<![CDATA[Predicting agricultural commodity prices is vital for food security and market stability, but price data is highly non-linear, high-cardinality, and severely imbalanced, rendering conventional accuracy metrics misleading. This study proposes a heterogeneous stacking ensemble of LightGBM, XGBoost, and Random Forest base learners with a LightGBM meta-learner, evaluated against five BorderlineSMOTE oversampling variants and four ensemble configurations under a 282:1 class imbalance. A controlled ablation across six configurations — including a new factorial design that isolates the contributions of oversampling and an MLP base learner — demonstrates that performance is driven by the stacking architecture rather than by synthetic oversampling or deep learning components. The ensemble without oversampling achieves MCC = 0.9509 and AUC-OvR = 0.9908 on the held-out test set, while the addition of an MLP base learner yields no consistent improvement (|ΔMCC| < 0.008 across both SMOTE and no-SMOTE conditions). These findings identify cost-sensitive heterogeneous stacking as an effective and computationally tractable strategy for highly imbalanced tabular agricultural data, and establish that synthetic oversampling and MLP-based components are unnecessary under these conditions.]]>
Copyrights © 2026
