Forecasting disconnected exchange rates

Catalyzed by the work of Meese and Rogoff (1983), a large literature has documented the inability of empirical models to accurately forecast exchange rates out-of-sample. This paper extends the literature by introducing an empirical strategy that endogenously builds forecast models from a broad set of conventional exchange rate signals. The method is extremely flexible, allowing for potentially nonlinear models for each currency and forecast horizon that evolve over time. Analysis of the models selected by the procedure sheds light on the erratic behavior of exchange rates and their apparent disconnect from macroeconomic fundamentals. In terms of forecast ability, the Meese-Rogoff result remains intact. At short horizons, the method cannot outperform a random walk, although at longer horizons the method does outperform the random walk null. These findings are found consistently across currencies and forecast evaluation methods.