The Non-U.S. Bank Demand for U.S. Dollar Assets

The USD asset share of non-U.S. banks captures the demand for dollars by these investors. An instrumental variable strategy identifies a causal link from the USD asset share to the USD exchange rate. Cross-sectional asset pricing tests show that the USD asset share is a highly significant pricing factor for carry trade strategies. The USD asset share forecasts the dollar with economically large magnitude, high statistical significance, and large explanatory power, both in sample and out of sample, pointing towards time varying risk premia. It takes 2-5 years for exchange rate risk premia to normalize in response to demand shocks.

Abstract

The USD asset share of non-U.S. banks captures the demand for dollars by these investors. An instrumental variable strategy identifies a causal link from the USD asset share to the USD exchange rate. Cross-sectional asset pricing tests show that the USD asset share is a highly significant pricing factor for carry trade strategies. The USD asset share forecasts the dollar with economically large magnitude, high statistical significance, and large explanatory power, both in sample and out of sample, pointing towards time varying risk premia. It takes 2-5 years for exchange rate risk premia to normalize in response to demand shocks.

I. Introduction

Market commentary often associates exchange rate movements with “changes in positioning” and more generally demand and supply effects. In fact, Froot and Ramadorai (2005) show that institutional investors’ order flow helps explain transitory discount rate news of exchange rates. Furthermore, it is well known that exchange rates can deviate substantially and persistently from macroeconomic fundamentals (Rogoff 1996). As a result, an emerging recent literature has focused on documenting that demand and supply effects help explaining exchange rate movements (e.g., Engel and Wu 2018; Jiang, Krishnamurthy, and Lustig 2018; Lilley, Maggiori, Neiman, and Schreger 2019).

In this paper, we present a new measure of foreign “safe asset” demand for the U.S. dollar (USD), constructed as the share of non-U.S. banks’ USD assets. We construct the “safe asset” demand metric for 26 advanced and emerging economies, and document that this metric is consistently related to exchange rate movements contemporaneously and can help forecast exchange rates both in sample and out of sample. Furthermore, average currency excess returns are systematically related to the cross-section of betas with respect to the USD asset share.

The USD asset share is computed for each economy as the ratio of assets denominated in USD to total assets of the non-U.S. banking system on nationality basis, based on where the ultimate parent of reporting banking-system is headquartered. The USD positions include international positions, that is the USD operations outside the U.S., and U.S.-based branch operations. We don’t include non-U.S. banks’ U.S.-based subsidiaries, because they are subject to local regulation in the U.S. similar to domestic legal entities.

Time-series regressions show that the non-U.S. bank demand for U.S. dollar assets significantly and with an economically large magnitude contributes to the explanation of contemporaneous movements in the USD exchange rate with a partial R2 of around 36 percent. The significance and economic magnitude are unchanged by including the U.S. Treasury premium and the home minus U.S. interest differential as controls. Furthermore, the USD asset share significantly correlates with 11 out of 16 bilateral USD exchange rate movements vis-à-vis individual foreign currencies with high R2s.

The USD asset share is an endogenous variable, and hence we establish a causal relationship using valid demand shifters. The demand of USD denominated safe assets by non-U.S. banks would be expected to be a function of the safety and liquidity of substitute assets and the balance sheet capacity of non-U.S. banks to bear risks. One instrumental variable is the Treasury premium of non-U.S. G10 countries, which is proportional to the convenience yield on substitute safe assets. A second instrumental variable is the sovereign risk of substitute Treasury securities from non-U.S. G10. These two instrumental variables used as demand shifters belong to the group that captures the safety and liquidity of non-U.S. G10 currencies that is the closest substitute for the USD assets. A third instrumental variable used as demand shifter is the shock to the leverage of non-U.S. banks (orthogonal to the leverage of U.S. banks), which relates to the balance sheet capacity of non-U.S. banks to bear risks.

All three sets of instrumental variables are shown to be valid in a statistical sense, and economic intuition suggests that these are exogeneous demand shifters as they are correlated with USD asset demand, but not with the supply of USD assets. We find that an increase in the Treasury premium of non-U.S G10 countries is significantly associated with decreasing demand for USD assets by non-U.S. banks; higher sovereign risk in Treasury securities of non-U.S. G10 countries is associated with a significant increase in the foreign demand for USD safe assets; and higher leverage of non-U.S. banks is significantly related to higher demand for safe USD assets. In addition, an overidentification test based on the Sargan X2 statistic confirms that all three instrumental variables are valid exogenous demand shifters. In terms of economic significance, these instrumental variables alone account for more than 20% of the variation in contemporaneous changes of the USD asset demand from non-U.S. banks. Based on the critical values reported in Stock and Yogo (2002), we can reject the null hypothesis that the proposed instrumental variables are weak instruments.

When we investigate the relationship between the USD asset share and the USD exchange rate in the second stage, we continue to find that the higher fitted values of USD asset shares of non-U.S. banks (obtained from first stage) account for a contemporaneous appreciation in the USD exchange rate significantly at the 1% level. Note that the magnitude of the coefficient point estimates on changes in USD asset shares obtained by two-step-least square (2SLS) is around two times as large as those obtained by ordinary least squares (OLS). The F -statistic of the Durbin-Wu-Hausman test of endogeneity rejects that the USD asset share is an exogenous variable. These results suggest that the slope of supply curve of USD assets is underestimated by OLS regressions because it uses equilibrium information of price and quantity. This result suggests that our IV strategy identifies the supply curve and corrects the coefficient point estimates on the slope due to the usage of valid exogenous demand shifters.

Perhaps more strikingly, those strong contemporaneous results are complemented with forecasting regressions. When there is a positive innovation to USD asset demand, the U.S. dollar appreciates contemporaneously, and then forecasts a U.S. dollar depreciation one, two, three, and five years out. These forecasting results are statistically significant and economically large. The R2 for the USD exchange rate changes eight quarters ahead is 36%, with the USD asset share significant at the 1% level. More strikingly, at the twenty-quarter-ahead horizon, the R2 is 67% and the USD asset share is again significant at the 1% level.

There may be complementary economic channels at play. First, there is clearly a “safe asset demand channel” that is also demonstrated by Engel and Wu (2018), and Jiang, Krishnamurthy, and Lustig (2018). Higher USD safe asset demand from non-U.S. banks is associated with a stronger USD exchange rate contemporaneously. Second, our result coincides with the prediction by intermediary asset pricing theory that higher USD safe asset demand from non-U.S. banks corresponds to lower balance sheet capacity of non-U.S. banks, thus being associated with a significantly negative price of risk in the cross-section of currency excess returns, indicating that these institutions are marginal investors in the foreign exchange (FX) market. This finding resonates with Adrian, Etula, and Muir (2014) and He, Kelly, and Manela (2017) who find that intermediary balance sheet capacity is a significant pricing factor for risky assets. Third, we show that higher demand for USD safe assets from non-U.S. banks predicts USD exchange rate depreciation over a longer horizon, but barely predicts the exchange rate dynamics at a shorter horizon. This suggests time variation in exchange rate risk premia. In particular, it reflects the stationary but persistent deviation from the uncovered interest parity that Duarte and Stockman (2005) and Engel (2014 and 2016) have document. Lilley, Maggiori, Neiman, and Schreger (2019) also find that the U.S. purchase of foreign bonds is associated with risk premia after the GFC, because there is a strong correlation between these flows with traditional risk measures. In summary, because the demand for USD assets driven by risk-bearing capacity of non-U.S. banks is associated with time varying currency risk premia, it helps predict exchange rate dynamics.

We follow the tradition established by Meese and Rogoff (1983) and find these forecasting results also hold out of sample using Diebold-Mariano and Clark-West statistics, which means our forecasting model significantly outperforms the random walk prediction. We document highly significant (i.e., at the 1% level) out-of-sample forecastability for the U.S. dollar exchange rates against a basket of 16 currencies, as well as for the vast majority of bilateral exchange rates. To our knowledge, the strength of out-of-sample forecasting power from a quantity variable is unprecedented in the literature on exchange rates.

As a robustness check, we augment the 9 popular exchange rate models surveyed by Rossi (2013) with the USD asset demand from non-U.S. banks in out-of-sample forecasting tests. We find that the out-of-sample forecasting performance improves after we incorporate the average USD asset demand for all 9 models.

The remainder of the paper is organized as follows. Section II briefly discusses the related literature. Section III describes the construction of the USD demand by non-U.S. banks. Section IV presents the main contemporaneous results, including the Two-Stage Least Squares (2SLS) regressions. Section V shows cross-sectional asset pricing evidence. Section VI gives the forecasting results. Section VII lays out robustness checks. Section VIII discusses implications for the literature. Section IX concludes.

II. A Brief Literature Review

We now provide a brief overview of the existing related work. A more detailed discussion of implications for the literature can be found in Section VIII. This paper is closely relevant to four threads of literature.

First, there is a large literature about the U.S.’ special role as the provider of international reserve currency.2 A number of papers document that there is a positive and countercyclical safety premium for the U.S. dollar (e.g., Lustig, Roussanov, and Verdelhan 2014; Du, Im, and Schreger 2018; Jiang, Krishnamurthy, and Lustig 2018; Verdelhan 2018). All these papers focus on the U.S. Treasury premium, which captures the scarcity of USD assets based on a single type of security (a 1-year government bond). Our paper first documents the counter-cyclicality in USD asset demand of marginal investors by using a quantity measure, i.e., the USD asset share of non-U.S. banks. Our paper shows that there is a “safe asset demand channel” of non-U.S. financial intermediaries at play – the non-U.S. financial intermediaries pay a safety premium to hold the USD denominated assets, which provides an insurance when negative shocks happen because the U.S. dollar appreciates contemporaneously.

Second, starting with the seminal contribution of Meese and Rogoff (1983), there is a prevailing view that exchange rates follow a random-walk-like process and hence are not predictable, especially at a short horizon. Several well-known exchange rate puzzles result from this view. First, the “exchange rate disconnect puzzle” shows that the nominal exchange rate is not robustly correlated, even contemporaneously, with macroeconomic fundamentals (e.g., Mark 1995; Cheung, Chinn, and Pascual 2005; Engel and West 2005; Rogoff and Stavrakeva 2008). Second, the “UIP puzzle” implies that on average the interest differential is not offset by a commensurate depreciation of the investment currency (e.g., Fama 1984; Brunnermeier, Nagel, and Pedersen 2009). Our paper documents the demand for safe USD assets from non-U.S. banks is associated with a time-varying currency risk premium, so it not only helps account for contemporaneous changes in exchange rates, but also helps predict exchange rate dynamics over the longer horizon both in sample and out of sample.

Third, there is a large literature about the “global financial cycle”, which is in fact a “global USD cycle”. For example, Rey (2013) and Miranda-Agrippino and Rey (2015) consider the VIX index and U.S. monetary policy as key drivers of global financial cycles, and Avdjiev, Du, Koch, and Shin (2019) find the broad USD exchange rate is a risk barometer in global capital markets. Our paper contributes to this literature by establishing a causal relationship between USD asset demand, the U.S. dollar exchange rate, the U.S. Treasury premium, and the covered interest parity (CIP) deviation with a novel instrumental variable approach. It relates to a recent literature that explores the failure of CIP for LIBOR (e.g., Ivashina, Scharfstein and Stein 2015; Du, Tepper and Verdelhan 2018) and for Treasuries (e.g., Engle and Wu 2018; Jiang, Krishnamurthy, and Lustig 2018).

Lastly, our paper also sheds light on the theory of intermediary asset pricing. Prominent examples of such theories include Brunnermeier and Pedersen (2009), He and Krishnamurthy (2012, 2013), Brunnermeier and Sannikov (2014), and Adrian and Shin (2014). Adrian, Etula, and Muir (2014) and He, Kelly, and Manela (2017) find that the leverage of U.S. securities broker-dealers possesses significant explanatory power for the cross-sectional variation of expected returns for a wide range of asset classes. This paper focuses on non-U.S. banks and finds that the average USD asset demand from non-U.S. banks to be a global risk factor which can help to explain currency excess returns in the cross section.

III. Measuring Demand for USD Assets by Non-U.S. Banks

In this section, we first define our measure of demand for USD assets by non-U.S. banks and describe our data sources. We also take a glance at how the measure relates to the U.S. dollar exchange rate.

The bilateral exchange rate of any currency vis-à-vis the U.S. dollar is the relative value of this currency with respect to the U.S. dollar, i.e., the U.S. dollar is the numéraire. To be comparable, we want to measure the USD asset demand within each foreign economy as the relative holding denominated in USD to the holding denominated in corresponding foreign currency. Given that non-U.S. banks from each foreign economy are the marginal investors for the foreign currency there, their demand should impact the bilateral exchange rate of the U.S. dollar vis-à-vis this specific currency. We assume the total assets of non-U.S. banks from economy j are dominated by assets denominated either in currency J used in corresponding economy j or in USD, as the U.S. dollar historically has played a prominent role in global trade and financial flows (e.g., Farhi and Maggiori 2018; Gopinath and Stein 2018). As a result, we define the relative demand for USD assets by non-U.S. banks as the share of USD denominated assets to total assets.3

Given that the analysis aims to measure the USD demand of non-U.S. banks, the nationality of which should matter most, which is determined by where ultimate owners of the banks are headquartered, as opposed to the domicile or residency basis. Due to data limitations on bank-level USD balance sheets, our analysis is performed by aggregating non-U.S. banks’ balance sheets at economy level on nationality basis. Relying on USD balance sheets by nationality of reporting non-U.S. banks, a graphical representation of the different aggregates is shown in Figure 1.

Figure 1.
Figure 1.

USD International Positions vs. Foreign Positions

Citation: IMF Working Papers 2020, 101; 10.5089/9781513547732.001.A001

Notes: ① corresponds to the U.S. dollar cross-border claims; ② refers to U.S. dollar local positions outside the U.S. (i.e., U.S. dollar is the foreign currency of the economy); ③ and ④ refer to the local positions in U.S. dollar in the U.S for non-U.S. banks’ branches and subsidiaries. ① and ② are constructed based on the BIS unpublished free and restricted database of Locational Banking Statistics (LBS) on nationality basis. Due to data limitations, ③ is constructed through a bottom-up aggregation based on nationality of balance sheet information available in FFIEC 002 Regulatory Filings. Similarly, ④ is constructed using data from the Call Report Filings (FFIEC 031/041) of non-U.S. -owned subsidiaries, obtained via the S&P Global Market Intelligence platform. The underlying assumption is that the operations of non-U.S. banks in the U.S. are dominated by USD-denominated claims and liabilities.

International Positions (IP) are the USD operations outside the United States, which capture ① the cross-border positions of non-U.S. banks, and ② the local positions of non-U.S. banks in foreign currency (i.e., USD) outside the United States. The analysis explores the USD-denominated claims from the BIS unpublished free and restricted database of BIS Locational Banking Statistics (LBS) on nationality basis.4 The Banks’ positions are broken down by currency, by sector (bank or non-bank), by country of residency of the counterparty, by country of residence of the reporting banks, and by nationality of the ultimate owner of reporting banks.5 Hence, the USD-denominated aggregates of IP positions can be constructed at the economy level on nationality basis, based on where the ultimate parents of reporting non-U.S. banks are headquartered.

Operations in the U.S. are the USD operations of non-U.S. banks within the U.S., which include both ③ non-U.S. banks’ branch operations and ④ non-U.S. banks’ subsidiary operations. In the BIS Locational Banking Statistics (LBS), the U.S. does not report local claims/liabilities of non-US banks (and of US banks too). Branch operations within the U.S. are constructed through a bottom-up aggregation based on nationality of balance sheet information available in FFIEC 002 Regulatory Filings. Subsidiary operations in the U.S. are similarly constructed using data from the Call Report Filings (FFIEC 031/041) of non-U.S.-owned subsidiaries, obtained via the S&P Global Market Intelligence platform. The underlying assumption is that the operations of non-U.S. banks in the U.S. are dominated by USD-denominated claims and liabilities.

In this paper, the non-U.S. banks’ balance sheet aggregates encompass the International Position (IP) and ③ U.S.-based branches. We don’t include U.S.-based subsidiaries, which constitutes the Foreign Position (FP), because non-U.S. banks’ U.S.-based subsidiaries are subject to local regulation in the U.S. similar to domestic legal entities, the funding and liquidity management decisions of which are at least partially (if not entirely) independent of the headquarter.6 Following the similar approach as USD assets, we construct the total assets of non-US banks (for all currencies) on a nationality basis. However, there are data limitations for local positions in local currencies from BIS Locational Banking Statistics (LBS), which start only from 2012 for most nationalities. The aggregates of local positions in local currencies on nationality basis, which enter the denominator of USD asset share of non-U.S. banks, are obtained based on a bottom-up consolidated approach using FitchConnect and Factset database.7 In the absence of the currency breakdown at bank-level data, we assume the share of local currency position to all currency position is equal to what is available in the Other Depository Corporations (ODCs) survey. Although the ODCs survey is based on residency basis as opposed to nationality, but this is the best proxy we could think of. The initial sample covers 26 economies, where USD operations are considered to be of systemic importance (based on the IMF 2019).8 For convenience, Germany is used as the representative for eight economies in the Euro Area, since the Euro has been adopted by most of those countries during our sample period (2001q2–2017q4).9 China and Russia are excluded, because the relevant USD aggregates only span over 2 years. We don’t include Hong Kong SAR China for its adoption of the fixed exchange rate regime. 10 Table 1 summarizes the sample of economies used in this paper, which covers 16 legal tender currencies with (largely) floating exchange rate regimes from 16 economies covering 10 advanced and 6 emerging economies.

Table 1.

Sample Coverage and Summary Statistics

article image
Notes: This table describes the sample of 16 economies (based on nationalities of ultimate owners of reporting non-U.S. banks) used in this paper. It also presents the summary statistics of key variables. Columns (1) – (16) report by economies, and Column (17) reports the cross-sectional average over all 16 economies. In the upper (lower) panel, we report the summary statistics in levels (first differences) for key variables, including the log nominal exchange rate per U.S. dollar (st); the USD asset share of non-U.S. banks (D$t); the U.S. Treasury premium (Φt); home – U.S. government bond yield differential (yt – y$t). All series are valued at quarter end. The longest sample covers 2001q1 to 2017q4 for the upper panel and 2001q2 to 2017 q4 for the lower panel.

Table 1 also presents the summary statistics of key variables that we are interested in in this paper. Columns (1) to (16) report the summary statistics by economies, and Column (17) reports those of the cross-sectional average across all 16 economies. In the upper panel, we report the summary statistics of key variables in levels. First, the variable st stands for log nominal exchange rate per U.S. dollar, an increase of which means that the U.S. dollar appreciates. Second, the variable D$t stands for the USD asset share of non-U.S. banks from specific economies (nationalities of reporting banks). For example, D$t in Column (1) stands for USD asset share of the Australian banks around the world, opposed to that of non-U.S. banks reside in Australia, because our measure is constructed on nationality basis. The Swiss banks around the world have the highest USD asset share standing at 27.8 percent on average over the sample period, whereas that of Brazil has the lowest USD asset share at 3.8 percent instead. The average USD asset share across those 16 economies is around 9.5 percent during the sample period.

Third, the variable Φt stands for the U.S. Treasury premium at the one-year horizon.11 The U.S. Treasury premium is positive across almost all economies, which suggests that investors are willing to pay a safety premium to hold U.S. Treasury securities relative to foreign government bonds in that economy. The only exception is that the U.S. Treasury has negative premium vis-à-vis Australia at around -6 bps on average.12 The U.S. Treasury has the largest premium vis-à-vis Brazil around 189 bps on average. The average U.S. Treasury premium, i.e., the U.S. Treasury premium against a basket of 16 government bonds, is around 38 bps over the sample period. The variable yt – y$t stands for the government bond yield differential between each economy and the United States. Emerging markets tend to have higher yield differentials against the United States. For example, the government bond yields in Brazil and Turkey are 9 percent higher than that of the United States on average. On the other hand, some safe heavens (e.g., Japan and Switzerland) have even lower yields than the United States. The average yield differential between the cross-sectional average of 16 economies and the U.S. is around 2.2 percent over the sample period. In the lower panel, we report the summary statistics of the quarterly changes of the key variables. The longest sample covers the period 2001q1–2017q4 for the levels in the upper panel and, correspondingly, 2001q2 to 2017q4 for the changes in the lower panel.

Figure 2 shows the time series properties of quarterly changes in the log USD exchange rate (Δst) and the quarterly changes in the USD asset share ((ΔD$t). The contemporaneous correlation is 58%, much higher than the correlation between the USD and alternative proxies of safe asset demand such as quarterly changes in the U.S. Treasury premium (ΔΦt) which has a 36% correlation or interest differentials (yt – y$t) which have a -32% correlation. Furthermore, and perhaps more strikingly, the USD asset share highly correlates with future exchange rate changes 3-years ahead (with a correlation of 55%, again much higher than the corresponding correlation of the U.S. Treasury premium of 29% or the interest differential of 7%). These bilateral correlations are presented in Table 2.

Figure 2.
Figure 2.

U.S. Dollar Asset Share and Exchange Rate

Citation: IMF Working Papers 2020, 101; 10.5089/9781513547732.001.A001

Notes: The figure plots the annualized quarterly changes in log nominal exchange rate and quarterly changes in U.S. dollar asset share of non-U.S. banks over time. All series are cross-sectional averages over non-U.S. G10 currencies and valued at quarter end. An increase in the exchange rate means the U.S. dollar appreciates against a basket of non-U.S. G10 currencies. The contemporaneous correlation between the quarterly changes in USD exchange rate and quarterly changes in U.S. dollar asset share of non-U.S. banks is 58% over time. In the non-USD G10 currencies sample, the USD is excluded as the numéraire and NZD is excluded because of data availability for the USD asset share of non-U.S. banks from the Bank of International Settlements (BIS). The sample period is from 2001q2 to 2017q4.
Table 2.

Correlation Between Exchange Rate and Determinants

article image
Notes: The upper panel shows the contemporaneous correlations between annualized quarterly changes in the log nominal exchange rate (Δst) and quarterly changes in selected determinants. The lower panel shows the correlations between annualized future changes in the log nominal exchange rate between t and t+12 quarters ahead (Δst12) and selected determinants. The variable D$ denotes the USD asset share of non-U.S. banks. The variable Φ denotes the U.S. Treasury premium. The variable (y – y$) denotes home – U.S. government bond yield differentials. All series are cross-sectional averages over non-USD G10 currencies and valued at quarter end. The USD is excluded as the numéraire, and NZD is excluded because of data availability for the USD asset share of non-U.S. banks from the Bank of International Settlements (BIS). The sample period is from 2001q2 to 2017q4.

IV. Exchange Rates and Demand for USD Assets

In this section, we first show that there is a strong contemporaneous relationship between the exchange rate and the demand for USD assets from non-U.S. banks in the time series. Second, we give a causal interpretation using a set of novel instrumental variables.

A. Time-Series Relationship

We examine the relationship between the demand for USD assets from non-U.S. banks and the exchange rate over time. We regress quarterly changes in the log nominal USD exchange rate on contemporaneous quarterly changes in the demand for USD assets, captured by USD assets as share of total assets from non-U.S. banks. Our benchmark regression specification is given by:

Δsj,t=α1+β1ΔD$,j,t+β2ΔΦj,t+β3Δ(yj,tGovty$,tGovt)+βXj,t+εt(1)

where the dependent variable Δsj,t is the annualized quarterly change in the log nominal exchange rate of the USD vis-à-vis currency J, an increase of which means USD appreciates against currency J between quarters t-1 and t. The variable ΔD$,j,t denotes the quarterly change in the USD asset share of non-U.S. banks from economy) between quarters t-1 and t. The variable ΔΦj,t denotes the quarterly change in the U.S. Treasury premium vis-à-vis economy j between quarters t-1 and t. The variable Δ(yj,tGovty$,tGovt) denotes the quarterly change in the interest differential of economy j over the United States between quarters t-1 and t. Finally, XJ,t is a vector of control variables.13

Table 3 shows our regression results. In Columns (1) to (7), we perform regressions by using the cross-sectional average series across 16 economies. It can be interpreted as a relationship between the annualized quarterly changes in the log USD exchange rate against a basket of 16 currencies and the cross-sectional average of demand for USD assets by non-U.S. banks from those 16 economies. The coefficient point estimates on ΔD¯$,t are positive and significant at the 1% level with similar magnitudes across all specifications. This is consistent with theory prediction (e.g., Engel and Wu 2018; Jiang, Krishnamurthy, and Lustig 2018), suggesting that an increase in USD asset demand from non-U.S. banks is associated with U.S. dollar appreciation against a basket of currencies. In terms of magnitude, the coefficient point estimate on ΔD¯$,t from the univariate regression in Column (1) implies that a one percent quarterly increase in average USD assets (as share of total assets) of non-U.S. banks across 16 economies is associated with an appreciation of the USD against a basket of 16 currencies by 6.8% in the same quarter.14 To provide a further sense of magnitudes, a positive shock to the USD asset share by one-standard deviation of 0.33% (Table 1) is associated with an appreciation in the USD vis-à-vis a basket of 16 currencies by around 2.2% in that quarter on average. In terms of economic significance, the variation in the USD asset demand of non-U.S. banks alone accounts for 36% of the variation in the contemporaneous movements of U.S. dollar vis-à-vis a basket of 16 currencies. Given the well-known exchange rate disconnect puzzle (Frankel and Rose 1995; Froot and Rogoff 1995), this magnitude of R2 is very meaningful and surprisingly high.

Table 3.

USD Assets Share and Exchange Rates

article image
Notes: The table reports the OLS regression results for the contemporaneous relationship between the USD exchange rate and the USD asset share of non-U.S. banks, with specification: Δsj,t=α1+β1ΔD$,j,t+β2ΔΦj,t+β3Δ(yj,tGovty$,tGovt)+βXj,t+εt, where the dependent variable Δsj,t is the annualized quarterly change in the log nominal exchange rate of the USD vis-à-vis currency J, an increase of which means the USD appreciates against currency J between quarters t-1 and t. The variable ΔD$,j,t denotes quarterly changes in the USD asset share of non-U.S. banks from economy j between quarters t-1 and t. The variable ΔΦj,t denotes quarterly changes in the U.S. Treasury premium vis-à-vis economy j between quarters t-1 and t. The variable Δ(yj,tGovty$,tGovt) denotes the quarterly changes in the interest differential of economy j over the United States between quarters t-1 and t. Finally, Xj,t is a vector of control variables including D$,j,t-1, Φj,t-1, and Δ(yj,tGovty$,tGovt). In Columns (1) to (7), we perform regressions by using cross-sectional average series across 16 economies, within which Columns (1) to (5) uses the full sample, and Columns (6) and (7) only use pre-GFC and post-GFC episodes, respectively. In Columns (1) – (7), all series are cross-sectional averages over 16 currencies and valued at quarter end. In Columns (8) – (23), we perform regressions by currencies. All specifications include a constant that is not reported in the table. The Newey-West heteroskedasticity-and-autocorrelation-consistent asymptotic standard errors are reported in parentheses. *, **, and *** denote significance levels at 10%, 5%, and 1%. The full sample period is from 200 1q2 to 2017q4.

Column (2) of Table 3 includes the contemporaneous quarterly change in the U.S. Treasury premium, which is an alternative measure of USD asset demand advocated by Engel and Wu (2018), and Jiang, Krishnamurthy, and Lustig (2018). Our measure is different from theirs in at least three aspects: 1) we use quantity information instead of price information; 2) we focus on all USD-denominated assets rather than only one specific USD-denominated assets (U.S. Treasury at 1-year maturity); and 3) we only focus on marginal investors (i.e., banks) rather than all investors. In Column (2), the coefficient point estimate for ΔΦ¯t is positive as theory predicts but it is insignificant and the R2 only increases marginally to 38%, which implies the additional information from this alternative measure documented in the literature is limited after controlling for the quantity measure of USD demand from non-U.S. banks, that is the USD asset share of non-U.S. banks.

Column (3) of Table 3 includes the contemporaneous quarterly change in the interest differential. The result is consistent with simple predictions from textbook models of exchange rate determination – interest rate increases in the U.S. relative to a basket of currencies are associated with USD appreciation. However, this relationship is not significant.15 Furthermore, the additional contribution to the R2 is negligible.

Column (4) of Table 3 includes both covariates of ΔΦ¯tandΔ(y¯tGovty$,tGovt) along with changes in the USD asset share (ΔD¯$,t). The R2 rises to 40%, and the coefficient point estimate on ΔΦ¯t becomes marginally significant at the 10% level, and coefficients and significance for other variables are nearly unchanged. In Column (5), we include other potential drivers (the one quarter lags of D¯,Φ¯,and(y¯Govty$Govt) at quarter t – 1), the ΔΦ¯t becomes significant at the 1% level and the R2 rises to 50%. This specification is equivalent to including the AR (1) innovations of the three factors discussed above. It is consistent with the finding of Jiang, Krishnamurthy, Lustig (2018) that the AR (1) innovations of ΔΦ¯t has additional information on contemporaneous movements of exchange rates.16 Note that the coefficient point estimates and significance of D¯ are nearly unchanged.

In Columns (6) and (7), we present the results for the pre-GFC (2001q2 to 2006q4) and the post-GFC (2010q1 to 2017q4) periods using the same specification as in Column (5). For the pre-GFC episode, the coefficient point estimate on ΔD¯$,t drops by 40% but is still significant at the 1% level, and ΔΦ¯t loses explanatory power and flips sign, and the coefficient point estimate on Δ(y¯tGovty$,tGovt) violates the textbook prediction.17 For the post-GFC episode, the magnitude of the coefficient point estimate on ΔD¯$,t is back to the magnitude of the whole sample presented in Column (5) and remains significant at the 1% level, and ΔΦ¯t gains explanatory power, and the coefficient point estimate on Δ(y¯tGovty$,tGovt) supports the text book prediction but becomes insignificant again. The R2 is above 50% for both episodes. The result for the pre-crisis episode distinguishes our findings from the recent development in exchange rate disconnect literature, which documents a reconnection between the dollar exchange rate and measures of risk or liquidity has emerged only after Global Financial Crisis in 2007–2009.18

In the right panel of Table 3, we adopt the same specification as in Column (5) but focus on the bilateral exchange rate of the USD vis-à-vis different individual currencies in our sample (see, Columns (8) to (23)). We find that the contemporaneous relationship between the USD asset share of non-banks from economy) and the exchange rate of the USD vis-à-vis currency/ significantly holds for 11 out of 16 cases.19 The significant coefficient point estimates on ΔD$,j,t vary from 6.5 (CHF) to 65.9 (BRL). The changes in the U.S. Treasury premium also have explanatory power for 10 out of 16 cases. The changes in interest differential have more explanatory power for bilateral exchange rates: 9 out of 16 cases support the textbook prediction. The cross-sectional average results are mostly contaminated by currencies in emerging markets, where they either have contradictive and significant signs (MXN, TRY, ZAR) supporting the UIP condition or have insignificant results (BRL, INR, and MYR).

Taken together, these results show that the USD asset share of non-U.S. banks is a robust explanatory variable for the variation in the USD exchange rate over time. The explanatory power doesn’t change before or after Global Financial Crisis. Adding additional drivers does not affect the explanatory power of quantity measure of USD safe asset demand captured by USD asset share, especially in the case of controlling for the price measure of USD safe asset demand (the U.S. Treasury premium).

Our findings suggest that the USD exchange rate vis-à-vis individual foreign currency of corresponding economy is, to a large extent, driven by a simple safe asset demand story from the perspective of non-U.S. banks. Higher demand for the USD by non-U.S. banks from that economy is driving down the currency where the respective banks are headquartered. This finding coincides with the earlier findings in the literature that order flows as quantity proxies for demand are significant determinants of exchange rates contemporaneously (e.g., Evan and Lyons 2002).

What is strikingly different from the existing literature is the magnitude of the correlations. The proxy for USD safe asset demand, i.e., the USD asset share of non-U.S. banks, accounts for 36% of the variation in the USD exchange rate over time (in terms of the R2), which is quantitatively beyond anything that has been documented in the literature so far.

Though striking, Table 3 is just the beginning of our empirical investigation. We next turn to an instrumental variable approach that allows us to establish a causal relationship between the USD safe demand and the USD exchange rate in subsection B. Then we move to asset pricing in Section V and to forecasting relationships in Section VI.

B. Causal Relationship

We now turn to establish the causal interpretation between the demand for USD safe assets from non-U.S. banks and exchange rates using novel instrumental variables. It is a classical challenge in economics to separate supply and demand effects. If the demand and supply curves shift simultaneously over time, the observed data on relative quantities (USD assets as percent of total assets) and relative prices (exchange rates) reflect a set of equilibrium points on both curves.

Consequently, an OLS regression of relative prices on relative quantities fails to identify or to trace out either the supply or demand relationship. Exogeneous curve shifters (i.e., instrumental variables) can be used to address this problem.20 Furthermore, we extend this strategy to the U.S. Treasury premium and LIBOR cross-currency basis, which helps us identify different roles played by the “safe asset demand channel” and the “financial intermediation channel”, two channels proposed in the literature.

In the previous subsection, the underlying assumption for the OLS regression to obtain a consistent estimate of the supply curve is that the observed changes in quantities and prices are solely due to demand shifts. This might not be a very strong assumption for any given individual economy but might be too strong for their average as a whole. In this subsection, we try to adopt a conservative approach – find exogeneous demand shifters for USD assets of non-U.S. banks in each economy. A valid demand shifter should affect demand conditions of USD denominated assets exogenously without affecting cost conditions of supplying USD denominated assets. Based on this principle, we propose two groups of instrumental variables:

Safety of substitute currencies. The costs of substitutes are often used as demand shifters in the literature (Angrist and Krueger 2001). As the closest substitute for USD denominated assets to serve as global safe assets, we only focus on sovereign securities denominated in six non-USD currencies from non-U.S. G10 economies – that is, Canadian Dollar for Canada; Swiss Franc for Switzerland; Euro for Euro Area economies; Japanese Yen for Japan; Pound Sterling for United Kingdom; and Swedish Krona for Sweden. We consider two different variables that can capture the safety of assets denominated in substitute currencies:

  • i. The first measure is the Treasury premium of substitute economies, which is proportional to the convenience yield on substitute safe assets. This is built on the work of Du, Im, and Schreger (2018) that defines the U.S. Treasury premium, which is proportional to the convenience yield on U.S. safe assets of non-U.S. investors. Similarly, the Treasury premium of substitute economy, Φj,tsub, is equal to (yj,tGovtρj,tysub,tGovt), where yj,tGovtandysub,tGovt denote 1-year Treasury yield of economy j and that of substitute economy (all issued in respective currencies), and ρj,t is the 1-year market-implied forward premium for hedging currency j against the substitute currency. 21 By construction, it captures the premium that investors are willing to receive less from holding the substitute Treasury securities relative to government bonds in economy j.

  • ii. The second measure is the sovereign CDS spread of substitute Treasury securities, which capture the sovereign risk of substitute safe assets that are denominated in substitute currencies. In fact, sovereign risk is a key component that influences the Treasury premium (Du, Im, Schreger 2018; Engel and Wu 2018). We obtain the sovereign CDS spreads from Markit for government bonds with 5-year maturity (the most liquid maturity).

These two measures of safety of substitute currencies are simple averages across six non-USD G10 currencies, including Canadian Dollar, Euro, Japanese Yen, Pound Sterling, Swedish Krona, and Swiss Franc. We would expect when the Treasury premium (sovereign CDS spread) of substitutes increases (decreases), the substitute currencies become safer – the demand for USD safe assets from non-U.S. banks would be substituted more by demand for other safe assets denominated in substitute currencies.

Balance sheet constraints of non-U.S. banks. The risk-taking capacities of non-U.S. banks should matter for their demand for safe assets. We build on the work of Adrian, Etula, Shin (2010) and Adrian, Etula, Muir (2014) to measure the risk-taking capacities of non-U.S. banks as the logarithm of the book leverage ratio. We use the bank balance sheet data from Fitchconnect and match with the bank ownership information from Bankscope to make sure that the leverage ratios that we construct share the same economies as USD asset share of non-U.S. banks where they are headquartered. We orthogonalize this economy-specific bank leverage measure with respect to the leverage of the U.S. banking system, so that the residual is irrelevant to the pricing kernel of U.S. financial intermediaries. We would expect that when the (orthogonalized) leverage ratio of non-U.S. banks increases, balance sheet capacity deteriorates, and the demand for USD assets increases.

Table 4 reports our Two-Stage Least Squares (2SLS) results using the two groups of instrumental variables discussed above. As before, we focus on time-series regressions of the cross-sectional average of 16 economies over the full sample period.

Table 4.

Two-Stage Least Squares: USD Asset Share and Exchange Rates

article image
Notes: The table reports the 2SLS regression results for the contemporaneous relationship between the USD exchange rate and_the USD assets_share of non-U.S. banks. The upper panel summarizes the regression results for the first stage: ΔD¯$,t=α11+B11ΔΦ¯tsub+B21ΔSovereignCDS¯tsub+β31ΔLeveraget¯+β1X¯t+εt1, where the variable ΔD¯$,t denotes average quarterly changes in USD asset share of non-U.S. banks; ΔΦ¯tsub denotes the average quarterly changes in Treasury premium of substitute government bonds; ΔSovereignCDS¯tsub denotes the average quarterly changes in sovereign CDS spread of substitute government bonds; ΔLeverage¯ denotes the average quarterly changes of the log leverage ratio of non-U.S. banks, orthogonalized to U.S. bank leverage; Finally, Xt is a vector of control variables, including D¯$,t1,Φ¯t1,and(y¯t1Govty$,t1Govt). The lower panel summarizes the regression results for the second stage: Δs¯t=α12+β12ΔD¯^$,t+β22ΔΦ¯t+β32Δ(y¯tGovty$,tGovt)+β2X¯t+ϵt2, where the variable ΔD¯^$,t denotes fitted value of average quarterly changes in USD assets share of non-U.S. banks obtained from the first stage. Other variables are defined the same way as those in Table 3. The constants are not reported in the table for both stages. For testing for weak instrumental variables, F-statistics are reported. According to Stock and Yogo (2002), the critical value for the weak instrument test based on 2SLS bias of 5% significant level is 13.91 for parameters setting with 1 endogenous regressor, 3 instrumental variables, with the desired maximal bias of the IV estimator relative to OLS at 0.05. For endogeneity, the Durbin-Wu-Hausman F statistics are reported. For overidentifying restriction, the Sargan x2 statistic is reported. All series are cross-sectional averages over 16 currencies and valued at quarter end. The Newey-West heteroskedasticity-and-autocorrelation-consistent asymptotic standard errors are reported in parentheses. *, **, and *** denote significance levels at 10%, 5%, and 1%. The sample period is from 2001q2 to 2017q4.

In the upper panel of Table 4, we investigate the first stage relationship – between the USD asset demand from non-U.S. banks, captured by USD asset share, and three different instrumental variables proposed above.22 In particular, we run the following regression:23

ΔD¯$,t=α11+β11ΔΦ¯tsub+β21ΔSovereignCDS¯tsub+β31ΔLeveraget¯+β1X¯t+εt1(2)

From Columns (1) to (5), the coefficient point estimates on ΔΦ¯tsub are negative with similar magnitudes across all specifications, and significant at the 10% level or above for 3 out of 5 specifications; the coefficient point estimates on ΔSovereignCDS¯tsub are positive and significant at the 1% level with similar magnitudes across all specifications. This is consistent with our prior, suggesting that non-U.S. banks tend to reduce demand for safe USD assets when the safety and liquidity of substitute currencies increases. The coefficient point estimates on ΔLeveraget¯ are positive and significant at the 1% level with similar magnitudes across all specifications. This suggests that non-U.S. banks tend to increase demand for safe USD assets when their balance sheet capacity to bear risk deteriorates. In terms of economic significance, the variation from those three instrumental variables alone accounts for around 20% of the variation in the contemporaneous changes of USD asset demand from non-U.S. banks. Based on the critical values reported in Stock and Yogo (2002), we can reject the null hypothesis that the proposed variables are weak instruments for all specifications at the 5% significance level or above.24

The lower panel of Table 4 presents the results for the second stage. Again, we investigate the relationship between USD asset demand from non-U.S. banks and the USD exchange rate by running:

Δs¯t=α12+β12ΔD¯^$,t+β22ΔΦ¯t+β32Δ(y¯tGovty$,tGovt)+β2X¯t+εt2(3)

Specifications in Columns (1) to (5) are the same as the ones in Table 3, except instead of the observed values (ΔD¯$,t) we use the fitted values of USD asset demand (ΔD¯^$,t), which are purely driven by demand conditions of USD denominated assets of non-U.S. banks obtained from the first stage, but not by the supply of USD denominated assets. From Columns (1) to (5), the coefficient point estimates on ΔD¯^$,t are positive and significant at the 1% level across all specifications.

Note that the magnitude of the coefficient point estimates on ΔD¯^$,t obtained by 2SLS are around two times as larger as those obtained by OLS in Table 3, and we can reject that the USD asset share is an exogenous variable from all specifications at the 10% level or higher based on the F statistic of Durbin-Wu-Hausman test of endogeneity. This result suggests the slope of the supply curve is underestimated by the OLS estimation, because observed data in relative quantities (USD assets as a percent of total assets) and relative prices (exchange rates) is a reflection of a set of equilibrium points on both supply and demand curves. In addition, we confirm that three instrumental variables together are exogenous based on inference from the Sargan x2 statistic of overidentification. This result suggests that our instrumental variable (IV) strategy manages to identify the supply curve and corrects the coefficient point estimates using valid demand shifters that are exogeneous to supply conditions.

In terms of economic significance, the variation from the predicted changes in the USD asset share alone accounts for 7% of the variation in the contemporaneous changes of the USD exchange rate (Column 1 of Table 4), which is much smaller than that obtained from the OLS regression (Column 1 of Table 3) – that is 36%. In Column (4), the specification includes both covariates of ΔΦ¯tandΔ(y¯tGovty$,tGovt) along with the fitted value of the USD asset demand (ΔD¯^$,t), the R2 rises marginally to 11% but both ΔΦ¯tandΔ(y¯tGovty$,tGovt) are insignificant. However, after Column (5) includes all other potential drivers (the one quarter lags of D¯,Φ¯,and(y¯Govty$Govt) at quarter t – 1), the R2 rises to 36%, the point estimate on ΔD¯^$,t from 2SLS drops to around 41.8 (still almost two times as large as that from OLS), and ΔΦ¯t becomes marginally significant at the 10% level. As we discussed before, this specification is equivalent to include the AR (1) innovations of three factors discussed above. Our quantity measure of USD asset demand, captured by USD assets from non-U.S. banks, has more explanatory power on exchange rates in the form of AR (1) innovation from the 2SLS regression.

C. Transmission Channels

An interesting extension is to investigate the relationship between the U.S. Treasury premium, the LIBOR cross-currency basis, and the quantity of USD asset demand from non-U.S. banks using our IV strategy. This can add value to the discussion of the “safe asset demand channel” and the “financial intermediary channel” for the global USD cycle (Du, Im, and Schreger 2018; Jiang, Krishnamurthy, and Lustig 2018; and Du 2019). Both the LIBOR cross-currency basis, which equals the LIBOR CIP deviation, and the U.S. Treasury premium are measures of the “specialness” of USD assets relative to non-USD assets on a currency hedged basis. The key difference is that U.S. financial intermediaries represent key suppliers of USD denominated assets in the interbank market, while the U.S. Treasury is the ultimate supplier of U.S. Treasury securities in the U.S. Treasury market. As we discussed earlier, the “safe asset demand channel” is about shifters to the demand curve. The “financial intermediation channel” is about shifters to the supply curve. Using exogeneous demand shifters from our IV strategy, we can identify the slopes of the supply curves for these two markets.

Table 5 only presents the results of the second stage, because the first stage results are very similar to that of the upper panel in Table 4. The specification of the second stage is given by25:

ΔΦ¯t=α12+β12ΔD¯^$,t+β22Δ(y¯tGovty$,tGovt)+β2X¯t+εt2(4)

Where instead of using average quarterly changes in log nominal exchange rate as dependent variable, we investigate average quarterly changes of the U.S. Treasury premium (ΔΦtTreasury) and average quarterly changes of the (-1)*LIBOR cross-currency basis (ΔΦtLibor) in Columns (1) – (2) and Columns (3) – (4), respectively.26

Table 5.

USD Asset Share, U.S. Treasury Premium, and Cross-Currency Basis

article image
Notes: The table summarizes the second-stage results of the 2SLS regression of the U.S. Treasury premium on the USD asset share in Columns (1) to (2), and of the (-1)* cross-currency basis on USD asset share in Columns (3) and (4) with the specification: ΔΦ¯t=α12+β12ΔD¯^$,t+β22Δ(y¯tGovty$,tGovt)+β2X¯t+εt2, where ΔΦ¯t is the average quarterly changes in U.S. Treasury premium in Columns (1) and (2), o’r average quarterly changes in (-1)*cross-currency basis in Columns (3) and (4), respectively. The variable ΔD¯^$,t denotes the fitted value of average quarterly changes in USD asset share of non-U.S. banks obtained from the first stage. Other variables are defined as the same as those in Table 3. All specifications include a constant that is not reported in the table. All series are cross-sectional averages over 16 currencies and valued at quarter end. The Newey-West heteroskedasticity-and-autocorrelation-consistent asymptotic standard errors are reported in parentheses. *, **, and *** denote significance levels at 10%, 5%, and 1%. The sample period is from 2001q2 to 2017q4.

In the regressions of the U.S. Treasury premium (ΔΦtTreasury), Column (1) shows that the 2SLS coefficient point estimate on the fitted value of the USD asset share (ΔD¯^$,t) is positive and statistically significant during the full sample (2001q2–2017q4), and it remains significant and becomes even larger since the GFC (2010q1–2017q4) in Column (2). These results coincide with findings from Du, Im, and Schreger (2018), and Jiang, Krishnamurthy, and Lustig (2018) that document the U.S. Treasury premium has been positive for a long time even before the GFC. Because the fitted value of the USD asset share (ΔD¯^$,t) comes from exogeneous demand shifters, it suggests the supply curve in the U.S. Treasury market identified by our IV strategy is upward sloping for both the full sample episode and the post-GFC episode, though the slope is slightly larger for the latter but insignificant in difference. On the contrary, in the regressions of the (-1)* LIBOR cross-currency basis (ΔΦtLibor), Column (3) shows that the 2SLS coefficient point estimate on the fitted value of USD asset share is not significantly different from zero during the full sample (2001q2–2017q4), however, it turns positive and statistically significant after the GFC (2010q1–2017q4) in Column (4). These results again echo the findings from Du, Tepper, and Verdelhan (2018) that CIP deviation has only become significantly negative since the GFC. In the LIBOR cross-currency swap market, we cannot reject that the supply curve is flat for the full sample episode, but we find that the supply curve becomes significantly upward-sloping in the post-GFC episode. These results suggest that U.S. financial intermediaries actively traded in the LIBOR cross-currency swap market and arbitraged away the CIP deviation pre-GFC, but they have had less balance sheet capacities to do so since the GFC. In contrast, the U.S. Treasury never exhibits active speculative trading behavior even if it has the balance sheet capacity to do so, which is because the U.S. Treasury acts according to fiscal mandates.

As far as we know, our paper is the first one to provide comprehensive empirical evidence to understand the premise that the “safe asset demand channel” and the “financial intermediation channel” play different roles in different markets. Our findings suggest that the “safe asset demand channel” is at play for both the U.S. LIBOR cross-currency swap market and the U.S. Treasury market (Jiang, Krishnamurthy, and Lustig 2018; Krishnamurthy and Lustig 2019). In addition, our findings suggest that the “financial intermediation channel” has started to play an important role in the U.S. LIBOR cross-currency swap market, because the global capital and liquidity requirements and specific regulations at the individual economy level may have tightened the supply of US dollar funding to non-U.S. banks since the GFC (Du, Tepper, and Verdelhan 2018; Iida, Kimura, and Sudo 2018). Our findings also suggest that the “financial intermediation channel” doesn’t play a critical role in the U.S. Treasury market, because the U.S. Treasury (as the ultimate supplier of U.S. government bonds) can but does not actively arbitrage away the U.S. Treasury premium when the demand for U.S. Treasury from non-U.S. banks is strong.

Overall, our novel IV strategy helps identify the causal relationship between the USD asset share of non-U.S. banks and the USD exchange rate, correcting the downward bias of the OLS estimator. Further, it sheds new light on the “safe asset demand channel” and the “financial intermediation channel”.

V. Cross-Sectional Asset Pricing

In addition to the strong contemporaneous causal relationship, we turn to asset pricing in the cross-section. In this section, we find that different loadings on the average USD asset demand of non-U.S. banks help to explain currency excess returns in the cross section. Our results suggest that average global demand of USD safe assets from non-U.S. banks acts as a risk factor in pricing currency excess returns.

The risk factor of global USD asset demand is defined as the cross-sectional average of annualized quarterly changes in the USD asset share of non-U.S. banks across 16 economies. The log currency excess return of currency J corresponds to a long position in Treasury bonds of economy j funded by borrowing in U.S. dollars:

Rj,t=Δyj,t1GovtΔsj,ty$,t1Govt(5)

To estimate the factor price and the loading betas, we use a two-stage Fama and MacBeth (1973) regression. We conduct this analysis both at the individual-currency level and at the currency-portfolio level provided by Lustig, Roussanov, and Verdelhan (2011). In the first stage, we run a time-series regression of the currency excess returns on a constant and global USD asset demand changes as pricing factor to estimate betas. The betas for individual currencies (currency portfolios) are presented in Figure 3 below.

Figure 3.
Figure 3.

USD Asset Demand Loading

Citation: IMF Working Papers 2020, 101; 10.5089/9781513547732.001.A001

Notes: This figure reports the coefficient point estimates of ΔD$,t in the following regression specification: Rj,t = αj + βjΔD$,t + εj,t, where Rj,t is the currency excess return of individual currency J from investing in 1-year government bonds’in jurisdiction j and shorting 1-year U.S. Treasury Bill at end of quarter t (in the left Panel), or stands for that of currency portfolio (in the right Panel). ΔD$,t is the average quarterly changes in the USD asset share of non-US. banks across 16 jurisdictions at end of quarter t. The blue dots indicate the coefficient point estimates of β^J, for individual currencies (in the left panel) and for currency portfolios (in the right panel), respectively. The blue bars indicate 95% confidence interval bands. The excess returns of currency portfolios are from Lustig, Roussanov, and Verdelhan (2011). As Lustig and Verdelhan (2007) suggest, we drop the last currency portfolio, which consists currencies with very high inflation. The sample period is from 2001q2 to 2017q4.

From the left panel of Figure 3, we can find the loadings on global USD asset demand are negative and significant at the 5% level or above. It suggests that higher interest rate currencies (e.g., BRL and ZAR) have higher loadings (in absolute term) than lower interest rate currencies (e.g., JPY and GBP) on average. The currency portfolios from Lustig, Roussanov, and Verdelhan (2011) are ranked from low to high interest rate – portfolio 1 contains currencies with the lowest interest rate, while portfolio 5 contains currencies with the highest interest rate.27 In the right panel, we also find higher interest rate currency portfolios tend to have higher loadings in general.28

In the second stage, we examine the cross-sectional relationship between the magnitude of exposure to USD asset demand and the average size of excess returns. To visualize this cross-sectional relationship, Figure 4 plots the average currency excess returns against the corresponding USD demand betas for individual currencies (in the left panel), and for currency portfolios (in the right panel), respectively. We can see a strongly negative correlation between the average excess returns and the USD demand betas with a bivariate correlation equal to -58% for the individual currencies and -76% for the currency portfolios.

Figure 4.
Figure 4.

Currency Excess Returns and USD Asset Demand Betas

Citation: IMF Working Papers 2020, 101; 10.5089/9781513547732.001.A001

Notes: The vertical axis in the left panel shows the average excess return of currency J average over time, defined as investing in 1-year government bonds in jurisdiction j from shorting the 1-year U.S. Treasury, while the horizontal axis indicates the regression beta of running a quarterly regression of currency excess returns on average changes in the USD asset demand of non-U.S. banks for 16 currencies. The vertical axis in the right panel shows the average excess return of currency portfolios over time, while the horizontal axis indicates the regression beta of running quarterly regression of currency excess returns on average changes in USD assets demand of non-U.S. banks for 5 currency portfolios from Lustig, Roussanov, and Verdelhan (2011). As Lustig and Verdelhan (2007), we drop the last currency portfolio, which consists of currencies with very high inflation. The sample period is from 2001q2 to 2017q4.

From the left panel in Figure 4, we can see that investing in high interest rate currencies (e.g. BRL and ZAR) gives U.S. investors on average higher excess returns than investing in low interest currencies (e.g. JPY and GBP). At the same time, high interest rate currencies have larger negative loadings on the global average USD asset demand factor than low interest rate currencies. This suggests that the demand of USD assets of non-U.S. banks is a global risk factor on the FX market. High interest rate currencies have higher systematic loadings on the USD demand factor (a more negative USD asset demand beta), which require higher on average excess returns to compensate this exposure to USD asset demand risk. On the contrary, low interest rate currencies provide a hedge against USD asset demand risk.

The right panel for carry-trade portfolios ranked by interest rate demonstrates an even stronger relationship.29 This finding relates to the discussion of “currency crashes” by Brunnermeier, Nagel, and Pedersen (2009). It also coincides with the conclusion of Clarida, Davis, and Pedersen (2009) and Menkhoff, Sarno, Schmeling, and Schrimpf (2012) that the ex post returns for high interest rate currencies are low during times of high volatility. High interest rate currencies are risky because they have poor payoffs when a measure of global volatility is high, which coincides with strong demand for USD denominated safe assets.

The cross-sectional relationship discussed above has some linkages with the cross-sectional relationship between the broad U.S. dollar index and the LIBOR cross-currency basis. Avdjiev, Du, Koch, and Shin (2019) find currencies with more systematic exposures to the broad U.S. index have higher excess returns from CIP trade. When the global USD asset demand from non-U.S. banks rises, the USD exchange rate tends to appreciate. As a result, the currencies with more exposures to the USD asset demand risk tend to have higher expected excess returns from CIP trade.

Our finding is consistent with recent exchange rate theories based on capital flows in imperfect financial markets that links global imbalances with currency premia (Gourinchas and Rey 2007; Corte, Riddiough, and Sarno 2016). These theories suggest net debtor countries (usually high interest rate countries) offer currency risk premia to compensate investors willing to finance negative external imbalances because their currencies depreciate in bad times, which are times of large negative shocks to risk-bearing capacity and global risk aversion inducing safe demand for USD assets.

In summary, our findings suggest that the average USD asset demand from non-U.S. banks acts as a global risk factor that is priced in the cross section of currency excess returns. The currencies with more systematic exposures to the USD asset demand risk factor tend to have higher expected returns to compensate risks of “currency crashes” in bad times, which are times coincides with strong demand for safe USD assets.

VI. Predictability

In this section, we focus on the performance of the USD asset share from non-U.S. banks (a quantity measure) in forecasting exchange rates. In the literature, it is well known that some FX market quantities like order flow are associated with contemporaneous returns (e.g., Evans and Lyons 2002), however, it is less known whether predictive information is contained.30

To begin with, we illustrate the predictive power using a simple vector autoregression (VAR). The impulse response functions (IPFs) from the VAR are shown in Figure 5. In the left panel, we find that the contemporaneous appreciation in the USD exchange rate in response to positive shocks in USD asset share is nearly entirely undone after five years. One of the economic interpretations is overshooting due to time varying risk premia, where demand for USD assets of non-U.S. banks drives up the U.S. dollar contemporaneously by compressing risk premia, but this contemporaneous impact gets reversed over 1–5 years as risk premia normalize. Lilley, Maggiori, Neiman, and Schreger (2019) also find that the U.S. purchase of foreign bonds is associated with risk premia after the GFC.

Figure 5.
Figure 5.

Impulse Responses from a Var

Citation: IMF Working Papers 2020, 101; 10.5089/9781513547732.001.A001

Notes: This figure presents impulse responses from a four-variable VAR in the USD asset share, the U.S. Treasury premium, the interest rate differential (home – U.S.), and the exchange rate. All series are cross-sectional average over 16 economies and valued at quarter end. An increase in the exchange rate means that the U.S. dollar appreciates against a basket of 16 currencies. The left panel presents the exchange rate response to a USD-asset-share shock. The middle panel shows the exchange rate response to a U.S. Treasury premium shock. The right panel shows the exchange rate response to an interest rate differential shock. The VAR is estimated in levels and the shocks are identified using a Cholesky decomposition with the following ordering: exchange rate, U.S. Treasury premium, interest rate differential, USD assets share. Red dashed lines represent 95% confidence intervals and blue solid lines represent impulse response function. The sample period is from 2001q2 to 2017q4.

An alternative explanation is that the USD asset share of non-U.S. banks could capture a dimension of investor sentiment that is not captured by other macro fundamentals. For example, López-Salido, Stein, and Zakrajšek (2017) and Brandão-Marques, Chen, Raddatz, Vandenbussche, and Xie (2019) document that risk metrics from credit markets can help to predict reversals of financial conditions and corporate spreads. In addition, this can also reflect the extraordinary ability of non-U.S. banks to process fundamental information, especially for the local market. For example, Menkhoff, Sarno, Schmeling, Schrimpf (2016) find that customer types vary markedly in terms of their predictive ability – the flows by long-term demand-side investment managers have the strongest predictive power for exchange rates, while flows originated from the other groups only predict transitory changes in exchange rates. However, fully disentangling the extent to which the relationship between USD asset share of non-U.S. banks and the USD exchange rate by any of these dimensions is beyond the scope of this paper.

In the middle panel, we find that the U.S. Treasury premium drives up the U.S. dollar contemporaneously and the induced momentum effect is short-lived for another quarter, then it depreciates over the next one-to-five-year period. This result is consistent with findings from Jiang, Krishnamurthy, and Lustig (2018). In the right panel, we find the U.S. dollar appreciates in response to positive shocks to interest differentials (home – U.S.) in the first two years, then it wanes down over the next three-year period.31

A. In-sample Forecasting

As suggested by the theoretical models by Engel and Wu (2018) and Jiang, Krishnamurthy, and Lustig (2018), a strong USD asset demand from non-U.S. banks today means that today’s dollar exchange rate appreciates, thus lowering the expected future returns from owning USD safe assets and inducing a subsequent depreciation in the future.32 In particular, we run the following regression:33

Δsj,th=α1+β1Dj,$,t+β2Φj,t+β3(yj,tGovty$,tGovt)+ϵt+h(6)

where the dependent variable Δsj,th is the annualized change in the log nominal exchange rate of USD vis-à-vis currency J, an increase of which means that the USD appreciates against currency J over the horizon of [t, t + h]. The variable Dj,$,t denotes the USD asset share of non-U.S. banks from economy j at end of quarter t. The variable Φj,t denotes the U.S. Treasury premium vis-à-vis economy j at the end of quarter t. The variable (yj,tGovty$,tGovt) denotes the interest differential between 1-year Treasury securities of economy ) and that of the U.S. Treasury at end of quarter t. We run this regression using quarterly data as before, but vary by horizons from one-quarter, one-year, two-year, three-year, to five-year.

Table 6 reports our forecasting results of the U.S. dollar exchange rates against a basket of 16 currencies. The USD asset share doesn’t help predict the USD exchange rate at the short end but has very strong predictive power over the longer horizon. This suggests time variation in exchange rate risk premia. In particular, it reflects the stationary but persistent deviation from the uncovered interest parity that Duarte and Stockman (2005) and Engel (2014 and 2016) have documented. The coefficient point estimates on the USD asset share of non-U.S. banks (D¯$,t) start from -0.27 (not significant) at the one-quarter horizon, and then stabilizes to a level between 2.02 and 2.75 per annum from one-year to five-year horizon at 5% significance level or above. This means that a positive shock to USD asset share by one-standard deviation (0.33% from Table 1) is contemporaneously correlated with an appreciation of the U.S dollar by 8.1% (Column 5 of Table 3) per annum in that year on average, but it is expected to depreciate by 0.67% (Column 5 of Table 6) per annum over the next five years on average to wane down the overshooting.

Table 6.

Forecasting Exchange Rates

article image
Notes: This table reports OLS results for in-sample forecasts of annualized percentage changes in the U.S. dollar exchange rates against a basket of 16 currencies at the one-quarter, one-year, two-year, three-year and five-year horizons. The specification is: Δsj,th=α1+β1Dj,$,t+β2Φj,t+β3(yj,tGovty$,tGovt)+ϵt+h, where Δsj,th is the annualized future change in the log of nominal exchange rate of the USD vis-à-vis currency J in economy j over the horizon [t, t + h], an increase of which means that the USD appreciates against currency J over the period [t, t + h]. The variable DJ,$,t denotes the USD asset share by non-U.S. banks from economy j at end of quarter t. The variable Φj,t denotes the U.S. Treasury premium vis-à-vis economy j at end of quarter t. The variable (yj,tGovty$,tGovt) denotes the interest differential at end of quarter t. We run this regression using quarterly data as before but vary the horizons ahead from one-quarter (h = 1), one-year (h = 4), two-years (h = 8), three-years (h = 12), to five-years (h = 20). All series are cross-sectional averages over 16 currencies and valued at quarter end. All specifications include a constant (not reported). The Newey-West heteroskedasticity-and-autocorrelation-consistent asymptotic standard errors are reported in parentheses. *, **, and *** denote significance levels at 10%, 5%, and 1%. The sample period is 2001q2 to 2017q4.

As Jiang, Krishnamurthy, and Lustig (2018), we also find the induced momentum effect from the U.S. Treasury premium (Φ¯t) is short-lived and the coefficient point estimate switches sign over the long run, although it is only statistically significant at the two-year horizon.

The Uncovered Interest Rate Parity (UIP) suggests the coefficient before interest differential equals to 1. However, numerous empirical studies consistently reject the UIP condition (see Engel 1996; Sarno 2005). This is the case if we only include non-U.S. G10 economies in our sample.34 However, the interest differential (y¯tGovty$,tGovt), which is a proxy for the risk premium of carry trade excess returns, positively and significantly forecasts the USD exchange rate appreciation for one-year, two-year, and five-year horizons at 10% significance level or above. At short horizon, all these variables don’t help predict the dynamics of the U.S. dollar exchange rate, where the one-quarter ahead R2 is as low as 6%. Gradually, these regressors jointly account for more variation in the moves of exchange rate at longer horizon – the R2 increases to 20% at the one-year horizon and strikingly reaches 67% at the five-year horizon.

In Table 7, we report our forecasting regression results of bilateral exchange rate over the five-year horizon. For better comparison, a simple univariate regression is used here:

Δsj,t20=α1+β1Dj,$,t+ϵt+20(7)

In Panel A, the regressor is Dj,$,t, that is the individual USD asset share of non-U.S. banks in economy) at end of quarter t. The forecasting results of the bilateral exchange rate confirms our finding: the USD asset share from each economy predicts the USD exchange rate at the 5% significance level or above for 10 out of 16 currencies. In Panel B, the regressor is D¯$,t, that is the cross-sectional average of USD asset shares over 16 economies at end of quarter t. The forecasting power is even stronger – it helps predict the bilateral exchange rate at the 5% significance level or above for all 16 currencies in our sample. The point estimates on D¯$,t vary from as low as 0.59% per annum (for MXN) to as high as 5.46% per annum (for MYR). While, the R2 varies from 10% (for MXN) to 69% (for MYR), which is on-average higher than that of Panel A using the USD asset share from individual economies as regressors. These results suggest that it is the common, global component in the USD asset share of non-U.S. banks that forecasts the USD exchange rate against individual currencies.

Table 7.

Forecasting Dollar Exchange Rates Against Individual Currency

article image
Notes: This table reports OLS regression results for in-sample forecasts of annualized percentage changes in the bilateral USD exchange rate at the five-year horizon. The univariate regression specification is: Δsj,t20=α1+β1Dj,$,t+ϵt+20, where Δsj,t20 is the annualized change in the log bilateral USD exchange rate against currency J in economy j over the horizon of [t, t + 20], an increase of which means USD appreciates against currency J. The variable Dj,$,t denotes USD asset share of non-U.S. banks from economy j at end of quarter t. In Panel A, the regressor is the USD asset share of non-U.S. banks from individual economy. In Panel B, the regressor is the cross-sectional average over the 16 currencies. All specifications include a constant (not reported). The Newey-West heteroskedasticity-and-autocorrelation-consistent asymptotic standard errors are reported in parentheses. *, **, and *** denote significance levels at 10%, 5%, and 1%. The sample period is 2001q2 to 2017q4.

Taken together, these results strongly suggest that exchange rate movements are associated with time varying risk premia, which covary with the demand for USD denominated safe assets by non-U.S. banks. The overshooting of Sections IV and VI.A. suggests that demand for USD assets leads to a sharp contemporaneous appreciation of the USD, which is reverted slowly over time. In addition, the results of cross-sectional asset pricing in Section V points to a coherent intermediary asset pricing story. The scarcity of safe USD assets is driven to an important extent by risk-taking capacity of non-U.S. banks, which enters the pricing kernel of these marginal investors on the FX market, thus driving time varying USD risk premia. We next turn to the out-of-sample exercise, where we show that the strong predictability holds even out of sample, thus, undermining one of the cornerstones of the exchange rate literature (Meese and Rogoff, 1983).

B. Out-of-sample Forecasting

In this subsection, we turn to the out-of-sample forecasting performance following the tradition established by Meese and Rogoff (1983). We find our forecasting model to significantly outperform the random walk prediction.

We consider an out-of-sample forecasting exercise with a period of K quarters to obtain in-sample coefficient point estimates and a period of P quarters to evaluate the model’s performance out of sample. We denote the end of the model evaluation as quarter T. To estimate model parameters, we regress the annualized change of the USD exchange rate (over an h- quarter horizon) only on the USD asset share to obtain in-sample estimates of the model parameters within a window spanning K quarters:

Δsj,th=α1+β1Dj,$,t+ϵt+h(8)

Where, t ε [T – P – K, T – P]. We then use the estimated coefficients α^1andβ^1, and the realized observation of the USD asset share to forecast the exchange rate change in the next period:

Δs^j,TP+1h=α^1+β^1Dj,$,t(9)

The model’s forecast error is the difference between the realized exchange rate change and the model-implied forecast:

errorj,TP+1h=Δsj,TP+1hΔs^j,TP+1h(10)

We repeat this procedure for P times using a rolling regression approach until we have an out-of-sample forecast for all quarters in the period between quarters T – P to T. We compute the mean-square-error (MSE) of our forecasting model and compare it against that of the random walk benchmark. First, we adopt the out-of-sample (OOS) statistic from Diebold and Mariano (1995) and West (1996) based on MSEr – MSEU to evaluate performance of the alternative model. Because the random walk benchmark is nested in the “unrestricted” specification, we can evaluate its performance using the inference procedure by Clark and West (2006) to adjust difference in mean squared errors: MSEr – (MSEU – Adj.), which accounts for the small-sample forecast bias.

Table 8 presents the out-of-sample forecasting performance of our exchange rate forecast model for the five-year horizon (h = 20). 35 In Panel A, the Diebold-Mariano difference in MSEs and the OOS-T statistic are reported.36 In Column (1), we report the result that uses the cross-sectional average series over 16 currencies. We find our forecast model outperforms the benchmark random walk model at the 1% significance level.

Table 8.

Out-of-sample Forecastability

article image
Notes: This table investigates the out-of-sample forecastability of the U.S. dollar exchange rates at the five-year horizon. We compare the forecasting performance of a model using the USD assets share as predictor against a benchmark random walk model. To estimate model parameters, we regress the annualized percentage changes in the USD exchange rate on the USD asset share to provide in-sample estimates in the period of [T – P – K, T – P] with the specification: Δsj,t20=α1+β1D$,t+ϵt+20, where Δst20 is the annualized percentage change in the USD exchange rate against currency J between quarters t and t+20, increases of which means that the USD appreciates against currency J over the horizon [t, t + 20]. The variable % is the cross-sectional average of the USD assets share of non-US. banks over 16 economies at quarter t. The estimated coefficients and the realized observations of the USD asset share are used to forecast the exchange rate change in next period. We repeat this procedure P times using a rolling regression approach until we have an out-of-sample forecast for all quarters in the period between T – P to T. We compute the mean-square-error (MSE) of our forecasting model and compare it against that of the random walk benchmark. Panel A reports the Diebold-Mariano difference in MSEs and the OOS-T statistic. Panel B reports the Clark-West adjusted difference in MSEs and the C-W statistic. The heteroskedasticity and autocorrelation consistent (HAC) standard errors are estimated to compute the OOS-T statistics and C-W statistics (Clark and West 2006). The corresponding 90%, 95%, and 99% critical values are 0.780, 1.111, and 1.784 (using κ2 = 1 and π = 0.2) obtained from Table 1 in McCracken (2007). In Column (1), the USD exchange rate against the cross-sectional average over 16 currencies is used. From Columns (2) – Column (17), the USD exchange rate against individual currency is used. All series are valued at quarter end. *, **, and *** denote significance levels at 10%, 5%, and 1%. The sample period is 2001q2 to 2017q4.

In Columns (2) to (17), we repeat this exercise for individual economy. We find that our forecasting model outperforms the benchmark random model for 11 out of 16 currencies at the 10% significance level or above. In Panel B, the Clark-West adjusted difference in MSEs and the C-W statistic are reported. Similarly, our forecasting model outperforms the benchmark random walk model at the 1% significance level for the USD exchange rate against a basket of 16 currencies in Column (1). More strikingly, our forecasting model outperforms the benchmark model in 15 out of 16 currencies at 10% significance level or above after adjusting for the small-sample forecast bias from Columns (2) to (17).37 To our knowledge, the strength of this out-of-sample forecasting power from a quantity variable is unprecedented in the literature on exchange rates.

Overall, our results suggest a strong USD asset demand from non-U.S. banks today means that the USD exchange rate appreciates contemporaneously, which induces an expected depreciation in the future. This is an overshooting result in the spirit of Dornbusch (1976). However, the economic interpretation is one relying on USD asset scarcity that is driven largely by non-U.S. banks as the marginal investor in the FX market. We also find the intermediary asset pricing factor (e.g. the leverage of non-U.S. banks) to be one of the candidate explanatory variables. Time varying FX risk premia are related to the demand of USD denominated safe assets from non-U.S. banks in a systematic fashion, thus creating forcastability of the exchange rate. The forecasting power of the USD asset share of non-U.S. banks in exchange rates over longer horizons is robust for both in-sample and out-of-sample regressions.

VII. Robustness

In this section, we discuss the robustness of our analysis. First, we tackle data limitations in the contemporaneous relationship between changes in the USD asset share and changes in the USD exchange rate. Second, we turn to our identification strategy and the robustness of our instrumental variables. Third, we compare the forecastiblity of our model with existing exchange rate prediction models.

A. Measurement

We want to acknowledge a caveat in our measure due to data limitations. Ideally, we would capture the relative demand for USD assets by non-U.S. banks of each economy from a pure quantity perspective. However, we can only measure the aggregated value of USD assets over aggregated values of total assets of non-U.S. banks at the nationality level, because the data we use do not allow us to separate the quantity of assets purchased from the exchange rates and prices at which they are purchased. As a result, contemporaneous explanatory power in exchange rate movements may reflect the mechanical influence of the exchange rate on the value of USD assets purchases.

We tackle this in three steps. First, we adjust for changes in outstanding claims using adjustments for all series obtained from BIS Locational Banking Statistics (LBS) throughout the whole paper. This helps eliminate the impact of methodological changes and exchange rate movements for those components. Second, the local positions in local currencies are obtained based on a bottom-up consolidated approach using the FitchConnect and Factset databases and the share of local currency position to all currency positions is based on the Other Depository Corporations (ODCs) survey, which is based on residency basis as opposed to nationality basis. We further avoid bias in the currency breakdown by not adjusting for exchange rate movements of the local positions in local currencies in the analysis of the contemporaneous relationship. Third, in this subsection, we investigate the sensitivity of the contemporaneous relationship when we adjust the exchange rate movements for local positions in local currencies. We follow the same method as the BIS Locational Banking Statistics (LBS) uses to adjust for exchange rate movements.38

Appendix Table 1 reports the results after we adjust for the exchange rate movements in the local positions in local currencies.39 For the cross-sectional average across 16 currencies, the contemporaneous relationship between quarterly changes in the USD asset share of non-US banks and quarterly changes in log USD exchange rate still holds. However, the magnitudes decline by around 30% relative to the ones without adjustment (see Appendix Table 3). Further, the contemporaneous relationship between USD asset share of non-US banks and USD exchange rate becomes insignificant prior to the Global Financial Crisis. For individual currencies, this contemporaneous relationship holds at the 10% significance level or above for 8 out of 16 individual currencies. We conjecture that these results are weaker due to the measurement error in the local currency adjustment.

Lilley, Maggiori, Neiman, and Schreger (2019) also acknowledge this limitation of aggregated data and explore the pure quantities of U.S. foreign bond flows using the U.S. mutual fund holdings data from Morningstar, which does allow decomposition of market-value positions into exchanges rates, prices and quantities. Their finding confirms that the reconnect between the USD exchange rate and U.S. foreign bond flows (denominated in USD) since the GFC is either not due to the mechanical influence of exchange rate or due to the influence of price on the value of U.S. foreign bond purchases. There are two key distinctions between our paper and Maggiori, Neiman, and Schreger (2019): 1) the power of USD asset share of non-U.S. banks in explaining the U.S. dollar dynamics is not restricted to the post-GFC sample; 2) we emphasize the role of marginal investors played by non-U.S. banks on the FX market rather than focus on the U.S. mutual funds. We view the two approaches as complementary.

B. Validity of IVs

We would like to acknowledge several drawbacks in terms of the validity of our identification strategy. First, our basic challenge is that observed changes in the USD asset share reflects both the demand and supply for USD asset share. This becomes worrisome especially for large economies, because large economies not only demand USD assets but also act as USD asset providers. For example, large European banks and Japanese banks are major USD liquidity providers in the USD funding market (see McGuire and von Peter 2012). Second, the validity of the IVs depends on the assumption that the IVs are exogeneous to the supply conditions of USD assets.

In this subsection, we propose to resolve those concerns by refining our IV strategy. First, we could restrict our sample in the 2SLS analyses to smaller countries which act more like price takers compared to major economies in the global USD market. Second, we consider to only use sovereign CDS as our IV. The non-U.S. banks’ balance sheet capacity is challenged by its correlation with non-US banks’ propensity to lend in USD. In addition, FX swaps supplied by banks, which are often subject to bank balance sheet constraints, are used to price the Treasury premium, so the Treasury premium may be correlated with the supply of USD assets. Third, instead of using the safety measures of six G10 currencies, we only focus on the safety measures for Canadian Dollar, Swiss Franc, and Swedish Korona as the substitute currencies. Compared to Euro Area, Japan, and the United Kingdom, those three economies are relatively small and less prominent in providing USD assets. As a result, the correlation between the USD asset share of non-US banks to the safety measures of those three economies are more likely related to substitute safety of those three currencies relative to USD rather than the supply of USD.

Appendix Table 2 reports the results of the 2SLS. In the upper panel, we present the results for the first stage with the refined IV strategy. The coefficient point estimates on ΔSovereignCDS¯tsub are positive and significant at the 1% level with similar magnitudes across all specifications. Based on the critical values reported in Stock and Yogo (2002), we can reject the null hypothesis that the proposed IV is a weak instrument for 4 out of 5 specifications at the 5% significance level or above. In the lower panel, we present the results for the second stage with refined IV strategy. Again, the coefficient point estimates on ΔD¯^$,t are positive and significant at the 1% level across all specifications. Note that the magnitude of the coefficient point estimates on ΔD¯^$,t obtained by the refined 2SLS are around three times as larger as those obtained by OLS in Appendix Table 3, and we can reject that the USD asset share is an exogenous variable from all specifications at the 10% level or higher based on the F statistic of Durbin-Wu-Hausman test of endogeneity. This result suggests the slope of the supply curve is underestimated by the OLS estimation and the refined IV strategy corrects the coefficient point estimates even further.

C. Forecastibility

We now turn to the robustness in forecastability of exchange rates. We compare the forecasting performance of our model with 9 exchange rate forecasting models discussed in the literature since the 1970s. More specifically, we compare the forecasting performance of augmented versions that include the USD asset share as an additional predictor in existing model discussed in the literature.

Guided by the excellent survey of exchange rate predictability in Rossi (2013), the models we consider for the horse race include: (1) the UIP model; (2) the monetary model with flexible prices (“Frankel-Bilson” model); (3) the monetary model with sticky prices (the “Dornbusch-Frankel” model); (4) the model with productivity differentials ( the “Balassa-Samuelson” model); (5) the Taylor rule model; and (6) the net foreign asset model (the “Gourinchas-Rey” model). In addition, we take into account of the most recent development in exchange rate forecast including: (7) the dollar liquidity model suggested by Adrian, Etula, and Shin (2010); (8) the Treasury premium model suggested by Jiang, Krishnamurthy, and Lustig (2018); and (9) the U.S. foreign bond flow model suggested by Lilley, Neiman, Maggiori, and Schreger (2019).

Appendix Table 3 compares the out-of-sample forecasting performance of the augmented exchange rate forecasting model by adding the USD asset share as an additional predictor to existing model against each existing model stated above. The Diebold-Mariano difference in MSEs and the OOS-T statistic are reported for each comparison. We find the MSEexisting – MSEaugmented is positive across all models and that the OOS-T statistic is significant at the 10% level or above across all 9 comparisons. This suggests that including the USD asset share as an additional predictor improves the performance of out-of-sample forecasting for each of the exchange rate models discussed above.

In summary, this section confirms that the contemporaneous relationship between changes in the USD asset share of non-U.S. banks and changes in the USD exchange rate hold if the exchange rate movements in the local positions in local currencies are taken into consideration, even though the relationship is somewhat weaker and also we could not fully purely isolate the quantity from value due to data limitations. In addition, the refined IV strategy resolves several concerns on the validity of IVs discussed above and give us more confidence on the causal relationship between contemporaneous changes in the USD safe asset demand and changes in the USD exchange rate. Finally, our findings suggest that the average USD asset demand from non-U.S. banks measured by the cross-sectional average of USD asset share from non-U.S. banks across 16 economies helps to improve the forecasting performance of 9 existing exchange rate models discussed in the literature since the 1970s. Clearly, exchange rates are deviating significantly from the random walk benchmark as a function of the USD asset demand share of non-U.S. banks, thus pointing strongly towards exchange rate overshooting as a function of time varying intermediary risk premia.

VIII. Implications for the Literature

We summarize our paper’s key implications for four strands of the literature.

First, our paper contributes to the literature on the United States’ special role as the provider of world safe assets (e.g., Gourinchas and Rey 2007; Gourinchas, Rey, and Govillot 2010; Farhi and Maggiori 2018). Maggiori (2017) shows a model with a simple asymmetry in global risk sharing and heterogeneity in financial development can rationalize the economic role of the U.S. in the global financial architecture. As financial intermediaries in the U.S. are better able to deal with funding problems following negative shocks, the United States consumes more relative to the rest of the world (RoW) and runs a trade deficit based on higher financial income that it earns as compensation for greater risks it takes. On the contrary, the financial intermediaries in the RoW accumulate precautionary long positions in USD safer assets in order to insulate their capitals from negative shocks. When bad time hits, capital losses on the external portfolio of the U.S. lead to a wealth transfer to the RoW. Naturally, the U.S. dollar emerges as the reserve currency because it appreciates during bad times, thus USD denominated assets represent global safe assets by providing a good hedge.40 Our paper supports this theory by providing empirical evidence that the financial intermediaries from the ROW pay a counter-cyclical safety premium to hold USD denominated safe assets. In addition, we demonstrate that demand for USD safe assets from financial intermediaries of the ROW increases strongly following negative shocks because of flight to safety, and the USD exchange rate appreciates contemporaneously, thus lowering the investors’ expected future return from owning U.S. safe assets.

Second, our paper sheds light on the large literature on the “exchange rate disconnect puzzle”. Rossi (2013) gives an excellent review of exchange rate predictability especially for models before the first decade of the 21st century. Our paper is directly related to recent developments in this area from two perspectives. On the one hand, foreign investors’ demand for safety, empirically captured by the Treasury premium, is linked to movements in exchange rates (e.g., Engel and Wu 2018; Jiang, Krishnamurthy, and Lustig 2018).41 On the other hand, FX market quantities, like order flows, can help understand currency returns (e.g., Evans and Lyons 2002, 2005; Froot and Ramadorai 2005; Rime, Sarno, and Sojli 2010). Lilley, Maggiori, Neiman, and Schreger (2019) use security-level data on U.S. portfolios and demonstrate that the reconnect of U.S. foreign bond purchases to exchange rates is largely driven by investment in USD-denominated assets. Our paper links these two directions together and finds that a quantity-based measure can capture the foreign safe asset demand, especially for the marginal investors. Furthermore, we use a novel instrumental variable approach to identify a causal relationship between safe asset demand from non-U.S. banks and the U.S. dollar exchange rate. Last but not least, our paper documents the demand for safety in USD assets from non-U.S. banks is associated with time-varying currency risk premia, thus it not only helps account for contemporaneous changes in exchange rates, but also helps predict exchange rate dynamics over the longer horizon both in sample and out of sample.

Third, our paper also contributes to the literature on the twin sisters of “the global financial cycle” and “the global USD cycle”. In an early version of Avdjiev, Du, Koch, and Shin (2019), they present a global banking model with a VaR constraint and show that a stronger USD coincides with a higher shadow cost of banks’ balance sheet capacity, so it can price the cross-section of CIP deviations and is associated with lower growth of cross-border bank lending denominated in the USD. Valchev (forthcoming) provides an open-economy model to relate the quantity of U.S. Treasury bonds outstanding to the U.S. Treasury premium, to the return of the U.S. dollar, and to the failure of CIP, jointly. Our paper adopts an instrumental variable approach and establishes a causal relationship between USD asset demand from non-U.S. banks, the U.S. dollar exchange rate, the U.S. Treasury premium, and the LIBOR CIP deviation. As far as we know, our paper is the first one that provides empirical evidence that balance sheet constraints of non-U.S. financial intermediaries play a critical role in the U.S. interbank market but not in the U.S. Treasury market by identifying different slopes of supply curves from both markets, respectively. In other words, we differentiate the importance of the “safe asset demand” channel from the “financial intermediation channel” in these two markets, which has been a premise in the literature (e.g., Jiang, Krishnamurthy, and Lustig 2018; Du 2019; and Krishnamurthy and Lustig 2019).

Fourthly, our paper also sheds light on the theory of intermediary asset pricing (e.g., Brunnermeier and Pedersen 2009; He and Krishnamurthy 2012, 2013; Adrian and Shin 2014; Brunnermeier and Sannikov 2014). Adrian, Etula, Shin (2010) provide a theoretical foundation for a funding liquidity channel in a global banking model and show that USD funding liquidity forecasts exchange rates because of its association with time-varying risk premia. Our paper focuses on the other side of the mirror – the USD denominated assets of non-U.S. banks rather than the USD denominated liabilities of the U.S. banking sector. As the marginal investors in the FX market especially for the local currency (by regarding USD as numéraire), our paper demonstrates non-U.S. banks’ leverage significantly accounts for the changes in their demand for USD denominated assets. It suggests exchange rate fluctuations are associated with the USD asset demand from non-U.S. banks, which is a function of non-U.S. banks’ balance sheet capacity. Furthermore, we contribute to cross-sectional asset pricing theory by finding evidence that average USD asset shares of non-U.S. banks can help explain the variation in currency excess returns.

IX. Conclusion

In this paper, we construct a measure of USD asset demand, the USD asset share of non-U.S. banks, to explain exchange rate movements. First, we use this variable to explain contemporaneous exchange rate movements. Second, we establish a causal relationship between the USD asset demand, the U.S. dollar exchange rate, the U.S. Treasury premium, and the LIBOR cross-currency basis. We distinguish the “safe asset demand channel” from the “financial intermediation channel” with our instrumental variable strategy. Third, we find that the USD asset demand of non-U.S. banks at the global level can help explain currency excess returns in the cross section. Fourth, the USD asset demand variable helps forecast exchange rates both in-sample and out-of-sample. For all of these four empirical exercises, we obtain statistically highly significant and economically large results. More importantly, the four empirical results taken together point to a coherent economic mechanism that significantly advances the exchange rate literature.

We find that the demand for the USD by non-U.S. banks drives up the value of the U.S. dollar contemporaneously with high statistical significance and in an economically large magnitude. This appreciation of the U.S. dollar reverts over the next five years. The reversal of the currency value leads to strong forecastability in exchange rates – a positive shock to the USD asset share by one-standard deviation (around 0.33 percent from Table 1) is associated with appreciation of the USD by 8.1 percent per annum on average in that same year (Column (4) of Table 3), and forecasts a depreciation of the USD by an average of 0.67 percent per annum over the next five years (Column (5) of Table 6). Importantly, the forecasting R2s are very large, with in sample R2s of 36 percent at the two-year horizon and 67 percent at the five-year horizon. This is an overshooting result in the spirit of Dornbusch (1976): a USD asset demand shock by non-U.S. banks drives up the dollar contemporaneously via a compression of the risk premia, and that compression then reverts over the coming five years. The economic interpretation is one relying on USD asset scarcity that is driven largely by non-U.S. banks as the marginal investor on the FX market, especially for local currencies.

The in-sample forecasting results are strongly suggestive of a risk premium mechanism. Time varying FX risk premia are related to the demand of USD denominated safe assets from non-U.S. banks in a systematic fashion, thus creating forcastability of the exchange rate. Indeed, in the cross-sectional Fama-MacBeth regressions, we find that currencies and carry-trade portfolios earn significant excess returns with respect to higher loadings on the global USD asset demand factor. Put differently, the price of risk of the USD asset demand factor is negative, and relatively high interest rate currencies (or high carry portfolios) have more negative exposure to the USD asset demand factor, thus earning a positive and large excess return on average.

Importantly, the strength of the out-of-sample forecasting power from a quantity variable is unprecedented in the literature on exchange rates. We clearly reject the random walk hypothesis, thus again confirming that a time varying risk premium embedded in the USD asset demand of non-U.S. banks is the lead explanation for our findings. Importantly, we significantly improve relative to 9 traditional exchange rate forecasting models that have been presented in the literature since the 1970s, suggesting that our USD asset demand channel in forecasting exchange rates is very robust.

References

  • Adrian, Tobias, Erkko Etula, and Hyun Song Shin. 2010. “Risk Appetite and Exchange Rates.” Federal Reserve Bank of New York Staff Report 361.

    • Search Google Scholar
    • Export Citation
  • Adrian, Tobias, and Hyun Song Shin. 2014. “Procyclical Leverage and Value-at-Risk.” Review of Financial Studies 27 (2): 373403.

  • Adrian, Tobias, Erkko Etula, and Tyler Muir. 2014. “Financial Intermediaries and the Cross-Section of Asset Returns.” Journal of Finance 69 (6): 25572596.

    • Search Google Scholar
    • Export Citation
  • Angrist, Joshua D., and Alan B. Krueger. 2001Instrumental Variables and the Search for Identification: From Supply and Demand to Natural Experiments.” Journal of Economic Perspectives 15 (4): 6985.

    • Search Google Scholar
    • Export Citation
  • Avdjiev, Stefan, Wenxin Du, Catherine Koch, and Hyun Song Shin. 2019. “The Dollar, Bank Leverage, and Deviations from Covered Interest Parity.” American Economic Review: Insights 1 (2): 193208.

    • Search Google Scholar
    • Export Citation
  • Brandão-Marques, Luis, Qianying Chen, Claudio Raddatz, Jérôme Vandenbussche, and Peichu Xie. 2019. “The Riskiness of Credit Allocation and Financial Stability.” International Monetary Fund Working Paper 19/207.

    • Search Google Scholar
    • Export Citation
  • Brunnermeier, Markus K., Stefan Nagel, and Lasse H. Pedersen . 2009. “Carry Trades and Currency Crashes.” In NBER Macroeconomics Annual 2008, edited by Daron Acemoglu, Kenneth Rogoff, and Michael Woodford, 31347. Chicago: University of Chicago Press.

    • Search Google Scholar
    • Export Citation
  • Brunnermeier, Markus K., and Lasse Heje Pedersen. 2009. “Market Liquidity and Funding Liquidity.” Review of Financial Studies 22 (6): 22012238.

    • Search Google Scholar
    • Export Citation
  • Brunnermeier, Markus K., and Yuliy Sannikov. 2014. “A Macroeconomic Model with a Financial Sector.” American Economic Review 104 (2): 379421.

    • Search Google Scholar
    • Export Citation
  • Burnside, Craig, Martin Eichenbaum, and Sergio Rebelo. 2007. “The Returns to Currency Speculation in Emerging markets.” American Economic Review 97 (2): 333338.

    • Search Google Scholar
    • Export Citation
  • Burnside, Craig, Martin Eichenbaum, and Sergio Rebelo. 2011. “Carry Trade and Momentum in Currency Markets.” Annual Review of Financial Economics 3 (1): 511535.

    • Search Google Scholar
    • Export Citation
  • Cerutti, Eugenio, Stijn Claessens, and Patrick McGuire. 2012. “Systemic Risks in Global Banking: What Available Data Can Tell Us and What More Data Are Needed?National Bureau of Economic Research Working Paper 18531.

    • Search Google Scholar
    • Export Citation
  • Cerutti, Eugenio M., Maurice Obstfeld, and Haonan Zhou. 2019. “Covered Interest Parity Deviations: Macrofinancial Determinants.” National Bureau of Economic Research Working Paper 26129.

    • Search Google Scholar
    • Export Citation
  • Cheung, Yin-Wong, Menzie D. Chinn, and Antonio Garcia Pascual. 2005. “Empirical Exchange Rate Models of the Nineties: Are Any Fit to Survive?Journal of International Money and Finance 24 (7): 11501175.

    • Search Google Scholar
    • Export Citation
  • Clarida, Richard, Josh Davis, and Niels Pedersen. 2009. “Currency Carry Trade Regimes: Beyond the Fama Regression.” Journal of International Money and Finance 28 (8): 13751389.

    • Search Google Scholar
    • Export Citation
  • Clark, Todd E., and Kenneth D. West. 2006. “Using Out-of-Sample Mean Squared Prediction Errors to Test the Martingale Difference Hypothesis.” Journal of Econometrics 135 (1–2): 155186.

    • Search Google Scholar
    • Export Citation
  • Corte, Pasquale Della, Steven J. Riddiough, and Lucio Sarno. 2016. “Currency Premia and Global Imbalances.” Review of Financial Studies 29 (8): 21612193.

    • Search Google Scholar
    • Export Citation
  • Diebold, Francis X., and Robert S. Mariano. 2002. “Comparing Predictive Accuracy.” Journal of Business & Economic Statistics 20 (1): 134144.

    • Search Google Scholar
    • Export Citation
  • Dornbusch, Rudiger. 1976. “Expectations and Exchange Rate Dynamics.” Journal of Political Economy 84 (6): 11611176.

  • Du, Wenxin, Joanne Im, and Jesse Schreger. 2018. “The US Treasury Premium.” Journal of International Economics 112: 167181.

  • Du, Wenxin, Alexander Tepper, and Adrien Verdelhan. 2018. “Deviations from Covered Interest Rate Parity.” Journal of Finance 73 (3): 915957.

    • Search Google Scholar
    • Export Citation
  • Du, Wenxin. 2019. “Financial Intermediation Channel in the Global Dollar Cycle.” Remarks on “Mind the Gap in Sovereign Debt Markets: The U.S. Treasury Basis and the Dollar Factor.” Paper presented by Arvind Krishnamurthy and Hanno Lustig at the Challenges for Monetary Policy Federal Reserve Bank of Kansas Symposium, Jackson Hole, WY.

    • Search Google Scholar
    • Export Citation
  • Duarte, Margarida, and Alan C. Stockman. 2005. “Rational Speculation and Exchange Rates.” Journal of Monetary Economics 52 (1): 329.

    • Search Google Scholar
    • Export Citation
  • Engel, Charles. 1996. “The Forward Discount Anomaly and the Risk Premium: A Survey of Recent Evidence.” Journal of Empirical Finance 3 (2): 123192.

    • Search Google Scholar
    • Export Citation
  • Engel, Charles. 2014. “Exchange Rates and Interest Parity.” In Handbook of International Economics, Vol. 4, edited by Gita Gopinath and Elhanan Helpman, 453522. Amsterdam: Elsevier Science.

    • Search Google Scholar
    • Export Citation
  • Engel, Charles. 2016 “Exchange Rates, Interest Rates, and the Risk Premium.” American Economic Review 106 (2): 43674.

  • Engel, Charles, and Kenneth D. West. 2005. “Exchange Rates and Fundamentals.” Journal of Political Economy 113 (3): 485517.

  • Engel, Charles, and Steve Pak Yeung Wu. 2018. “Liquidity and Exchange Rates: An Empirical Investigation.” National Bureau of Economic Research Working Paper 25397.

    • Search Google Scholar
    • Export Citation
  • Evans, Martin DD, and Richard K. Lyons. 2002. “Order Flow and Exchange Rate Dynamics.” Journal of Political Economy 110 (1): 170180.

    • Search Google Scholar
    • Export Citation
  • Evans, Martin DD, and Richard K. Lyons. 2005. “Meese-Rogoff Redux: Micro-Based Exchange-Rate Forecasting.” American Economic Review 95 (2): 405414.

    • Search Google Scholar
    • Export Citation
  • Fama, Eugene F. 1984. “Forward and Spot Exchange Rates.” Journal of Monetary Economics 14 (3): 319338.

  • Fama, Eugene F., and James D. MacBeth. 1973. “Risk, Return, and Equilibrium: Empirical Tests.” Journal of Political Economy 81 (3): 607636.

    • Search Google Scholar
    • Export Citation
  • Farhi,