Explaining Exchange Rate Behavior

In an era characterized by increasingly integrated national economies, the exchange rate is the key relative price in open economies. As such, a great deal of attention has been focused on characterizing its behavior. Unfortunately, it is unclear how much success there has been in predicting this critical relative price. As recently remarked, "There may be more forecasting of exchange rates, with less success, than almost any other economic variable."¹ While this characterization may be quite apt -- a point I will return to later -- it should not prevent us from attempting to identify the empirical determinants of exchange rates, an enterprise separate from forecasting exchange rates.

The Impact of Productivity Changes

The first major line of inquiry I've followed links changes in productivity to changes in nominal and real exchange rates. There is a long and venerable literature that links these two variables theoretically, most notably associated with Balassa and Samuelson. ² In these models, differences in productivity levels between traded and nontraded sectors affect the relative prices of these goods. Further, with traded goods prices equalized in common currency terms, real exchange rates -- which incorporate the prices of nontraded good -- will be affected.

The post-War yen has been the traditional candidate for explanation by this type of model. ³ In addition, the model typically is applied to economies experiencing rapid growth, since such growth often is associated with rapid productivity change in the tradable (manufacturing) sector. Hence, a natural application of the model is to the East Asian countries. Unfortunately, the data necessary for a direct test of the model do not readily exist. Instead, most analyses rely on observations on relative prices to infer the validity of the approach. In order to conduct a direct test, I compiled sector-specific employment and output data for China, Indonesia, Japan, Korea, Malaysia, Philippines, Singapore, Taiwan, and Thailand, and estimated the implied relationships. The time-series evidence did not support the model except in a few cases. Using panel regression techniques adapted to persistent time series, ⁴ I find that the model applies to the set of countries including Indonesia, Japan, Korea, Malaysia, and the Philippines. ⁵ Part of the reason for the limited extent of the finding may be that measured traded goods prices do not appear to be equalized, especially when the prices pertain to bundles of goods that are changing rapidly. After all, the composition of exports of Malaysia today bears little resemblance to that of forty years ago.

Interestingly, there is some evidence that the productivity effect applies even for more developed economies. Louis D. Johnston and I examined sector-specific productivity levels for 14 OECD countries. Using panel cointegration methods, we found that productivity levels did matter for dollar-based real exchange rate levels in the long run, although other factors mattered as well. These other factors included government spending and the terms of trade. In a closely related paper, we found that the same conclusions held for trade-weighted OECD real exchange rates. ⁶

More recently, Ron Alquist and I have examined the behavior of the euro/dollar exchange rate, drawing inspiration from the large literature that ascribed the strength of the dollar and the weakness of the euro to the differing prospects for accelerated productivity growth rates associated with the diffusion of the New Economy. Using aggregate productivity data from 1985 to 2001, we found that productivity was strongly related to the euro/dollar rate. One of the paradoxes of the results is that according to the estimates, each one percentage point increase in the productivity differential between the United States and the eurozone economies results in a real dollar appreciation of between 2 and 5 percent. While other studies have detected effects of a similar nature, the magnitude is somewhat larger than has been found previously. Furthermore, it is hard -- although not impossible -- to rationalize the magnitude of the effect theoretically. A combination of demand side effects and an increase in productivity, localized to the technologies used in the United States, is one interpretation. ⁷

Overvaluation

Models of purchasing power parity, or the Balassa-Samuelson hypothesis, naturally lend themselves to the exercise of determining whether a currency is "overvalued" or otherwise misaligned. Indeed, one implication of the Balassa-Samuleson hypothesis is that the standard practice of measuring misalignments as deviations from linear trends is likely to provide misleading conclusions. In work I conducted in the wake of the crises of 1997-8, I asked whether the East Asian currencies were overvalued, given the possibility that the standard practice was applied inappropriately.

A long-run relationship between exchange rates and relative prices exists for all currencies, with respect to at least one reference currency (dollar or yen) or price deflator (CPI or PPI). My results indicate that the Malaysian ringgit, Philippine peso, and Thai baht were overvalued a month before the baht devaluation in July 1997. On the other hand, the Indonesian rupiah, Korean won, and Singapore dollar appear to have been undervalued. Of these results, the implied undervaluations of the rupiah and won are the most counter-intuitive, since these two currencies suffered precipitous declines in value. Consequently, the widely held view that currency overvaluation was at the heart of each of the East Asian currency crises lacks credibility (although overvaluation probably did play some role).

Real Exchange Rate Behavior and Market Characteristics

A large body of work has sought to characterize the adjustment of the real exchange rate toward its long-run value. Often, the long-run real exchange rate is thought to be what sets the price of identical baskets of goods to be equal, when expressed in common currency terms; this condition often is termed purchasing power parity. The mystery arises from the stylized fact that the adjustment takes longer than what can be rationalized by sticky prices. ⁸ Yin-Wong Cheung, Eiji Fujii, and I merge the literature on real exchange rates with that on industrial organization factors suggested by the New Keynesian literature. ⁹ We calculate real exchange rates sector-by-sector (for example, for chemicals, or for fabricated metal products), and relate the pace at which these real exchange rates revert to their long-run values to the characteristics of those sectors, including the amount of intra-industry trade, size of price-cost margins (a proxy measure for the degree of substitutability of goods) and other factors thought to be important including distance, exchange rate volatility and inflation rates.

The econometric results reveal considerable evidence for the hypothesis that market imperfection is associated with high persistence in deviations from purchasing power parity. In general, the two measures of market imperfection -- a price-cost margin and an index of intra-industry trade -- are significant across different specifications and have a positive impact on real exchange rate persistence. The robustness of the market structure effects stands in stark contrast with the results pertaining to the macroeconomic variables, which can yield coefficient estimates that vary across model specifications, and occasionally have a sign different from what the theory predicts. Overall, our analysis uncovers positive evidence of market structure effects on real exchange rate persistence.

Interest rates, exchange rates and expectations

A common method of predicting asset prices uses market-based indicators. For instance, futures prices often are cited as forecasts of commodities. Forward rates -- agreements set today for a trade of currencies in the future -- would seem to be an ideal indicator for the future exchange rate. Equivalently, according to a no-arbitrage profits condition, when financial capital is free to move, the forward rate equals the current exchange rate adjusted by the interest differential. In reality, forward rates for developed economy currencies typically are biased predictors of future spot rates; indeed, when interest differentials point to a dollar depreciation, the dollar on average appreciates. This well known fact led Jeffrey Frankel and Kenneth Froot to use survey data to assess whether, for the major currencies, this bias was attributable to the existence of a risk premium or to biased expectations. In two important works, they conclude that the expectations of market participants were biased, and further that there was little evidence that the bias in the forward rate was caused by the presence of an exchange risk premium. ¹⁰ Frankel and I examined a larger number of currencies and once again found evidence of biased expectations for 25 currencies over a three-year period. ¹¹ Interestingly, in examining forward rate bias in a larger set of currencies (17), we find somewhat more evidence in favor of an exchange risk premium. ¹² To the extent that one believes that such risk premiums arise from the differentiated nature of the bonds issued by separate governments, the result makes sense. For instance, U.S. and German bonds may be more substitutable than U.S. and Swedish bonds.

Guy Meredith and I ¹³ investigate whether interest rate differentials point in the wrong direction for exchange rate changes for horizons much longer than typically studied: five and ten years, versus the one month or one year used in earlier studies. We find that at these horizons, this perverse correlation largely disappears. While this finding appears robust to a number of variations, its statistical significance has been disputed, given the small number of independent observations in the post-Bretton Woods era (for example, five non-overlapping five-year horizons). Hence, we use panel regressions and confirm the finding.

The interpretation of these results is complicated by the lack of agreement on the origins of the forward rate bias. We propose a model wherein shocks to the interest rate parity relationship (perhaps because of noise traders) spur a central bank reaction function that serves to make exchange changes correlated negatively with interest differentials. Because central banks only can control short-term interest rates, the effect is most pronounced at short horizons. Since long-term interest rates are a weighted average of short-term interest rates, the effect is muted at longer horizons. Further research may illuminate alternative explanations.

What Do Market Participants Think?

The work previously recounted uses empirical methods to discern the determinants of exchange rate movements. Taking a different tack, Cheung and I conduct a survey study of foreign exchange traders in the United States. ¹⁴ Our results indicate that: more than half of market respondents believe that large players dominate in the dollar-pound and dollar-Swiss franc markets, and technical trading best characterizes about 30 percent of traders, with this proportion rising from five years ago. The responses also suggest that news about macroeconomic variables is incorporated rapidly into exchange rates, although the relative importance of individual macroeconomic variables shifts over time. Finally, economic fundamentals appear to be more important at longer horizons, while short-run deviations of exchange rates from their fundamentals are attributed to excess speculation and institutional customer/hedge fund manipulation.

Perhaps, unsurprisingly given the mixed findings regarding purchasing power parity, traders do not view the parity condition as a useful concept, even though a significant proportion of them believe that it affects exchange rates at horizons of over six months. Interestingly, these particular findings do not appear to be location-specific. Ian W. Marsh, Cheung, and I conducted a similar survey of U.K.-based foreign exchange dealers in 1998. ¹⁵ We confirm that many of these characteristics also pertain to that market. Moreover, we find that there is clear heterogeneity of traders' beliefs, but it is not possible to explain the source of these disagreements in terms of institutional detail, rank, or trading technique (for example, technical analysts versus fundamentalists).

Are Exchange Rates Predictable?

One of the key issues dominating the empirical literature is whether exchange rates can be predicted. Previous assessments of nominal exchange rate determination have focused on a narrow set of models typically of the 1970s vintage. The canonical papers in this literature are by Meese and Rogoff, who examined monetary and portfolio balance models. ¹⁶ These papers established the stylized fact that it is extremely difficult to beat a random walk on a consistent basis. Succeeding works by Mark and by Chinn and Meese overturned these results, but only at long (three or four year) horizons. ¹⁷

More recently, several studies have re-evaluated the long-horizon results. Faust, Rogers, and Wright argue that the success of long-horizon regressions is specific to the particular time period examined by Mark and Chinn and Meese. ¹⁸ In work co-authored with Cheung and Pascual, ¹⁹ I also re-assess exchange rate prediction. Using a wider set of models that have been proposed in the last decade -- interest rate parity, productivity based models, and a composite specification incorporating sticky-price, productivity, and portfolio balance models -- we compare these models against a benchmark, the Dornbusch-Frankel sticky price monetary model.

We examine the model's performance at various forecast horizons (1 quarter, 4 quarters, 20 quarters) using differing metrics (mean squared error, direction of change), as well as the "consistency" test proposed in Cheung and Chinn. ²⁰ About half of the estimates are based upon specifications that use contemporaneous information (that is, the forecast of December 2000 uses December 2000 data on the right-hand side variables), while half use only lagged information (that is the December 2000 forecast uses data either one quarter, one year, or 5 years prior.) Consequently, half of our specifications are at a great informational disadvantage.

We find that no model consistently outperforms a random walk, by the conventionally adopted mean squared error measure. However, along a direction-of-change dimension, certain structural models do outperform a random walk with statistical significance. We also find that interest rates predict quite well, although only at the longest horizon.

These forecasts are tied to the actual values of exchange rates in the long run, although in a large number of cases the elasticity of the forecasts with respect to the actual values is different from unity. Overall, we find that model/specification/currency combinations that work well in one period do not necessarily work well in another period. ²¹

While the results are not very positive, they do suggest that along some dimensions, structural models have predictive power. And, it is important to recognize that we have stacked the deck against these models having good predictive power, in that half of our estimates do not rely upon contemporaneous information. So, while useful forecasting models remain elusive, the identification of key empirical factors remains a productive, albeit challenging, enterprise.