Predictive models and directional strategies for the USD/COP pair

Abstract

This paper presents a comparative study of machine learning models for the directional prediction of financial time series, taking the USD/COP pair as a case study. From 29 macrofinancial and market variables, nine models belonging to three families were evaluated: regularized linear models, gradient boosting  models and hybrid models aimed at mitigating the collapse of variance over seven prediction horizons between 1 and 30 business days. The evaluation combined statistical and machine learning metrics with economic performance indicators, prioritizing the directional utility of the signals over the magnitude estimation error. The results show that Ridge obtained the best average performance, with a Direction Accuracy of 52.3 percent, an average Sharpe Ratio of 1.60 and an average Profit Factor of 1.29. Likewise, the  pure gradient boosting  models presented Variance Ratio values below 0.1 and negative economic performance, which shows relevant limitations in financial contexts with a low signal level. A central finding is that traditional metrics such as the coefficient of determination are not very informative to evaluate strategies based on return prediction, so it is recommended to complement them with directional and economic metrics. The entire analysis is based on a reproducible processing and experimentation flow that integrates intake, training, recording of experiments and generation of predictions.