Forecasting Crude Oil Futures Prices: A Stacking Model With Mixed-Frequency Predictors, Machine Learning, and Forecast Combination

Abstract

Low- and high-frequency data are widely used in crude oil futures price forecasting, and machine learning techniques exhibit strong performance. To enhance predictive accuracy, this study presents a stacking model that integrates mixed-frequency predictors, machine learning models, and forecast combination methods. Specifically, 18 low-frequency and 20 high-frequency predictors are selected, and 15 distinct machine learning models are applied for prediction. The forecasts of the top 10 individual models are screened based on out-of-sample performance over the past decade, then synthesized using the discounted mean squared prediction error combination method. Against 26 competing benchmarks, empirical results show that the proposed model consistently outperforms all benchmarks, underscoring the advantages of mixed-frequency data, machine learning, and combination methods. It further explores the model's dynamic forecast selection to clarify operational logic. In addition, we confirm its robustness across the business cycle and provide preliminary, small-sample evidence from the COVID-19 pandemic.