This paper examines the forecasting power of models based on direct and indirect forecasting methods as applied to the real exchange rate, which is a defined variable. The real exchange rate is a defined variable in the sense that it is related by an identity to three other variables (the nominal exchange rate, the domestic price level and the foreign price level). Direct forecasting is based on modelling a time series on the real exchange rate, whereas indirect forecasting is based on modelling the time series on the individual defining variables. Two models are applied to three U.S. dollar exchange rates: the autoregressive model and Harvey's structural time series model. The bootstrap after bootstrap is employed to correct for bias when the AR model is used and to obtain prediction intervals. The results show that the indirect method does not produce superior forecasting results as compared with the direct method.