Forecasting Exchange Rates with Machine Learning Models: Revised and Updated Estimates

With the development of new machine learning techniques, artificial intelligence is increasingly being used to forecast financial time series. Exchange rates are known to be nonlinear, exhibiting stochastic trends. In part for this reason, many models still fail to outperform a random walk. This paper runs forecasting tests for five major currencies at a monthly resolution, over horizons of 1–4 months. Prior to the forecasting experiments, regressions are run for causal factors. Few are found to be statistically significant. Instead, the exchange rates are dominated by serial correlation. There is also evidence of serial dependence in the rate of change, arguing for including lagged differences in the models. Three basic methods are tested, nonlinear regression, neural networks and support vector machines (SVM). All the models are estimated over moving windows of observations. The tests demonstrate that while it is possible to surpass the accuracy of a random walk, the magnitude of the improvement is small. Further, the accuracy of the machine learning models is only somewhat better than that of the regressions. Finally, given the way in which exchange rates have evolved it is essential to retrain the models as new data becomes available.