Exchange rate predictability: fact or fiction?

The present study investigates the factors that affect the forecasting performance of several models that have been used for exchange rate prediction. We provide a quantitative survey collecting 8,413 reported forecast errors and we investigate which forecasting characteristics tend to improve forecasting ability. According to our evidence, predictions can beat random walk when certain types of models and econometric methods are used. In particular, linear specifications based on PPP outperform random walk. Furthermore, higher data frequency and longer forecasting horizon also improve forecasting performance. In this way, we identify under which conditions it is feasible to solve the `Meese-Rogoff' puzzle.