A comparison between neural networks and chaotic models for exchange rate prediction

Abstract

Forecasting currency exchange rates is an important financial problem which is receiving increasing attention especially because of its intrinsic difficulty and practical applications. During the last few years, a number of nonlinear models have been proposed for obtaining accurate prediction results, in an attempt to ameliorate the performance of simple random walk models. Among them, neural networks and chaotic models have been used with encouraging results. It is the aim of this paper to provide a comparative evaluation of these two models over common data sets and variables and verify whether they are able to predict better than chance under the same experimental conditions. In particular, the data used in this study were the monthly exchange rates of the four major European currencies from 1973 to 1995. The prediction performance is measured in terms of the well-known normalized mean square error (NMSE) as well as in terms of the statistical significance of the forecasts obtained. To this end, a test statistic proposed by Mizrach has been considered. The experimental results obtained show that neural networks compare favorably with chaotic models, in terms of NMSE and, in turn, both models perform substantially better than that based on the random walk hypothesis. From the statistical significance standpoint, instead, it was found that neural networks' forecasts are statistically equivalent to those yielded by chaotic models and, in most cases, both turn out to be statistically better than those obtained by the random walk.