On Forecasting Exchange Rates Using Neutral Networks

In this paper we consider forecasting daily exchange rate returns using neutral network models (NNs). Based on simulations, we argue (i) that neglected GARCH does not lead to spuriously successful NNs and (ii) that if there is nonlinearity in the conditional mean, NNs will exploit this for improved forecasting. In contrast, for our sample data we find that NNs do not yield favorable results. Our results raise several quesitons about the nature of possible nonlinearity in exchange rates.