Considerable evidence supports the view that real exchange rates do no
t have unit roots but rather show mean reversion. This is the first pa
per to present systematic results on if and how this mean reversion is
useful for forecasting real rates. Longer time horizons improve the p
erformance of the estimated forecast models relative to the benchmark
driftless random walks. Longer estimation periods particularly improve
performance relative to the benchmark. The forecasting ability in sim
ple statistical models seems genuine, but the models resolve only a sm
all part of the uncertainty regarding the revolution of real exchange
rates. (C) 1998 Elsevier Science Ltd. All rights reserved.