Forecasting the Exchange Rate: A Comparison Between Econometric and Neural Network Models - Proceedings AFIR 1996 - Nürnberg, Germany

In this paper the performance of four linear models of the exchange rate Spanish peseta/US dollar is compared with that of Artificial Neural Networks. The models are a random walk process and three different specifications based on the purchasing power parity (PPP) theory. The aim is to examine whether potentially highly nonlinear neural network models outperform traditional methods or give at least competitive results. The comparative exercise has been conducted both insample and out-of-sample. In general, the results confirm the difficulty in forecasting exchange rates, and reaffirm those obtained in previous literature which show that the performance of econometric models of the exchange rates is inferior to that of a random walk. A similar result is found for the neural networks. In the direct comparison between linear and nonlinear models, the experiment with quarterly data indicates that there is no advantage in the use of NNs for forecasting the exchange rate, while the performance of the NNs clearly improves when they are trained on monthly data.