The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
Working Paper 9009 INTERVENTION AND THE FOREIGN EXCHANGE RISK PREMIUM: AN EMPIRICAL INVESTIGATION OF DAILY EFFECTS by Owen F. Humpage and William P. Osterberg Owen F. Humpage is an economic advisor and William P. Osterberg is an economist at the Federal Reserve Bank of Cleveland. Kyle Fleming provided able research assistance. The authors also gratefully acknowledge the technical assistance of Ralph Day. Working papers of the Federal Reserve Bank of Cleveland are preliminary materials circulated to stimulate discussion and critical comment. The views stated herein are those of the authors and not necessarily those of the Federal Reserve Bank of Cleveland or of the Board of Governors of the Federal Reserve System. August 1990 www.clevelandfed.org/research/workpaper/index.cfm Abstract Currency markets have witnessed a sharp increase in government intervention since 1985. Many observers believe that this intervention promoted the dollar's depreciation between 1985 and early 1987, and that intervention has since helped to stabilize dollar exchange rates. This paper tests for a systematic effect of daily dollar intervention on exchange rate risk premia. We test for both portfolio balance effects and signaling influences by using daily data on central bank intervention (in dollars) against both the yen and the West German mark. Following work by Dominguez (1989) and Loopesko (1984), we measure the daily risk premium in terms of the deviation from uncovered interest parity. However, we follow other empirical analyses of exchange rates and allow for generalized conditional autoregressive heteroscedasticity (GARCH). Some evidence is found for both the portfolio balance and signaling channels. www.clevelandfed.org/research/workpaper/index.cfm I, Introduction The recent emphasis on foreign exchange intervention by several large industrialized nations has renewed an interest in the study of channels through which intervention may operate. Some research suggests that intervention may be responsible for the failure of exchange market efficiency models. From a policy standpoint, if intervention has an impact on exchange rates, then the channel of its influence must be identified in order to determine whether it is an independent policy tool. For this reason, most studies focus on sterilized intervention, which by definition does not affect the monetary base. Sterilized intervention may operate through either the portfolio balance or signaling channel. Most empirical studies have found little support for the portfolio balance channel. Evidence of a signaling role is somewhat stronger; however, disentangling the two effects is difficult. This paper uses confidential daily data on G-3 central bank intervention to test for the presence of both portfolio balance and signaling effects of intervention on exchange rate risk premia. Use of high-frequency daily data allows us to capture the relationships among intervention, volatility, and excess returns. The existence of a risk premium is one possible explanation for the poor out-of-sample forecasting performance of exchange rate models. Variances of exchange rates seem to show persistence, with distinct periods of low and high volatility. Various researchers have suggested that policy shifts may be related to volatility in asset prices. Thus, it may be useful to think of the impact of intervention as operating specifically through a risk premium. www.clevelandfed.org/research/workpaper/index.cfm Unfortunately, there is no consensus on how to model risk in foreign exchange. A widely used approach is to analyze the relationship between forward rates and spot rates. Hodrick (1989), for example, relates the forward premium to conditional means and variances of market fundamentals. One disadvantage of approaches that relate risk premia to fundamentals is that they do not permit testing with high-frequency data. However, a method that can be applied to daily analyses of intervention is to analyze the measure of realized excess returns suggested by the uncovered interest parity (UIP) condition. Two previous studies (Loopesko [I9841 and Dominguez [1989]) have taken this approach. This paper differs by using more recent data and modeling the conditional variance of the excess returns. We take advantage of recent advances in modeling conditional variances in asset returns (generalized autoregressive conditional heteroscedasticity [GARCH]), particularly as applied to exchange rates. Baillie and Bollerslev's 1989 study is one of many to find evidence for GARCH in exchange rates. To allow for the possibility that the conditional variance of the excess return influences its mean (GARCH-M), and that intervention influences the conditional variance, we utilize a variant of GARCH-M that allows the error term to have a conditional student-t distribution. In previous applications to exchange rate data, the student-t distribution has explained leptokurtosis (Baillie and Bollerslev). Our analysis confirms the existence of portfolio balance and signaling channels, but differs from other studies in regard to which countries' operations had significant impacts. Although evidence of GARCH is present, the conditional variance does not influence the conditional mean (no GARCH-M). In addition, we find evidence of day-of-the-weekeffects. www.clevelandfed.org/research/workpaper/index.cfm 11. Related Literature Theorv: Channels of Influence for Intervention Theory has focused on sterilized intervention for two reasons.l First, the effects of unsterilized intervention may be indistinguishable from those of monetary policy. Second, most large industrialized nations claim that intervention is sterilized. Most analyses of intervention utilize the portfolio balance approach (Branson and Henderson [1985]). With risk-averse investors and imperfect substitutability of assets of differing currencies, shifts in the relative supplies of assets may induce changes in rates of return via the exchange rate. However, under Ricardian equivalence, sterilized intervention would have no impact, even with imperfect substitutability (Backus and Kehoe [I9881) . The other channel through which intervention may operate is signaling, or the provision of new information to the market (Obstfeld [I9891 and Dominguez [1989]). Intervention can provide an effective and credible signal about future monetary policy if 1) the central bank has inside information and the incentive to reveal it truthfully and 2) the market has the ability to determine the credibility of that information. Intervention may be preferable to other signals because it does not require the central bank to change the monetary base. implied policy. On the other hand, this may make it easier to renege on the The fact that the central bank puts its own money on the line by intervening has been cited by some as a reason why intervention may have www.clevelandfed.org/research/workpaper/index.cfm credibility. If intervention operates through a signaling channel, then coordination may either strengthen its signal, or it may give some the incentive to "free ride," if such actions are undetectable by the market. Evidence While most investigations of the portfolio balance channel conclude that changes in the currency denomination of bond holdings do not influence exchange rates (see Weber [I9861 for a survey), Danker et al. (1984), Loopesko (1984), Johnson (1988), and Ghosh (1989) find evidence supportive of such a channel. However, even if changes in the relative stocks do influence exchange rates, intervention still may have no meaningful impact, since the volume of sterilized intervention is small relative to the total stock of assets. Evidence for a signaling channel is somewhat more consistent. Dominguez (1988) finds that, between 1977 and 1981, the relationship between intervention and money-supply surprises is consistent with the idea that intervention conveys information about future monetary policy. The response of exchange rates to intervention suggests that whether the market bets for or against intervention depends on the central bank's credibility in conveying such information. Using daily data, Humpage (1988) finds evidence that initial intervention has an effect on exchange rates, but subsequent intervention does not. Dominguez (1989) looks at the impact of official sterilized intervention and coordination from 1985 to 1987, and attempts to distinguish longer-term influences by using one-month and three-month interest and exchange rates. Results indicate that coordinated intervention may have a longer-term influence than unilateral intervention, the impact of which was www.clevelandfed.org/research/workpaper/index.cfm less consistent. On the other hand, Humpage (1984) finds that U.S. monetary authorities react to smooth, unanticipated exchange rate movements, but notes no evidence of an expectations effect. Dominguez and Frankel (1990) attempt to disentangle portfolio balance from expectation influences through the use of exchange rate expectations data and newspaper accounts of intervention. They find evidence for both effects. Loopesko (1984) and Dominguez (1989), the two studies that take the approach closest to that of this paper, use the UIP condition to test the impact of daily intervention. Loopesko examines the joint hypothesis of perfect asset substitutability and exchange market efficiency using daily data from 1975 to 1981. Cumulative central bank intervention that could have been known to market participants is the independent variable used to test for a portfolio balance effect. Lagged values of the realized profits and exchange rate are included to test for market efficiency. Although the joint hypothesis is resoundingly rejected, identification of the influence of the independent variables is clouded by the possibility that variables may have been omitted, or that not all of the measured intervention has been observed. 111. Risk Premia in Exchange Rates There is no consensus as to the appropriate theoretical framework for exchange rate risk premia. Lucas's (1982) intertemporal dynamic two-country model implies that risk premia should be related to preferences and to the stochastic behavior of the driving processes, such as monetary policy. The intertemporal capital asset pricing model (Engel and Rodrigues [1987], Giovannini and Jorion [1989], and Mark [1988]) suggests that risk premia should be related to covariances among asset returns. The consumption-based www.clevelandfed.org/research/workpaper/index.cfm capital asset pricing model (Hodrick [1989], Cumby [1988]) has specific implications for covariation between asset returns and intertemporal marginal rates of substitution in utility. Option pricing theory implies that risk premia are imbedded in foreign currency options prices (Lyons [1988], McCurdy and Morgan [1988]). Tests of all of these approaches have had mixed results. Hodrick (1987) and Baillie and McMahon (1989) provide excellent overviews of this literature. Evidence favoring the existence of a risk premium in foreign exchange rates is indirect. Violation of the UIP condition, rejection of unbiasedness in the forward market, and poor out-of-sample forecasting performance of log-linear models that rely on first moments suggest that a risk premium may exist. However, most tests of UIP or of the relationship between forward and future spot rates are joint examinations of market efficiency, perfect substitutability, and capital mobility. Nonetheless, evidence of conditional heteroscedasticity in exchange rates naturally leads to attempts to explain time variation in the conditional variance of exchange rates. Many of the theoretical approaches mentioned above imply that the conditional variance of exchange rates should be related to time-varying conditional covariances that involve exogenous processes such as money or output. However, testing these theories would require using data of no greater than monthly frequency. As Baillie and Bollerslev (1989) point out, evidence of time variation in conditional variance is weaker with such data. Most efforts to model the conditional variances of exchange rates utilize ARCH (autoregressive conditional heteroscedasticity) or its variants (GARCH, GARCH-M). ARCH allows for conditional normality combined with a leptkurtic, symmetric unconditional distribution consistent with the typical fat-tailed www.clevelandfed.org/research/workpaper/index.cfm nature of asset return data. Baillie and Bollerslev find that a version of GARCH in which the conditional distribution is student-t successfully models heteroscedasticity in the first-difference of the logarithm of daily exchange rates. Hsieh (1989) confirms the ability of ARCH or GARCH, in combination with various assumptions regarding nonnormality, to remove heteroscedasticity from similar data. Both Baillie and Bollerslev and Hsieh (1988) find day-of-the-weekeffects in exchange rate data. The limitations of ARCH as a vehicle for explaining conditional variance are pointed out by Pagan and Hong (1988), Nelson (1987), and others. Hodrick (1987, p. 110) argues that ARCH may be inappropriate for analyzing volatility in exchange rates. If high-risk premia are rooted in policy uncertainty, then clarification by policymakers should reduce them. However, ARCH implies persistence in conditional variance, so the implied risk premia would only be reduced after a period of lower ex-post volatility. The role of policy regime shifts in explaining exchange market volatility is explored empirically by Lastrapes (1989). IV. Interest Paritv and Excess Return We use the UIP condition to generate our measure of the exchange rate risk premium. An alternative would be to use the covered interest parity condition, which involves forward contracts. However, forward contracts are intended for delivery at least one month in the future, which, with daily data, would entail a loss of degrees of freedom in order to account for serially correlated errors induced by overlapping forecast intervals. UIP suggests utilizing equation (1). www.clevelandfed.org/research/workpaper/index.cfm where Rt Rt* St RETt = -- = domestic interest rate, foreign interest rate, exchange rate (foreign currency price of U.S. dollars), and excess return. Here, the investor does not cover the transaction by selling forward, but instead forms expectations of the spot rate (EISt+l] for a one-day investment), which is uncertain at the time of the transaction. We utilize daily data on interest rates, exchange rates, and intervention Timing conventions in the foreign exchange markets require the buying and selling of currency to be completed prior to the investment. Consider an investor who places funds overnight. This investor buys West German marks on day t-2 for delivery on day t. On day t-1, he sells the marks for dollars that are to be delivered on day t+l. On day t, his marks are collected and invested overnight. On day t+l, he receives his marks, which he had previously contracted to sell. These considerations, together with the assumption that EISt+l] = St+l, imply equation (2). where the excess return has been decomposed into a risk premium (RP) and a forecast error (FE). Since we utilize St+l instead of its expectation, an MA(1) term is introduced into FEt. A regression of RETt on variables that would be in the investor's information set at transaction time provides a joint test of informational efficiency and absence of a risk premium. Hence, www.clevelandfed.org/research/workpaper/index.cfm if our measure of intervention captures its influence on portfolio balance at t-3 and explains RETt, we would have evidence of an influence on risk premia if this market were informationally efficient. We introduce intervention in two forms. To test for its influence through the portfolio balance channel, the total of the two countries' cumulative intervention is entered at t-3. If this measure captures a portfolio balance effect, then the identity of the countries should be immaterial. To examine this, each country's cumulative total is entered separately, as well. As indicated above, this is a joint test of efficiency and the existence of a risk premium. In addition, in the absence of a portfolio balance channel, this test may indicate a signaling role for intervention. To further test for a signaling effect, we distinguish between coordinated and unilateral intervention at t-3. V. An Empirical Model A substantial body of literature suggests that a martingale process aptly describes movements in exchange rates, and that the variances of the first differences of exchange rates are heteroscedastic. Here, we model the forecast error with the GARCH procedure used by Baillie and Bollerslev (1989). The residuals from the conditional mean equation for RETt are assumed to be generated by a conditional student-t distribution, and the conditional variance of the residuals, ht, is modeled as an ARMA process. www.clevelandfed.org/research/workpaper/index.cfm In equation (3), Yt is the measured excess return and Xt is the vector of explanatory variables, which includes intervention, an intercept, four day-of-the-weekdummies, and dummies for missing data and vacation days. In equation ( 4 ) , the error ut is allowed to follow an MA(1) process. The term rhtPallows for the conditional variance (p-1) or the conditional standard deviation (p=.5) to influence the excess return. Although we do not present a theoretical model for this effect, it is implied by models such as that in Hodrick (1989). Equation (5) indicates that the distribution of the et conditional on the information set It-1 is student-t,with mean zero, variance ht, and distributional parameter v. If v exceeds 30, this distribution is approximately normal. Equation (6) shows that we utilize a GARCH(1,l) parameterization, with an intercept. Preliminarv Tests and Procedures A standard ARMA analysis of RETt did not help us to distinguish between AR(1) and MA(1) representations. Since overlapping forecast intervals suggest an MA(1) form, that is the one with which we proceed. Augmented Dickey-Fuller tests reject the hypothesis of a unit root in Yt. In order to examine the sensitivity of our results on the significance of intervention on excess returns, we omit the daily dummies, the MA(1) term, and the ARCH-in-mean term (ht or ht . 5, from the mean equation. The extent to which the residuals are nonwhite is indicated by the reported Q statistics (Q[15] indicates that 15 lags were utilized). ~ / ( h ~ -adjusts ~) the usual Q www.clevelandfed.org/research/workpaper/index.cfm statistic for heteroscedasticity, and Q~ is the standard Q statistic for the squared residuals, which may indicate ARCH effects. It, too, is adjusted for heteroscedasticity, and then reported as Q2fit. The parameter v indicates the extent to which the distribution deviates from normality. The sample measure of skewness aids in indicating the success of our distributional assumptions in modeling the conditional variance. to kurtosis, 3(v-2)/(v-4), Finally, we report the sample analogue where appropriate. Data The sample period is August 3, 1984 to February 19, 1990, and there are 1,770 daily observations, excluding lags. We obtained the exchange rate and interest rate data from the Paris market through DRIFACS PLUS (1988). The ultimate source is Credit Lyonnais, Paris. Yen-dollar and mark-dollar exchange rates are constructed as cross-rates for each currency quoted against the French franc. The exchange rate data are averages of bid and ask quotes as of 2:00 p.m. in Paris. Interest rates are overnight Eurocurrency deposit rates, quoted on a 360-day basis, as of 9:30 a.m.; they are converted to a daily basis. The market chosen is the only one in which we found overnight Euroyen deposit rates. Intervention data are daily net purchases of dollars by the United States, West Germany, and Japan, provided by the Board of Governors of the Federal Reserve System. Since the data are measured in dollars, we avoid the need to construct dollar measures of intervention using the exchange rate, which would imbed simultaneity into our analysis. Over the period investigated, virtually all U.S. intervention was against the mark or yen. The single exception was a purchase of $16.4 million equivalent British pounds in www.clevelandfed.org/research/workpaper/index.cfm February 1985 (see Cross [1985, p. 581). We include West German and Japanese intervention, but not intervention by other large central banks, which tend to focus intervention on their own currency's exchange rate rather than on the yen-dollar or mark-dollar rate. Moreover, their currency's relationship against the dollar need not be the objective of the intervention: Many participants in the European Monetary System (EMS) intervene in dollars to maintain their currencies within EMS limits. Although third-party intervention may affect the yen-dollar or the mark-dollar exchange rate, the impact is often caused by the aggregation of purchases and sales of dollars undertaken independently by many different countries. Results The portfolio balance channel and cumulative intervention If the portfolio balance channel is operative, the total change in relative portfolios should be important to the investor. In Table I, we use as our intervention measure the total of U.S. and West German purchases of U.S. dollars against the mark as of date t-3. Since intervention is measured at the end of the day, this is information that investors could have had. Table I indicates that an increase in dollar purchases tends to result in significantly increased (at the 1 percent level) dollar excess returns. In the absence of an agreed-upon theory of the determination of exchange rate risk premia, it is unclear how we should interpret the sign of the impact of intervention. However, the portfolio balance approach suggests that the excess return on dollar assets must increase in order to compensate investors for holding a greater stock of dollar assets. The positive coefficient implies that an increase in the stock of dollar assets (a negative www.clevelandfed.org/research/workpaper/index.cfm value) is associated with a decrease in the risk premium. This is inconsistent with what the portfolio balance approach implies. The significance of the cumulative intervention measure is in agreement with Loopesko (1984), who unfortunately does not report the direction of the effect that she finds. Of course, we cannot claim to have distinguished between a portfolio channel and the possibility that intervention has had a role in signaling new information to the market. For example, an examination of equation (2) confirms that, ceteris paribus, RET and St-1 are positively correlated. The risk premia would be reflected in E[St-I]. However, the forecast error may be correlated with new information that intervention could provide. In Table 11, we split the total cumulative intervention measure utilized in Table I into U.S. and West German purchases. If a portfolio balance channel is operative, the identity of the purchaser should be inconsequential. Thus, we would expect both variables to be significant. Results indicate, however, that only West German purchases of dollars have a significant impact on excess returns (about 10 percent). The sign of the effects is again positive. Tables I11 and IV indicate the results for the excess return of dollars over yen. There is no evidence that intervention has a significant influence. The signal.ingchannel and coordinated versus uncoordinated intervention If intervention works through providing signals to the market, then it need not be cumulative, and it might be ~.?cessaryto distinguish between uncoordinated and unilateral interventions; this study measures both at t-3. If intervention is coordinated (both countries intervene in the same www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm www.clevelandfed.org/research/workpaper/index.cfm
@FedFRASER