Applying Machine Learning Techniques to Improve Exchange Rate Forecasting

Abstract

The majority of literature on exchange rate forecasting uses different econometric models and economic variables to attempt to predict the future exchange rates of individual countries. Though, there have been many studies in recent years that have strayed away from these econometric models, and instead have opted to predict using a type of “black box” machine learning approach. A noticeable gap in the literature is that there are no studies that directly compare the usage of econometric techniques with machine learning techniques to determine which are better at prediction. Further, there are no studies that attempt to optimize machine learning models based on the economic theory of exchange rate determination, tweaking variables for different countries based on their specific characteristics. This paper fills these gaps, and examines 14 currency pairs to see whether or not prediction can be improved when employing machine learning techniques compared to econometric techniques. Specifically, it attempts to predict 1-month currency fluctuations using a VAR model and an LSTM network, determining which method is better for prediction in different countries. I find that the LSTM results significantly improve upon the results of the VAR model for most of the currency pairs considered in this paper.