Comparative analysis of machine learning and time series models for crop volume prediction: evidence from AGMARKNET data for wheat and maize in India

Abstract

Accurate forecasting of crop market arrivals is essential for agricultural supply chain management, food security planning, and insurance premium design in India. This study presents a comprehensive comparative analysis of eight forecasting models — including six machine learning approaches and two classical time series models — applied to monthly wheat and maize market arrivals data sourced from the Agricultural Marketing Information Network (AGMARKNET) covering April 2008 to March 2025. The wheat dataset encompasses 2,426,543 transaction records spanning 24 states, while the maize dataset covers 1,014,415 records across 26 states. Following standardised preprocessing and monthly aggregation, lag-based feature engineering was applied (lags: 1, 3, 6, 12 months) alongside a seasonality indicator. Models evaluated include Ridge Regression, Decision Tree, Random Forest, Gradient Boosting, XGBoost, Extra Trees (machine learning category), SARIMA(1,1,1)(1,1,1)₁₂ and SARIMAX with modal price as an exogenous variable (time series category). For wheat, the Extra Trees Regressor achieved the best performance (MAE = 0.448, RMSE = 0.740 million tonnes), outperforming all other models. For maize, Ridge Regression achieved the best performance (MAE = 0.285, RMSE = 0.606 million tonnes). The contrasting best-model selection between the two crops reveals important insights: wheat exhibits strong non-linear seasonal spikes (Rabi harvest, April–May), favouring ensemble tree methods, while maize arrivals exhibit a more gradual, trend-driven pattern amenable to linear regularised regression. The findings contribute to the emerging literature on machine-learning-based agricultural forecasting in India and have direct implications for revenue protection product design, food procurement planning, and price stabilisation policy.