Multifactor Model with Deep Learning for Currency Investments

In this study, we tackle multivariate currencies return prediction for the major 5 currencies against the U.S. dollar with a deep learning based multi-factor model. There are two approaches for predicting financial asset returns: time-series and cross-sectional analysis. While time-series analysis is better for short-term absolute return predictions, cross-sectional analysis is known to be effective for medium to long-term periods of relative return prediction. In cross-sectional analysis, ‘factors’ explain asset returns in the cross-section, and multi-factor models, which involve regression between these factors and future returns, are commonly used. We propose cross-sectional analysis with deep learning based multi-factor model for exchange rate prediction. In order to make it possible to determine whether the exchange rate will rise or fall against the U.S. dollar, we add USD/USD data for the baseline prediction. This approach should be effective for predicting exchange rates because foreign exchange prices are decided depending on the relative strength of each country. We confirm that the multi-factor model predicts whether exchange rates will rise or fall with statistically significant accuracy. Additionally, we ensure incorporating the correlation between the gold price and exchange rate as a feature allowed for maintaining predictive accuracy in high-risk scenarios and gaining a high-performing long-short portfolio.