Forecasting the U.S. Dollar in the 21st Century

We find that the level of the exchange rate appears to have strong forecasting power for dollar exchange rates against major currencies post-2000 at medium and long horizons in-sample and out-of-sample, using conventional asymptotic statistics correcting for serial correlation biases and non-standard distribution of parameters. However, we find using bootstrap methods that consider the small-sample biases in inference, the random walk cannot be rejected. Inferring small-sample properties of forecasts based on economic models is problematic since they require modeling the data generating processes of macroeconomic variables under the null. We propose two alternatives. First, using asymptotic statistics, can model-based forecasts “beat” the forecasts based on the level of the exchange rate? Second, we suggest a method that is not subject to some of the problems of modeling macro variables under the null. Forecasts based on “global risk” measures exhibit little support for exchange-rate predictability using our alternative tests