and

Our results complement the recent ¯ndings of real exchange rates as stationary processes. The standard procedure of applying a battery of unit root tests can be problematic since the tests are sensitive to the speci¯cs of the time series process. The novelty of the approach we apply is in emphasizing the information content of the data in distinguishing between the competing processes. Stationary and nonstationary ARIMA processes are ¯tted to the US/UK real exchange rate series covering 134 years. Arti¯cial data following these two processes are generated, and the small sample distributions of the chosen test statistics (including the most powerful point optimal tests with both the unit root and the stationarity as a null) are computed under each of the two hypotheses. The values of the actual sample statistics are shown to be more likely to come from the stationary process than from the non-stationary one