Forecasting Crude Oil Prices with a Structural Machine Learning Model

This study proposes a structural machine learning methodology that integrates both linear and nonlinear relationships for crude oil price forecasting. By employing a partially linear machine learning model that explicitly captures the linear effects of key variables influencing crude oil prices, the approach enhances interpretability while evaluating predictive performance. In addition, this study investigates the impact of hyperparameter selection on forecasting accuracy, with a particular emphasis on subsampling and random seed effects–factors that have received limited attention in existing empirical research. Subsampling is actively utilized as a hyperparameter to explore variations in predictive performance, and instead of relying on a single fixed random seed, multiple seeds are used to assess the model's stability and robustness. Based on monthly forecasting experiments spanning approximately 11 years, the results demonstrate that the partially linear machine learning model, when optimized through appropriate subsampling and hyperparameter tuning, outperforms benchmark models in one- to three-step-ahead forecasts. Furthermore, an analysis of prediction error distributions across different random seeds confirms the robustness of the model's predictive performance.