Ebook: Empirical Modeling of Exchange Rate Dynamics

Structural exchange rate modeling has proven extremely difficult during the recent post-1973 float. The disappointment climaxed with the papers of Meese and Rogoff (1983a, 1983b), who showed that a "naive" random walk model distinctly dominated received theoretical models in terms of predictive performance for the major dollar spot rates. One purpose of this monograph is to seek the reasons for this failure by exploring the temporal behavior of seven major dollar exchange rates using nonstructural time-series methods. The Meese-Rogoff finding does not mean that exchange rates evolve as random walks; rather it simply means that the random walk is a better stochastic approximation than any of their other candidate models. In this monograph, we use optimal model specification techniques, including formal unit root tests which allow for trend, and find that all of the exchange rates studied do in fact evolve as random walks or random walks with drift (to a very close approximation). This result is consistent with efficient asset markets, and provides an explanation for the Meese-Rogoff results. Far more subtle forces are at work, however, which lead to interesting econometric problems and have implications for the measurement of exchange rate volatility and moment structure. It is shown that all exchange rates display substantial conditional heteroskedasticity. A particularly reasonable parameterization of this conditional heteroskedasticity, which captures the observed clustering of prediction error variances, is developed in Chapter 2.

---

This book uses the methods of statistical time-series analysis to characterize the stochastic structure of seven major dollar spot exchange rates, at both weekly and monthly frequencies, during the recent floating-rate regime 1973-1985. While the conditional-mean behaviour of each exchange rate is close to a random walk, the conditional variances are found to have strongly time-varying volatility. Models of autoregressive conditional heteroskedasticity (ARCH) are estimated and used to explain unconditional exchange-rate leptokurtosis (as well as convergence to normality under temporal aggregation), and to provide superior interval predictors. The results are extended to real exchange rates and deviations from purchasing power parity.