This paper provides a robust statistical approach to testing the unbiasedness hypothesis in forward exchange market efficiency studies. The methods we use allow us to work explicitly with levels rather than differenced data. They are statistically robust to data distributions with heavy tails, and they can be applied to data sets where the frequency of observation and the futures maturity do not coincide. In addition, our methods allow for stochastic trend non-stationarity and general forms of serial dependence. The methods are applied to daily data of spot exchange rates and forward exchange rates during the 1920s, which marked the first episode of a broadly general floating exchange rate system. The tail behaviour of the data is analysed using an adaptive data-based method for estimating the tail slope of the density. The results confirm the need for the use of robust regression methods. We find cointegration between the forward rate and spot rate for the four currencies we consider (the Belgian and French francs, the Italian lira and the US dollar, all measured against the British pound), we find support for a stationary risk premium in the case of the Belgian franc, the Italian lira and the US dollar, and we find support for the simple market efficiency hypothesis (where the forward rate is an unbiased predictor of the future spot rate and there is a zero mean risk premium) in the case of the US dollar.
