Out-of-Sample Exchange Rate Prediction: A Machine Learning Perspective

Abstract

We establish the out-of-sample predictability of monthly exchange rates via machine learning techniques based on 70 predictors capturing country characteristics, global variables, and their interactions. To better guard against overfitting in our high-dimensional and noisy data environment, we adjust “off-the-shelf” implementations of machine learning techniques to induce adequate shrinkage. The resulting forecasts consistently outperform the no-change benchmark, which has proven difficult to beat. Variable importance analysis indicates that country characteristics are important for forecasting, once they interact with global variables. Machine learning forecasts also markedly improve the performance of a carry trade portfolio, especially since the Global Financial Crisis.

Keywords: Short-horizon exchange rate predictability, Panel predictive regression, Elastic net, Deep neural network, Variable importance, Partial dependence plot, Individual conditional expectation curves, Carry trade

JEL Classification: C45, F31, F37, G11, G12, G15