A0371
Title: Sweet, crude, and golden: Superior commodity price forecasting with penalized linear methods
Authors: Juan Laborda - Universidad Politecnica Madrid (Spain) [presenting]
James Chen - Michigan State University (United States)
Charalampos Agiropoulos - University of Piraeus (Greece)
Abstract: A structured framework is proposed for commodity price forecasting, balancing interpretability, predictive accuracy, and computational efficiency. A four-stage progression is constructed: (1) A CAPM+4 model extending the capital asset pricing model beyond volatility to skewness and kurtosis; (2) penalized methods of linear regression (Ridge, Lasso, ElasticNet) to mitigate multicollinearity and overfitting; (3) Gaussian process regression (GPR) for quantifying uncertainty; and (4) machine learning ensembles (random forests, extra trees, XGBoost). Prices for five commodities are forecasted: Brent, natural gas, copper, gold, and sugar. Penalized linear methods, notably a Bayesian implementation of the L2 penalty, outperform both OLS and machine ensembles, while maintaining computational tractability and economic interpretability. Additional innovations include RAFE (regularization-adjusted factor estimation), a novel diagnostic derived from Bayesian ridge hyperparameters that identifies periods of market stress. Furthermore, stacked generalization produces a meta-forecast. Findings challenge the presumption that complex ML models dominate financial forecasting. Penalized linear methods hit the "sweet spot", combining performance, speed, and transparency. The framework provides clear criteria for model selection based on interpretability, data heterogeneity, and computational constraints.
