Comparison of Feature Selection Methods for Wheat Price Forecasting

A systematic comparison of 15+ feature selection methods and a baseline Lasso regression for weekly forecasting of CME Wheat futures (ZW). Rolling-window validation demonstrated that Spearman correlation provides the best balance of accuracy and stability: achieving a MAPE of 2.49%, Directional Accuracy of 59.5%, and a Composite Score of 0.79 — outperforming both simple statistical filters and complex ML wrappers (XGBoost, SHAP). This study provides practical, data-driven recommendations for quantitative researchers on selecting a method based on priorities: accuracy, stability, or computational speed.