Machine Learning Approaches for Predicting United States Crude Oil Prices: Model Selection and Key Determinants

Abstract

This study applies machine learning algorithms to predict U.S. crude oil prices using monthly data from January 2010 to December 2023, incorporating 25 independent variables spanning supply, demand, inventory, speculation, monetary markets, stock markets, commodity markets, and technology indicators. A total of three regularization techniques, six ensemble methods, a Neural Network, and two hybrid models were implemented to evaluate predictive performance. Models were initially run with all variables; significant factors were identified, and parameter tuning was applied to enhance predictive accuracy. Ridge Regression, XGBoost, and a hybrid Neural Network-XGBoost model emerged as the top performers, with XGBoost identifying 21 significant contributors. The hybrid Neural Network-XGBoost model achieved the highest predictive accuracy, with an R² of 99.46% and a Mean Squared Error of 2.31, demonstrating its robustness and effectiveness in forecasting crude oil prices.