Machine Learning in Exchange Rate Forecastingand Portfolio Evaluation

This paper provides a comprehensive analysis of machine learning methods applied in short-term exchange rate forecast and portfolio evaluation. The models we used include: random forest, gradient boosting decision tree (GBDT) and extreme gradient boosting (XGBoost). We provide evidence that machine learning models, particularly for the GBDT model, have general advantages over either random walk or linear regression models in predicting out-of-sample exchange rates and directional changes. We further construct portfolios based on the model forecasts, and find that the GBDT model with all fundamental-based variables would appear to generate the largest benefits for investors. Additionally, we demonstrate the factor of asymmetric Taylor as being the only influencing one of all the machine learning models we used in the paper