Exchange rate forecasting: the errors we've really made

We examine the forecasting performance of standard macro models of exchange rates in real time, using dozens of different vintages of the OECD's Main Economic Indicators database. We calculate out-of-sample forecasts as they would have been made at the time, and compare them to a random walk alternative. The resulting \"time series\" of forecast performance indicates that both data revisions and changes in the sample period typically have large effects on exchange rate predictability. We show that the favorable evidence of long-horizon exchange rate predictability for the DM and Yen in Mark (1995) is present in only a narrow two-year window of data vintages around that used by Mark. In addition, approximately one-third of the improved forecasting performance of Mark's monetary model over a random walk is eventually undone by data revisions. Related to this, we find the models consistently perform better using original release data than using fully revised data. Finally, we find that model-based exchange rate forecasts are sometimes better when using Federal Reserve Staff forecasts of future fundamentals instead of actual future values of fundamentals. This contradicts a cherished presumption in the literature that dates all the way back to Meese and Rogoff (1983).