This paper sought to address the question as to whether the exchange rate can be forecasted more accurately by a monetary model of exchange rate determination or the random walk in the case of the Japan-U.S. exchange rate. The evidence of Meese and Rogoff (1983) on the out-of-sample forecasting performance of structural exchange rate models in comparison to the random walk model portrays a disappointing picture of structural models. I re-considered the issue for the Japanese yen for a more recent period. Besides out-of-sample evidence, within-sample evidence was also examined. The recent work of Phillips and Perron was employed so as to verify that the exchange rate series is well approximated by a random walk model without drift but with time dependent heteroscedasticity. Having established this benchmark, structural monetary models are constructed to see whether one can obtain better within-sample and/or out-of-sample results. It appeared that the random walk can be beaten.