Exchange Rate Prediction with Machine Learning and a Smart Carry Portfolio

Abstract

We establish the out-of-sample predictability of monthly exchange rate changes via machine learning techniques based on 70 predictors capturing country characteristics, global variables, and their interactions. To guard against overfi tting, we use the elastic net to estimate a high-dimensional panel predictive regression and find that the resulting forecast consistently outperforms the naive no-change benchmark, which has proven difficult to beat in the literature. The forecast also markedly improves the performance of a carry trade portfolio, especially during and after the global financial crisis. When we allow for more complex deep learning models, nonlinearities do not appear substantial in the data.