This paper re-examines the standard ‘unbiasedness’ hypothesis in foreign exchange markets, according to which the forward premium should be an unbiased predictor of the future change of the spot exchange rate. If traders are heterogeneous, they may consist of ‘fundamentalists’ who correctly forecast the future spot rate changes, and ‘chartists’ who use a different information set and pay attention to past exchange rate changes. In this case, the coefficient on the forward premium will deviate from its theoretical value, and will converge to one only if past spot rates are taken into account for the calculation of exchange rate changes and forward premia. The magnitude of the deviation and the speed of convergence are found to depend both on the proportion of each type of traders in the market and on the memory length of ‘chartists’. The analysis is used to explain the bias in the Deutsche Mark/US Dollar exchange rate in the early 1990s and estimates the proportion of traders that is found to be consistent with the deviation from the unbiasedness hypothesis.