The out-of-sample forecasting performance of nonlinear models of real exchange rate behavior

We analyze the out-of-sample forecasting performance of nonlinear models of U.S. dollar real exchange rate behavior from the extant empirical literature. Our analysis entails a comparison of point, interval, and density forecasts generated by nonlinear and linear autoregressive models. Using monthly data from the post-Bretton Woods period, there is little evidence to recommend either band-threshold or exponential smooth transition autoregressive models over simple linear autoregressive models in terms of out-of-sample forecasting performance at short horizons. Nonlinear models appear to offer more accurate point forecasts at long horizons for some countries. Overall, our results suggest that any nonlinearities in monthly real exchange rate data from the post-Bretton Woods period are quite ''subtle'' for band-threshold and exponential smooth transition autoregressive model specifications. Further evidence of this is provided by in-sample comparisons of the conditional densities implied by nonlinear and linear autoregressive models.