The Ability of Artificial Neural Networks to Exploit Non-Linearities by Data Mining Models Compared to Statistical Methods

This paper discusses the question whether Artificial Neural Networks(ANNs) have the capacity to exploit (additional) non-linear information on the basis of exchange rate forecast models selected on linear criteria. This includes a comparison of exchange rate forecasts between ANNs and linear statistical methods and it is asked whether linear relationships serve as an approximation. The focused forecast models are selected by a data mining approach which combines fundamentals from economic theory, respectively building blocks from economic theory, with “fundamentals” derived solely by statistical criteria. This combination of theoretical and statistical relationships in data mining makes sure that both long run determinants on exchange rate behaviour as well as current influences are integrated in the forecast models. The results are evaluated for five exchange rates on a monthly frequency. The results favour the use of ANNs as they slightly improve the out-of-sample performance of the forecast models. What is more, linear as well as non-linear methods can be applied and the advantages from both methods can be used, which means that statistical methods allow a more detailed analysis of the results whereas ANNs offer a slightly better forecast performance.