The new insight on exchange rate prediction and exposure

Transcription

1 The new insight on exchange rate prediction and exposure Abstract In this paper, a simple, but novel forecasting procedure is proposed to examine two intriguing issues in international economics from the market integn prospective: (i) the prediction of exchange rates, and (ii) the pricing of exchange rate risk in response to the macroeconomics activities or global events. This forecasting procedure, focusing on the structural linkage between the market integn and connectedness and exchange rates, includes a new indicator to measure market integn and connectedness and two recent developed estimations for the spurious and imbalanced regressions. Empirical results indicate that (i) our procedure outperforms the conventional random walk model in terms of forecasts of exchange rates, even the sample observations cover crises (2008 subprime crisis, European debt crisis and oil crash) and international monetary policy change (Quantitative easing policy, QE) period; (ii) the predictive ability of this new indicator is powerful on exchange rates; (iii) the changing roles of hedging currencies vary with different economics activities or significant global events, especially for the US dollar, Japanese Yen, and New Taiwan dollar. Keyword: Market integn; Exchange rate prediction; Capital flow; Currency Exposure

2 1. Introduction A well-known puzzle in international economics illustrates the best prediction of exchange rates by the conventional random walk model when compared to the structural economics models. To be specific, in an influential series of papers, Meese and Rogoff (1983a, 1983b, 1988) show that the forecasts of future nominal and real exchange rates generated by a simple random walk process always outperform those by structural macro-economic models. In what follows, two questions have been arisen here: Are exchange rates predictable? and what are predictors exhibiting powerful predictive ability in forecasting exchange rates? In defending fundamental-based exchange rate models, various combinations of economic variables and econometric methods have been used in attempt to challenge Meese and Rogoff's findings. For example, Mark (1995), Chinn and Messe (1995), Chen and Mark (1996) and Rossi (2005), among others, apply the long-horizon regression approach to investigate whether or not economic fundamentals are useful to forecast several leading U.S. dollar spot rates and report that monetary fundamentals may contain predictive power for movements in U.S. dollar exchange rate. In practice, exchange rate forecasts are quite crucial to policymakers in central bank and business practitioners in private sectors, because exchange rates fluctuations eventually influence several macroeconomic aggregates or firm s future cash flow of their own countries markets such as current account projections and real GDP growth. In addition, exchange rates are strongly forward looking, which would respond to the path of the policy and further affect capital flows among financial assets. Thus, how the exchange rate movements interact with the other financial assets and what role do foreign currencies play in a diversified investment portfolio are two important dimensions in line with international finance. An old controversial issue to demonstrate the impact of foreign exchange rates on stock returns is the pricing of exchange rate risk in firm values. This issue focuses on whether expected stock returns are affected by foreign exchange rate returns, an idea that has considerable theoretical, backing but mixed empirical support. Jorion (1990) shows that the co-movement between stock returns and the value of the dollar is positively related to the percentage of foreign opens of U.S. multinationals. On the contrary, Jorion (1991) notes 1

3 that the sensitivity of prices of multinational U.S. stocks to changes in the dollar exchange rate is not significant, while Griffin (2002) arrives at a similar conclusion, namely that pricing errors are magnified with the inclusion of exchange risk factors. Using the cross-section of U.S. stock returns and foreign exchange rates between 1973 and 2002, Kolari et al. (2008) provide evidence that stocks most sensitive to foreign exchange risk (in absolute value) have lower returns. However, Du and Hu (2012) conduct two sets of tests to determine whether Kolari et al.'s (2008) results are spurious and conclude that the exchange rate risk measured by contemporaneous exchange rate changes is not priced in the U.S. stock market. Nevertheless, the financial globalization and international investing has become important, especially in terms of being a crucial role in the recent financial crisis (e.g., You and Daigler, 2010; Mendoza and Quadrini, 2010). In addition, the increasing degree of financial globalization has been noticed by researchers and practitioners since it improves the economics efficiency and productivity by bringing capital, knowledge, and discipline from one country to others (Tong and Wei, 2010). Thus, financial returns depend on common market or systemic risks and result in enhancement of global market integn. In particular, since equity market valuations would respond to future economic activities and international capital flows, international stock market correlations have attracted more attention on account of the integn and globalization of financial markets. For example, Eun and Lee (2010) show international market integn by examining the international convergence in risk-return characteristics during By doing so, the pricing of exchange rate risk in global financial market integn could be also called into the question in this study. More importantly, to our best knowledge the literature on the pricing of exchange rate risk in global market integn has thus far rather limited. Accordingly, this paper fills this gap by considering a global common variable, which could represent the market integn. In particular, the intuition behind this idea is to examine the exchange rate exposure to international market risk and further understand how the investor use foreign currency to manage the risk of the portfolio when the market risk or crisis and global economics events occur. 2

4 Under the current globalized circumstance, this paper put interest in the influence of global factors on two aforementioned intriguing issues. We tackle them with a procedure combining the most up-to-date data, predictors, estimations and evaluation techniques. We provide a new variable, CDI AR, as a predictor in forecasting exchange rate returns and a measure in examining exchange rate exposure to international asset markets. The CDI AR, an indicator of market integn, bases on the AR-filtered CD test statistics proposed by Wang et al.(2017). This indicator could measure the interaction or contagion effect among international asset markets, monitor the incoming systemic risk and global common shocks by economics policy changes as well as indicate the pattern of capital flows. We here construct CDI AR by equity returns of seven developed and seven emergent markets and returns of oil, commodity, and Reits in a rolling window scheme to capture the time-varying market integn. Two recent developed estimations for spurious and imbalanced regressions by Wang and Hafner (2017) and Wang and Xie (2018) are concerns in this procedure. The criteria for evaluating the forecast performance are the mean squared predictive error () and the - test by Clark and West (2007). Our implementation and empirical results show novelty in several aspects. First, the predictor, CDI AR, is useful to forecast exchange rates. The forecast errors generated by our procedure are smaller than those by random walk model without drift. That implies the predictor CDI AR has significant predictive ability in exchange rate forecasts. The possible explanation to this result is because CDI AR indeed captures the time varying international market integn and contains unobservable and potential information in regard to market fluctuation by global common shocks or change in international policies, such as 2008 subprime crisis and the quantitative easing (QE) policy issued by US, Europe and Japan during the period from 2009 to Moreover, it could fellow from the fact that the exchange rate is forward looking and embodies information regarding to the future comovement among financial markets that cannot easily be measured by simple time series models. This CDI AR performs such an alternatives. Its early warning behavior of detecting the systemic risk is also provided in Wang et al. (2018). Second, the impact of exchange rate risk on the integn of global financial asset markets is significant. This finding demonstrates a new perspective on exchange rate risk pricing in market risk. That is to say, 3

5 the market (systemic) risk or the patterns of market integn could be affected and predicted by exchanges rate fluctuations. Briefly speaking, diversified investors could reduce the loss resulting from the market or systemic risk when managing portfolios through allocating their currencies position. Third, the hedging roles of several currencies alter in response to different global events or economic activities. For example, during the 2008 subprime crisis, the Swiss Franc, Japanese Yen and US dollar index appreciate and perform the hedging roles in reducing the loss arising from market risk. However, the Japanese Yen depreciate and lose its hedging role when oil crash in June, Fourth, empirical evidence of Rossi (2006) shows instabilities in models of exchange rate determinations and the impact of structural breaks on the out-of- sample forecasting performance. In our framework, the CDI AR itself contains the market instability pattern, thus the structural break effect for the outcome of out-of-sample forecasting ability is less. Last but not least, all results confirm our predictability conclusions. In other words, our procedure could be treated as useful tools for setting up portfolios since all implications conducted by our procedure could be the guidance to investor to ust the portfolio in reacting to the market economics conditions. The set-up of this paper could be organized as follows. Section 2 outlines models and econometrics methodologies. Data, all empirical results and analysis are reported in Section 3. Section 4 concludes. 2. Methodology 2.1 Predictive models There are many models that have been utilized to investigate the mutual nexus between exchange rates and macro-economic variables. We here consider to assess the predictive ability of market integn indicator ( CDI AR ) using a simple time series model. The coefficients are estimated by several estimations with several rolling window schemes. Note that the details of market integn indicator and econometrics analysis tools would be subsumed rigorously in the following. Let s t denote the logarithm of the exchange rate. All of currencies considered in our paper are priced against the US dollar(foreign currencies value per US dollar); thus, an increase in s t implies the depreciation of its resulting currency. 4

6 Moreover, Ferraro et al. (2015) states that forecast exchange rate movement with contemporaneous asset prices would not be useful for practical issue, we then consider a lagged variable as our predictor. Ultimately, we aim at estimating the following models: y i,t+h = β 0i + β 1i x q,t j + ε t, t=1 T j=1 k, (1) where x q,t j =CDI AR,q,t j, the growth rate of CDI AR changes on investing assets q in period t, is used as a predictor for exchange rate movement; y i,t equals the exchange rate returns of currency I; h represents the forecast-horizon. We here generate one-month-ahead forecasts (ŷ t+h ). We further explore the reverse relationship. Estimates of the exposure coefficients can be obtained from the following time-series regression, y i,t = β 0,i + β 1i x j,t, t=1 T (2) where X j,t is the exchange return on currencies j and y i,t is CDI AR,i,t. 2.2 Predictor We base on the new approach to measuring the market integn among global financial markets in our analysis on the AR-filtered CD test statistics, denoted as CD AR. as the following, where ρ ij,ar = CD AR = T 1 Σ T t=1 e i,te j,t T 1 Σ T t=1 e i,t 2 T 1 Σ T t=1 e j,t test statistics CD AR is a modification of CD test. 2T Σ N(N 1) i=1 N 1 Σ N i=j ρ ij,ar (3), e i,te j,t are the AR filtered estimates of residuals. The Wang et al. (2017) propose a methodology to correct this size distortion by using the autoregressive (AR) approximation to filter each component first and then construct the CD test statistics with each fitted residual, that being CD AR. Specifically, this CD AR embodies the information about the market fluctuation and interaction and capital flows as well as include the unobservable future information. In practice, the intuition behind using the CDI AR to measure market integn is its characteristics of being a test for error cross-sectional dependence in panels, i.e., the stronger the cross sectional correlation among markets, the 5

7 larger the value of the CDI AR. In other words, the increase in the CDI AR implies that all participating markets move to the same direction. For those reasons, we would like to see whether or not the CDI AR offer an ideal laboratory for cutting edge work on the exchange rate model. 2.3 Estimations In this study, a class of recent developed methodologies by Wang and Hafner (2017) and Wang and Xie (2018), the two-stage Cochrane-Orcutt Autoregressive Approximation (COAR) and the generalized Cochrane-Orcutt Autoregressive Approximation (GCOAR), are considered to address the above two important international financial issues, which are designed for spurious and imbalanced regressions, respectively. We now briefly describe the implementation of the two-stage CO-AR estimator. As follows: Stage.1 We first take the first difference of dependent and explanatory variables, and calculate the OLS estimate of β and β FD in the following model y t+1 = β x t + u t, (4) where = 1 L denote as the difference operator. Stage.2 Approximate the errors û t = y t+1 β FD x t by a finite order AR(k) model, i.e., û t = Σ k j=1 b ju t j + e tk. Then, conduct the following Cochrane-Ocrutt transformation of x t and y t+1, k y t+1 = y t Σ j=1 b jy t+1 j, x t = x t Σ t j b jx t j, (5) and compute OLS estimate of β of the regression y t+1 = βx t + u, t denoted as T β CO AR = (Σ t=k+1 k ) 1 T Σ t=k+1 x t y. t+1 (6) We then state the GCO-AR estimator for imbalanced regression, where the integrated order of regressor displays more persistent than the dependent variable and error term, its computational steps are similar to those of the COAR estimator for the spurious regression. More specifically, the only difference between two estimators is the following step 1, when we approximate the regressor by the AR model. 6

8 Step 1. We first approximate the regressor X t by an AR(k) model (see Poskitt, 2007), i.e., X t = Σ k j=1 b jx t j + e tk and ê tk = e t + Op(1). Again, we then conduct the following Cochrane- Oratt transformation of the variable X t and y t to balance the integrated orders of y t+1 and x t : y t+1 = y t+1 Σ k j=1 b jy t+1 j, x t = x t Σ k j=1 b jx tj. (7) Step 2. Compute the OLS estimate of the regression y t+1 = βx t+1 + u, t+1 then the GCO-AR estimate of β is denoted as: 2.4 Evaluation methods T β GCO AR = (Σ t=k+1 x 2) 1 T t Σ t=k+1 x t y. t+1 (8) We evaluate both of our testing model and benchmark model s out-of-sample forecasting ability according to their mean squared error (), which is the of testing model s to the benchmark model s, = Σ t=r T (y t+1 y 2t+1 ) 2 (9) Σ T t=r (y t+1 y 1t+1 ) 2 The value of smaller than unity denotes that the model (ŷ 2t ) forecasts better than the benchmark (ŷ 1t ) statistic For further confirmation, we adopt the - test proposed by Clark and West (2007) We first set our benchmark model, random walk, as the parsimonious model( y 1t ) and the testing model as the larger model( y 2t ). Write the predictions and prediction error as follow ŷ 1t+1 = X 1t+1 β 1t, e 1t+1 = y t+1 y 1t+1, y 2t+1 = X 2t+1 β 2t, e 2t+1 = y t+1 y 2t+1 the -usted is 2 2 f t+1 = ê 1t+1 [e 2t+1 (y 1t+1 y 2t+1 ) 2 ] (10) Let f be the corresponding sample average and p be the number of predictions we used, f = p 1 T Σ t=r f t+1, then, our test statistic is pf [sample variance of (f t+1 f )] 1 2 (11) 7

9 The null hypothesis is equal, and the alternative is that testing model generate small than those of benchmark model. By regressing f t+1 on a constant and using the resulting t-statistic for a zero coefficient, we reject the null if this test statistic is greater than (for a one-sided 0.10 significance level) or ( for a one sided 0.05 significance level). 3. Data and Empirical Result 3.1 Data All of our data are extracted from Bloomberg. The dataset consists of three investing asset prices (Oil, Commodity, and REITs), stock indices for fourteen countries (U.S., U.K., France, Canada, Germany, Japan, Australia, Brazil, Russia, China, India, Malaysia, Thailand, Taiwan) and six hedging currency series relative to US dollar (the Swiss franc, the Japanese yen, the Canadian dollar, the British pound, the US dollar, and the Euro dollar). The sample period spans from January 1999 to August 2017 at monthly frequency. We calculate return as the change in the log price on investing assets and stock indices. Notice that exchange rate series represent foreign currency value per unit of the US dollar. 3.2 Empirical Findings Market integn In a comparison, we do forecasts of exchange rates by a simple random walk process without drift and a predictive regression with the predictor CDI AR. Four types of CDI AR are considered in our framework, such as (i) CDI AR.E, by returns of fourteen countries equities; (ii) CDI AR,C, by returns of fourteen countries equities and by return of commodity; (iii) CDI AR,R, by returns of fourteen countries equities and by return of Reits; (iv) CDI AR,oil, by returns of fourteen countries equities and by return of oil. We construct all of them using a rolling window scheme to catch time varying market information. For the sale of addressing the predictive ability of the CDI AR variable, we first take a glance at time-varying patterns of those CDI AR -type indicators. Figure1 depicts four time-varying CDI AR type indicators. It appears that all of them surge to a peak in the beginning of the 8

10 subprime crisis and stay in this scenario until In fact, several subsequent economics activities occurred during the period from 2008 to the early of 2013, such as Quantitative easing (QE) policy by central banks of the US, Europe and Japan. On the other hand, in response to aforementioned noticeable economic activities, the connectedness among all considered markets becomes stronger which implies that those markets flock toward same direction once a common event or trend occur. In brief, the CDI AR indicated the market risk without any doubts. Otherwise, a sudden drop happened in the CDI AR when the common trend effect vanishes. Moreover, the CDI AR plummet sharply again as a response to the great oil crash in Exchange rate prediction In general, the changing patterns of currencies in response to the performance of equity returns (e.g., Nieh and Lee, 2001; Reboredo et al., 2016). Two scenarios are addressed in the existing literature. First, from a flow-oriented model perspective, Dornbusch and Fisher (1980) shows the drop in the domestic currency result in the increase of domestic firms exports and future cash flows. Second, in the view of a stock-oriented model, Branson et al. (1977) examine the currency movement by a portfolio-balance model. The result of this work illustrates by changing the demand and supply of international financial asset holdings of a given portfolio diversified across countries, the ustment of foreign exchange rate could be influenced by equity prices. In spite of the fact that afterwards these two scenario demonstrate substantial correlation between equity and currency markets, there are many few of them focuses on the interaction between the market risk (systemic risk) and exchange rates. In this regard, our empirical analysis intends to emphasize the interdependence between the hedging currencies and the connectedness of financial markets. With loss of generality, we consider the random walk without drift to be the benchmark model because it is the toughest benchmark to beat, since many researchers confirm the works by Messe and Rogoff (1983a, b) showing that forecasts of exchange rates by a simple random walk are better than those by any economics models. Our analysis focuses on two evaluation measures of predictive ability computed over several estimation window size. The choice of the estimation window size has always been a concern when speaking of the out-of-sample forecasting since the use of different window size may lead to different 9

11 empirical results. As a consequence, arbitrary choices of window size are concerned with how the sample is split into in-sample and out of sample portions. Rogoff and Stavrakeve (2008) even questioned the robustness of existing empirical results to the choice of the staring out-of-sample period. On the analysis of this work, the two considered models parameters are recursively re-estimated with a rolling window which is equal to half of the total sample size in the baseline analysis. This criterion satisfies the finding by Hansen and Timmerman (2012) illustrating the optimal size of out of sample forecasts. This criterion is also considered in Rossi (2013). To check the robustness analysis, following several literatures, including Chinn (1991), Cheung et al. (2005) and Clark and West (2006, 2007), we evaluate the model s forecasting performance with rolling windows of 36, 60, 75, 93, 120 observations. Ferraro et al. (2015) illustrate it is not necessary to consider the contemporaneous linkage between the exchange rates and predictive variables when predicting exchange rates. Thus, in line with Ferraro et al. (2015), we would like to address the several-lags correlation between the exchange rate and our four CDI AR type indicators. That means considering 1- lag and 2-lag predictive variable (CDI AR ) to forecast the exchange rates. In this study. Three estimations are concerned in our analysis, such as the OLS, COAR and GCOAR estimations. The latter two are proposed by Wang and Hafner (2017) and Wang and Xie (2018). And for the forecast horizon, different from previous studies that agree on lack of short-horizon predictability for traditional monetary fundamentals (see Cheung et al. 2005), we attempt to prove that by utilizing appropriate predictors and forecasting methods, it could present forecasting ability even with one-period-ahead horizon. We then apply the and the statistics to evaluate the one-period ahead forecasting performance of our procedure and the random walk without drift. A smaller than unity denotes that our model performs better than a random walk. For statistics, the null posits that the from the parsimonious model, benchmark model, is smaller than that of the larger model. That is, when statistic is larger than (for one sided 0.05 significance level) or (for one sided 0.05 significance level), we reject the null. For simplification, we only demonstrate our best performing case. Most forecasts of exchange rates using 1-lag CDI AR predictor perform poor when compared to those by 10

12 a simple random walk, only the GCOAR case could perform fragmentary predictive ability. When the 2-lag CDI AR is considered in the model, clearly, the results found more empirically in favor of Imbalanced regression case with larger window scheme. Both and statistics of GCOAR demonstrate more reliable result than those of the other two estimating methods. Additionally, an optimal window size for our framework is window size 75, which exhibits the strongest evidence of predictive ability in all six currencies. We could observe the and statistics are perfectly consistent with each other in window size=75 case, while those of other cases are distorted. Preliminary concludes, although we do not attempt to suggest a best estimation and evaluation window size contributing to forecasting literature, the predictive ability of the 2-lag CDI AR on exchange rates is in strong preference to a window size of 75 and slight preference to that of 120 in our investigation. To some degree, our forecasting framework could be the solution of Messes and Rogoff Puzzle since this framework combines several important procedures, such as the up-to-date data, the optimal estimation methods, a suitable predictor and reliable evaluation techniques. It follows from the fact that the exchange rate is forward-looking and embodies information regarding future co-movements among market, for example, the market integn and connectedness which cannot easily be represented by simple time series models Hedging role of currency In the past, most studies documented the pricing of exchange rate risk in returns of equity markets. For example, Campbell et al. (2010) proposes optimal hedging to hedge the currency risk in stock market. But under the internationalized circumstance, the interdependence among financial markets are more crucial to investors in portfolio allocations, since the markets are integrated increasingly. Therefore, we propose a more insightful way to examine the exchange rate exposure to equity markets, i.e., checking the relationship between the exchange rates and market risk or integn (or CDI AR ). Again, we estimate the coefficient using rolling window scheme with 75 observations in order to gauge the time varying influence of exchange rate movements on market risk or integn. A positive estimated coefficient means the increase return of exchange rate results in larger 11

13 value of CDI AR, which demonstrate stronger cross-sectional correlations among markets or more connectedness within markets through capital flows. Otherwise, markets are teared apart from each other when the coefficient displays negative. Fig. 2 illustrates the estimated coefficient β Plot for regressing CDI AR on currency returns. Similar patterns occur in the GBP, the CAD, the USD and the EURO case, however, we only show CHF, JPY, USD, and the EURO because of simplification. When looking at the overall trend, we can observe radically change interfacing previous results during financial crisis and 2014 oil price collapse. The novelty of the methodologies we propose is that comparison between β Plots and currencies movements demonstrates the implication to which we need to know the degree of market integn and exchange rate changes over our sample period in order to understand capital flow directions and how changing nature of currencies hedging role drove world economy. Our results fall along a spectrum. For episode, the Swiss franc and the Japanese yen return exhibit negative trend while the US dollar return performed reversely, which indicates that these three currencies appreciate during crisis period and serve as the roles of hedging currencies. That implies a rise in currencies results from instantaneous international capital to flow into countries, and subsequently stimulate stock markets interaction. For 2014 episode, markets detached from each other and currencies hedging effect is subtle due to perpetual bilateral transaction. As a result, except for the Swiss Franc, none of these countries behave main recipients or resources of capital flow. In other words, hedging roles of Japanese Yen and US dollar altered during this event. According to the above evidence, we can confirm the argument of this article. i.e., our framework could successfully detect hedging currencies when markets move together to the same direction; whereas market connectedness becomes less substantial, investors frequently moved out their capitals, and eventually the hedging currencies lose their capability in escaping from currency risk. In a conclusion, the fact the changing roles of hedging currencies vary with different economics activities or events could be monitored by pricing the exchange rate risk. 3.3 Four Asian emerging countries 12

14 This session repeats the previous analysis of exchange rate prediction and exposure, except that we turn our focus to four Asian emerging countries (Taiwan, South Korea, Hong Kong and Singapore). The results substantially correspond to previous finding in two aspects. First of all, Table 2-4 records s for comparing our forecasting model to a benchmark random walk without drift model. We have found that the New Taiwan dollar, the South Korean Won and the Singapore dollar display out-of-sample forecasting ability at short forecast horizon with one period lagged predictors. It should be noted that although in our advanced countries case, model with 2-lag predictors demonstrate stronger predictive power, our result in Asian emerging markets suggest more out-of-sample predictive ability in favor of model with 1-lag predictors. Furthermore, the predictive power of the new Taiwan dollar, the South Korean won and the Singapore dollar shows robustness across different rolling window scheme from 0.2T(36) to 0.4T(75), where T equals to input observations. Secondly, Fig. 3 depicts the estimated coefficient β Plot for regressing CDI AR on emerging currencies returns. The time-varying exchange rate fluctuation effect on financial market integn is observed. In our result, the appreciation of the New Taiwan dollar, the South Korean won, and the Singapore dollar have tended to strengthen the degree of market integn during the international monetary policies change period, but certainly we find the absence of the Hong Kong dollar s hedging characteristic. Our findings confirm the expectation that international capital flows into appreciating currencies and eventually brings markets together. 4. Conclusion This paper successfully possesses the prediction of exchange rates and the pricing of exchange rate risk in global asset markets by a new procedure including an indicator of market integn and two recent developed econometrics tools. More importantly, out-ofsample forecasts of exchange rates generated by our procedure perform better in comparison with those by the conventional random walk model. This finding is in the marked contrast to that in most existing studies. That is to say, the predictive power of this new market integn indicator substantially sustains for forecasts of exchange rates. Moreover, an instability in the relationship among financial markets due by the unexpected 13

15 economics shock or macro-economic activities could be also examined by analyzing the exchange rate pricing in global financial markets with our procedure. Two well-known crises occurred in 2008 and 2014 could be detected by our framework in an accurate way. In addition, the new hedging roles of several hedging currencies are altered depending on different economics activities and events, especially for the Japanese Yen, US dollars, and the New Taiwan dollar. In a conclusion, the applicability and feasibility of our procedure are established through these convincing empirical evidence. Reference Branson W, Halttunen H, Masson P (1977) Exchange rates in the short run: The dollardentschemark rate. European Economic Review 10(3): Campbell JY, Medeiros KS, Viceira LM (2010) Global Currency Hedging. Journal of Finance 65(1): Cheung YW, Chinn MD, Pascual AG (2005) Empirical exchange rate models of the nineties: are any fit to survive? Journal of international money and finance 24(7): Chinn, MD (1991) Some Linear and Nonlinear Thoughts on Exchange rates. Journal of International Money and Finance. 10(2): Chinn MD, Meese RA (1995) Banking on currency forecasts: how predictable is change in money? Journal of international economics 38(1-2): Clark TE, West KD (2006) Using out-of-sample mean squared prediction errors to test the martingale difference hypothesis. Journal of Econometrics 135(1-2): Clark TE, West KD (2007) Approximately Normal Tests for Equal Predictive Accuracy in Nested Models. Journal of Econometrics 138(1): Dornbusch R, Fischer S (1980) Exchange Rates and Current Account. The American Economic Review 70(5): Du D, Hu O (2012) Foreign exchange volatility and stock returns. Journal of International Financial Markets, Institutions and Money 22(5): Eun CS, Lee J (2010) Mean-variance convergence around the world. Journal of Banking & Finance 34(4):

16 Griffin JM (2002) Are the Fama and French factors global or country specific? The Review of financial studies 15(3): Hansen PR, Timmermann A (2012) Choice of Sample Split in Out-of-Sample Forecast Evaluation. Working Paper Jorion P (1990) The Exchange-Rate Exposure of U.S. Multionationals. The journal of Business 63(3): Jorion P (1991) The pricing of Exchange Rate Risk in the Stock Market. Journal of Financial and Quantitative Analysis 26(03): Kolari JW, Moorman TC, Sorescu SM (2008) Foreign exchange risk and the cross-section of stock returns. Journal of International Money and Finance 27(7): Mark NC (1995) Exchange rates and fundamentals: evidence on long-horizon prediction. American Ecoenomic Review 85(1): Meese RA, Rogoff KS (1983a) Empirical Exchange Rate Models of the Seventies: Do they fit out of sample? Journal of International Economics 14(1-2):3-24 Meese RA, Rogoff KS (1983b) The Out-of-Sample Failure of Exchange Rate Models: Sampling Error or Misspecification? In Exchange Rates and International Macroeconomics Meese RA, Rogoff KS (1988) Was it Real? The Exchange Rate-Interest Differential Relation over the Modern Floating-Rate Period. Journal of Finance 43(4): Nieh CC, Lee CF (2001) Dynamic relationship between stock prices and exchange rates for G-7 countries. Quarter Review of Economics and Finance 41(4): Pesaran MH (2004) General diagnostic tests for cross section dependence in panels. CESifo Working paper series no. 1229, CESifo Group, Munich Qi M, WU Y (2003) Nonlinear Prediction of Exchange Rates with Monetary Fundamentals. Journal of Empirical Finance 10(5): Reboredo JC, Rivera-Castrob MA, Ugolinibc A (2016) Downside and upside risk spillovers between exchange rates and stock prices. Journal of Banking and Finance 62:76 96 Rogoff KS, Stavrakeva V (2008) The Continuing Puzzle of short horizon exchange rate forecasting. NBER Working Paper Rossi B (2005) Testing Long-Horizon Predictive Ability with High Persistence, and the Meese Rogoff Puzzle. International Economic Review 46(1):

17 Rossi B, Inoue A (2012) Out-of-Sample Forecast Tests Robust to the Choice of Window Size. Journal of Business and Economic Statistics 30(3): Rossi B (2013) Exchange rate predictability. Journal of Economic Literature 51(4): Tong H, Wei SJ (2010) The Composition Matters: Capital Inflows and Liquidity Crunch During a Global Economic Crisis. The Review of Financial Studies 24(6): You L, Daigler R (2010) Is international diversification really beneficial? Journal of Banking and Finance 34(1) TABLE 1- and - windows size=75, 2-lag CDI AR CHF JPY X(t-2)\Y(t) OLS COAR GCOAR OLS COAR GCOAR CD_ OIL * * 1.31 * COMMODITY * 1.66 ** * 2.02 ** REITS * 1.49 * * 1.92 ** MARKET * 1.62 * * 2.00 ** GBP CAD CD_ OLS COAR GCOAR OLS COAR GCOAR OIL * 1.65 ** COMMODITY * 1.41 * REITS MARKET USD EURO CD_ OLS COAR GCOAR OLS COAR GCOAR OIL * 1.45 * COMMODITY REITS * 1.68 ** * 1.39 * MARKET * 1.52 * * 1.37 * 16

18 TABLE 2- Ratio windows size=36, 1-lag CDI AR X(t-1)\Y(t) TWD KRW CD_ OLS COAR GCOAR OLS COAR GCOAR OIL * * COMMODITY * * REITS * * MARKET * * HKD SGD CD_ OLS COAR GCOAR OLS COAR GCOAR OIL * COMMODITY * REITS * * MARKET * * TABLE 3- Ratio windows size=54, 1-lag CDI AR X(t-1)\Y(t) TWD KRW CD_ OLS COAR GCOAR OLS COAR GCOAR OIL * * COMMODITY * * REITS * * MARKET * * HKD SGD CD_ OLS COAR GCOAR OLS COAR GCOAR OIL * COMMODITY * REITS * MARKET * 17

19 TABLE 4- Ratio windows size=75, 1-lag CDI AR X(t-1)\Y(t) TWD KRW CD_ OLS COAR GCOAR OLS COAR GCOAR OIL * * COMMODITY * * REITS * * MARKET * * HKD SGD CD_ OLS COAR GCOAR OLS COAR GCOAR OIL * COMMODITY * REITS * MARKET * Fig. 1 CDI AR plot 18

20 Fig. 2 19

21 Fig. 3 20