Feature Engineering and Selection for Predicting Foreign Exchange Rates

We propose feature engineering and selection methods tailored for predicting foreign exchange (FX) rates. We selected 132 macroeconomic fundamentals relevant to FX markets and derived technical indicators that capture various aspects of the fundamentals. Next, we selected indicators with the lowest average p-value of multiple Granger causality tests and used them as model input. Experiment results demonstrate that our methods achieved directional accuracy of over 73% for predicting USD/KRW rates in 3-month and 6-month.