Forecasting Crop Yield Anomalies on Panel Data via Spatially Demeaned Ensembles

Accurate forecasting of crop yield anomalies is essential for global food security and effective agricultural policy-making, as it enables the separation of climate-driven impacts from baseline regional productivity. This study proposes a transparent and lightweight ensemble forecasting pipeline, termed Lightweight Exponential-Weighted Regression (LEWR), for predicting such yield deviations. To account for time-invariant country-specific characteristics, a cross-sectional demeaning procedure is first applied to the panel dataset. The procedure then integrates heterogeneous regression models—including Random Forest, XGBoost, and Support Vector Regressors—through exponentially decaying weights derived from validation errors, emphasizing higher-performing predictors without employing a second-layer meta-learner. The performance of the proposed method is evaluated on a hold-out test set (2009–2013) drawn from a panel dataset covering multiple countries and years (1990–2013) and benchmarked not only against individual base models but also against established ensemble methods. Results show that the procedure achieves competitive accuracy in forecasting yield deviations compared to more complex methods, while offering notable advantages in simplicity and computational efficiency. Furthermore, it supports incremental updates and provides a straightforward mechanism for quantifying predictive uncertainty—a critical factor in modern environmental risk assessment.