx i,t+1 rx t+1 )} fpi ˆ fp ˆ ) 1 [ var fpi ˆ = E E it {rx i,t+1 rx t+1 ) f ˆp i fp ˆ )}] )] ) 1 var fpi ˆ = cov π it, fp ˆ ) ) 1 i var fpi ˆ, ) 1 var fpi ˆ where E it ) is the rational expectations operator conditional on all information available about currency i at time t. The second equality applies the law of iterated expectations. The third equality uses the fact that fp ˆ i and fp ˆ are known at time t. The proofs for β dyn = 12

13 fp)) and β fp)) dol = covπ it,fp t ) varfp t) are analogous. Regressions 7)-9) thus estimate the covariance of risk premia and forward premia in a covπ it,fp it fp t ) ˆ fp i ˆ varfp it fp t ) ˆ fp i ˆ world with many foreign currencies. In an affine model of currency returns or in any first-order approximation of a non-linear model), the three coefficients β stat, β dyn, and β dol measure the elasticity of risk-premia with respect to forward premia in the cross-currency, between time and currency, and cross-time dimension, respectively. Columns 1-4 of Table 3 estimate specifications 7)-9) using our 1 Rebalance sample. The standard errors for β stat and β dol are clustered by currency and time, respectively, while the standard errors for β dyn are Newey-West with 12, 18 and 24 lags for the 1-, 6- and 12-month horizons respectively. An asterisk indicates that we can reject the null hypothesis that the coefficient is equal to zero at the 5% level. The specifications in column 1 use monthly forward contracts and show a highly statistically significant estimate for β stat of 0.47 s.e.=0.08). The estimate of β dyn is about the same size 0.44 s.e.=0.25), but statistically indistinguishable from zero, as is the much larger estimate for β dol 3.11, s.e.=1.60). The same column also reports estimates of the slope coefficients of equivalent specifications for the returns on the carry trade β ct ) and the FPT β fpt ), where in each case we regress currency returns in the relevant dimension on the portfolio weights used to implement the trading strategy rx i,t+1 rx t+1 = β ct fp it fp t ) + ɛ ct i,t+1, 10) rx i,t+1 rx i = β fp fpt it fp ˆ ) i + ɛ fpt i,t+1. 11) As expected, the coefficients in both regressions are positive and statistically significant. The coefficient in the carry trade regression is 0.68 s.e.=0.27), while the one in the FPT regression is 0.86 s.e.=0.34). In both regressions we use Newey-West standard errors with the appropriate number of lags, following the convention outlined above. As with the portfolio-based decomposition in Table 2, the coefficients β ct and β fpt are linear functions of β stat, β dyn and β dol, β dyn, respectively. 10 Column 1 of Table 3 thus also reports the partial R 2 of the static trade in the carry trade regression 62%) and the partial R 2 of the dollar trade in the FPT regression 90%). 11 The remaining columns report variations of the same estimates, using the same structure as Table 2. Columns 2-4 use returns adjusted for the bid-ask-spread and forward contracts at the 1, 6, and 12 month horizon. The remaining columns and panels repeat the same estimations using our 3, 6, and 12 Rebalance samples. 10 See Appendix A for the analytical expressions. 11 We calculate the partial R 2 as ESS d ESS d +ESS dyn, d = {stat, dol} where ESS dyn refers to the explained sum of squares in specification 8) and ESS stat, ESS dol refer to the explained sum of squares in specifications 7) and 9), respectively. 13

14 Table 3: Estimates of the Covariance of Risk Premia with Forward Premia 1) 2) 3) 4) 5) 6) 7) 8) Panel A 1 Rebalance 3 Rebalances Horizon months) Static T: β stat 0.47* 0.37* 0.56* 0.60* 0.26* 0.18* 0.26* 0.25* 0.08) 0.09) 0.10) 0.10) 0.05) 0.05) 0.04) 0.06) [0.000] [0.001] [0.000] [0.000] [0.000] [0.001] [0.000] [0.000] Dynamic T: β dyn * ) 0.25) 0.32) 0.26) 0.15) 0.15) 0.15) 0.15) [0.075] [0.095] [0.253] [0.043] [0.069] [0.108] [0.167] [0.078] Dollar T: β dol ) 1.58) 2.19) 2.16) 1.18) 1.18) 1.25) 1.20) [0.052] [0.050] [0.142] [0.085] [0.438] [0.481] [0.251] [0.138] Carry Trade: β ct 0.68* 0.55* 0.62* 0.71* 0.57* 0.45* 0.42* 0.43* 0.27) 0.26) 0.29) 0.26) 0.19) 0.18) 0.21) 0.19) % Partial R 2 Static T Forward Premium T: β fpt 0.86* 0.83* 0.85* 1.09* 0.41* * 0.60* 0.34) 0.34) 0.42) 0.40) 0.20) 0.20) 0.21) 0.21) % Partial R 2 Dollar T p-valβ dol β dyn ) N Panel B 6 Rebalances 12 Rebalances Static T: β stat 0.23* 0.15* 0.25* 0.24* 0.34* 0.23* 0.31* 0.30* 0.05) 0.05) 0.04) 0.05) 0.08) 0.08) 0.08) 0.08) [0.000] [0.005] [0.000] [0.000] [0.000] [0.011] [0.001] [0.001] Dynamic T: β dyn ) 0.14) 0.12) 0.06) 0.11) 0.11) 0.09) 0.05) [0.164] [0.251] [0.398] [0.757] [0.172] [0.268] [0.532] [0.773] Dollar T: β dol * ) 2.60) 2.22) 0.70) 2.26) 2.27) 2.20) 1.35) [0.737] [0.772] [0.410] [0.026] [0.448] [0.478] [0.993] [0.868] Carry Trade: β ct 0.56* 0.45* 0.45* * 0.52* 0.57* ) 0.17) 0.19) 0.14) 0.16) 0.16) 0.16) 0.17) % ESS Static T Forward Premium T: β fpt ) 0.19) 0.17) 0.08) 0.16) 0.16) 0.14) 0.05) % ESS Dollar T p-valβ dol β dyn ) N Bid-Ask Spreads No Yes Yes Yes No Yes Yes Yes Note: Estimates of the covariance of currency risk premia with forward premia in the cross-currency β stat ), between time and currency β dyn ), and cross-time dimension β dol ) using specifications 7), 8), and 9), respectively. The table also shows the slope coefficients from specifications 10) and 11) and the partial R 2 ESS, calculated as d ESS d +ESSdyn, d = {stat, dyn} where ESS dyn refers to the explained sum of squares in specification 8) and ESS stat, ESS dyn refer to the explained sum of squares in specifications 7) and 9), respectively. Standard errors are in round parentheses. P-value on null that coefficient is equal to zero in square parentheses. See text for details. 14

15 The pattern that emerges from the range of variations in Table 3 is similar to the results in column 1. In all samples, the coefficient on the static trade is a precisely estimated number between zero and one point estimates range from 0.15 to 0.6), and this coefficient usually explains about two thirds of the systematic variation driving the identification of β ct. We thus always reject the null that currency risk premia do not vary with unconditional differences in forward premia across currencies. The coefficient on the dollar trade is imprecisely estimated and statistically distinguishable from zero in one out of 16 specifications. Point estimates range from to We thus rarely reject the null that there is no co-variance between risk premia and forward premia in the cross-time dimension. However, the dollar trade always explains more than half, often more than 90% of the variation driving the identification of β fpt. In contrast, the dynamic trade often explains less than 10% of the variation identifying β fpt. Finally, we reject the null that currency risk-premia vary between time and currency in only one of our 16 specifications. These results have a number of surprising implications. First, the fact that we cannot reject the hypothesis that currency risk premia do not vary in the between time and currency dimension means that the FPP and the carry trade are not significantly related phenomena in the data. The FPP does not appear drive or motivate the carry trade, contrary to what most textbooks and many papers on the subject suggest. Models that are designed to fit the FPP thus do not automatically explain the the carry trade and vice versa. As a result, the two phenomena may require separate theoretical explanations. Second, throughout the table, the evidence that currency risk premia vary over time is quite weak. While both the Dynamic and the Dollar trade appear to yield positive expected returns in Table 2, the systematic part of the returns on these strategies are not statistically distinguishable from zero in most specifications. Recall that in 6) the terms α dyn and α dol result from expectational errors, such that risk premia on both the Dynamic and the Dollar trade are positive if and only if β dyn and β dol are strictly greater than zero, respectively.) In contrast, the most robust feature of the data appears to be the feature that has received least attention in the literature a significantly positive risk premium on the Static Trade, i.e. a significant covariance between currency risk premia and unconditional differences in forward premia across countries. 4.1 Model Selection Summing across equations 7), 8), and 9) and taking conditional expectations on both sides yields the generic affine model of currency risk premia with three parameters π it = γ + β dyn fp it f ˆp i fp t + f ˆp) + β stat f ˆp i f ˆp) + β dol fp t f ˆp). 15

16 A theorist wishing to focus her energy on the most salient features of the data may want to begin with the null hypothesis that each of these parameters are equal to zero and include them if and only if they significantly improve the model s fit to the data. Based on the results from Table 3, she might thus start with the simplest model that is not clearly rejected by the data {β stat > 0, β dyn = 0, β dol = 0}. This model explains returns on the carry trade as the result of static, unconditional, differences in risk premia across currencies. While this model explains most of the significant correlations shown in Table 3, it may not be satisfactory to discard the mean returns to the FPT, and thus the FPP itself, as a statistical fluke. Columns 1-5, 7, and 8 of Panel A, show significantly positive returns to the FPT. Although neither β dyn nor β dol are by themselves usually statistically distinguishable from zero, their weighted sum is statistically significant in these seven specifications. We might thus want to relax our model by adding an additional parameter that can explain this pattern. The three simplest options to extend the model are {β dyn > 0, β dol = 0}, {β dyn = 0, β dol > 0}, and {β dyn = β dol > 0}. Table 4 performs χ 2 difference tests, asking which of the three extensions is best able to explain the mean returns on the FPT observed in the data, under the assumption that the coefficients estimates of β fpt, β dyn, and β dol are normally distributed. The two columns in the table use the coefficient estimates and standard errors from columns 1 and 5 of Panel A in Table 3, respectively. As the linear relationship between the three coefficients holds only in the absence of transaction costs these are the only two relevant specifications.) In both cases, we cannot reject β dyn = 0 or β dyn = β dol, while we can reject β dol = 0 at the 5% level. The two simplest models that can explain all the statistically significant correlations in Table 3 are thus {β stat > 0, β dyn = 0, β dol > 0}, and {β stat > 0, β dyn = β dol > 0}. The main conclusion from this section is that the data strongly reject models in which β stat = 0 and, to the extent that the FPP is a robust fact in the data, also reject models in which β dol = 0. A parsimonious affine model of currency risk premia thus need only allow for variation in currency risk premia in the cross-currency and cross-time dimensions. Whatever we assume about β dyn, it does not significantly affect the model s ability to fit the data. 4.2 In Sample vs. Out of Sample Regressions The estimation in the previous section is based on out of sample regressions in the sense that fp ˆ i, fp ˆ are estimated in the pre-period. This approach came naturally as we used these regressions to analyze the statistical properties of the portfolios from section 3, where investors also needed to estimate fp ˆ i, fp ˆ in order to be able to form their portfolios. The following proposition shows that this is not an accident: any unbiased estimate of the covariance of riskpremia with forward premia in the cross-currency and between time and currency dimensions must take a stand on investors expectations of fp i and fp. 16

17 Table 4: χ 2 Difference Tests 1) 2) Sample 1 Rebalance 3 Rebalances Null Hypothesis p-values β dyn = β dol = β dyn = β dol Note: χ 2 difference tests of the ability of restricted linear models of currency risk premia to explain the returns on the FPT documented in columns 1 and 5 of Panel A in Table 2, under the assumption that the coefficients estimates of β fpt, β dyn, and β dol are normally distributed. Proposition 2 In a finite sample, in-sample estimates of β stat and β dyn are biased. The insample estimate of β dyn always over-estimates the true covariance of currency risk premia with forward premia in the between time and currency dimension. Proof. By definition, β dyn = E E it rx i,t+1 rx t+1 rx i rx)) [fp it fp t f ˆp i f ˆp)]) var fp it fp t f ˆp i f ˆp)) 1 Taking iterated expectations, adding and subtracting fp i fp) in the square brackets, and multiplying and dividing with var fp it fp t fp i fp)) yields β dyn = β dyn in sample + E rx ) i,t+1 rx t+1 rx i rx)) [fp i fp) f ˆp i f ˆp)]) var fpit fp t fp i fp)) var fp it fp t fp i fp)) var fp it fp t f ˆp i f ˆp)), where β dyn in sample = Erx i,t+1 rx t+1 rx i rx))[fp it fp t fp i fp)]) varfp it fp t fp i is the in-sample estimate from the fp)) specification rx i,t+1 rx t+1 rx i rx) = β dyn in sample [fp it fp t ) fp i fp)] + ɛ i,t+1. Now note that the second term in the round brackets is equal to zero and write β dyn = β dyn in sample var fp it fp t fp i fp)) var fp it fp t f ˆp i f ˆp)). The proof follows from the fact that var fp it fp t fp i fp)) > var fp it fp t f ˆp i f ˆp)) if f ˆp i fp i or f ˆp fp. The proof for β stat proceeds analogously where β stat = β stat in sample var fp i ) var f ˆp i ) + E [rx i rx) f ˆp i f ˆp fp i fp))]. var f ˆp i ) 17

18 An interesting corollary to this result is that in-sample estimates of β fpt i.e. versions of Fama s specification 1) that pool across currencies without introducing a full set of currency fixed effects) also return inflated estimates relative to those in Table 3. In contrast, the distinction between in-sample an out of sample regressions makes no difference when estimating the covariance of risk premia with forward premia in the cross-time dimension or in a two-country world: equations 1) and 9) both have intercepts α i and α) that absorb any errors in predicting fp i or fp. 4.3 Dynamics of Bilateral Currency Risk Premia Given the large literature that analyzes the dynamics of bilateral currency risk-premia using currency by currency regressions 1), a natural question is whether our three parameter model is too restrictive by imposing the same between time and currency dynamics for all foreign currencies. In this section, we relax this assumption by allowing for the possibility that the risk premia of specific currencies vary with deviations of forward premia from their time and currency specific means. In particular, we can re-write the decomposition in 6) as β stat var fpi ˆ fp ˆ ) + } {{ } i Static Trade where β dyn i cov rx i,t+1, fp it ) = β dyn i var i fp i,t fp t ) + α dyn } {{ } Dynamic Trade are currency-specific coefficients from the regression rx i,t+1 rx t+1 rx i rx) = α i + i D i β dyn i + β dol var fp t fp ˆ ) + α dol α dol. } {{ } Dollar Trade 12) [ fp it fp t ) fpi ˆ fp ˆ )] + ɛ dyn i,t+1 13) and D i is a currency fixed effect. Table 5 shows the coefficients from this regression for our 1, 3, 6, and 12 Rebalance samples. To save space we show only the coefficients using one-month forwards, without taking into account bid-ask spreads. An asterisk again denotes significance at the 5% level, where standard errors in columns 1-3 are again Newey-West with 12 lags. Column 4 uses robust standard errors. 12 The table shows that we cannot reject the null that cov i π it, fp it fp t ) = 0 for most currencies. In fact, looking across columns, we do not appear to robustly reject this null for any currency, with the possible exception of the Indian rupee and the South African rand. 12 Newey-West standard errors may be problematic in our 12 Rebalance sample which allows entry every 24 months and thus features some relatively short time series for individual currencies. 18

19 While we remain open to the possibility that risk-premia of these, and potentially a few other, currencies may co-move with deviations of forward premia from their time- and currency specific mean, the evidence does not appear overwhelming. In addition, note from comparing equations 6) and 12) that β dyn var fp i,t fp t fpi ˆ fp ˆ )) = i β dyn i var i fp i,t fp t ). 14) Allowing for different covariances between the risk premia of specific currencies and deviations of forward premia from their time and currency specific means thus does not change the model s ability to match the expected returns to the carry trade and the FPT as defined in 3) and 4). 4.4 Is the Dollar Special? Throughout the paper we account for returns in terms of US dollars. While this convention is standard practice in the literature, it is also somewhat arbitrary. How would our results change if we had chosen a different base currency? Given a large enough sample of currencies, our estimates of the returns on the Dynamic and the Static trades, as well as our estimates of β stat and β dyn would not change at all, as both strategies are neutral with respect to the base currency i.e. their returns are not affected by the returns on the base currency). However, our estimates β dol might be different as the Dollar Trade is not neutral with respect to the returns on the dollar. Proposition 3 In a large sample of convertible currencies, the covariance of the risk premium on any base currency j with the average forward premium on all other foreign currencies equals the covariance of currency j s risk premium against the US dollar with deviations of its forward premium against the US dollar from its time and currency specific mean, β j = β dyn j. Proof. First, we generalize our notation to account for returns in units of different currencies. Denote by fp j i,t the forward premium of currency i against currency j at time t, where for the US dollar we maintain fp dol i,t = fp i,t. By convertibility we have fp j i,t = fp i,t fp j,t, Δs j i,t+1 = Δs i,t+1 Δs j,t+1, and thus rx j i,t+1 = rx i,t+1 rx j,t+1, where we again use the convention that Δs j i,t+1 and rxj i,t+1 refer to values in terms of currency j. If the number of currencies is large, we can also write fp j t = fp t fp j,t and consequently, f ˆp j = f ˆp f ˆp j. 19

20 Table 5: Bilateral Dynamic Trade Regressions, one month horizon 1) 2) 3) 4) 1 Rebalance 3 Rebalances 6 Rebalances 12 Rebalances Australia Austria 4.29* 4.40* Belgium 2.95* 3.79* Canada Czech Rep Denmark ECU * Euro France Germany * Hong Kong Hungary 6.06* 8.69* 6.27* Iceland India 3.66* 3.44* 3.59* Indonesia 2.67 Ireland Italy Japan Korea Kuwait Malaysia * Mexico Netherlands New Zealand Norway Philippines Poland * Saudi Arabia 2.72* Sweden 3.08* Singapore Slovak Rep * Spain Switzerland Taiwan Thailand Turkey UAE * United Kingdom South Africa * 2.95* 0.92 β dol p-valβ dol = i ω iβ dyn i ) Note: Estimates of the currency-specific covariance of risk premia with forward premia β dyn i using specification 13). An asterisk denote statistical significance at the 5% level, where standard errors not shown) are robust to heteroskedasticity. The dependent variable in all columns are calculated using 1 month forward contracts. In the test, ω i = var ifp i,t fp t) varfp i,t fp t fp i fp)). 20