Forex forecasting: The critical role of feature selection

Abstract

With the evolution of machine learning (ML), forecasting exchange rate dynamics has gained considerable research attention. This study innovates by integrating ML models with adaptive feature techniques to optimize prediction precision. Using principal component analysis for dimensionality reduction and kernel regression for modeling, we analyze daily closing prices of 20 major currencies in foreign exchange (forex) markets. Results demonstrate that the framework achieves accurate predictions for all currencies, with minimal errors. Notably, prediction accuracy does not correlate linearly with feature quantity. Our hybrid ML-feature selection approach presents a scalable framework with potential applicability to diverse financial forecasting domains beyond exchange rates.