Abstract

Agricultural Supply–Demand systems are shaped by dynamic and interdependent factors, including climate variability, market fluctuations, and policy interventions, which challenge traditional forecasting models in both accuracy and interpretability. This study proposes CLIF (Closed-Loop Adaptive Interpretable Forecasting), a framework in which SHAP (SHapley Additive exPlanations)-based explanations actively guide model learning. By incorporating explanation-driven feedback and a temporal forgetting mechanism, the framework enables adaptive feature relevance under evolving market conditions. LASSO is employed for sparse feature selection, while Random Forest captures nonlinear relationships, forming a feedback-driven predictive system responsive to structural shocks and policy changes. Evaluation on a decade dataset of Indian wheat and rice markets demonstrates the feasibility of improving forecasting performance compared with ARIMA, SVM, LASSO, and standalone Random Forest models, achieving the lowest RMSE (0.55) and highest R² (0.91). Feature attribution analysis identifies fertilizer prices, climate variability, and export tariffs as key drivers, providing actionable insights. The framework demonstrates potential as an interpretable and adaptive approach for Agricultural Price Index forecasting and Supply–Demand analysis in dynamic agricultural environments.

Authors

B. Rajmohan, A. Latha
Takshashila University, India

Keywords

Supply Chain, Feature Selection, Interpretable Machine Learning, SHAP, LASSO, CLIF