This article reports research that critically examines the qualitative effect of training set size on neural network foreign exchange rate forecasting models. Training data sets of up to 21.75 years of data are used to predict day future spot rates for several nominal exchange rates