Exchange Rate Forecasts at Long Horizons: Are Error-Correction Models Superior?

The out-of-sample forecasting performance of error-correction models of exchange rates is tested on recent monthly data and on annual data from 1900 to 1995. The results for the monthly data set strongly favor the naive model, even when the series are pooled. In the annual data, the best model is usually a regression model of some form, but there is no evidence that a researcher can pick a regression model that outpredicts a naive model more often than not, either by choosing at random or by selecting the model that best fits past data.