Forecasting Exchange Rates with Neural Networks: Time Variation, Nonstationarity, and Causal Models

Abstract

There are two major issues in using artificial intelligence to forecast exchange rates, choice of methodology and choice of causal models. A further complication is the nonstationarity of the data. This study compares artificial neural networks, nonlinear regressions and recurrent neural networks, using seven econometric models, in forecasting four major exchange rates over horizons of 1–3 months. The models are trained over moving windows and estimated in both levels and differences. There are three key findings. First, the multilayer perceptron nearly always achieves the most accurate forecasts, with the regressions in second place. The recurrent neural network places a distant third. Second, at horizons of 1 and 2 months, the perceptron is usually better in differences. At the 3-month horizon, however, the accuracy in differences deteriorates. Third, the perceptron favors models including international differentials in price levels, interest rates and yields, which achieve the best forecasts in the majority of cases. Several other models are competitive. One is the familiar Dornbusch-Frankel equation which uses differentials in inflation, output, interest rates and money supplies. Another is a combined model, the Dornbusch-Frankel equation with an additional term for the yield differential. Models using differentials in real interest rates do well in one instance.