Forecasting agricultural commodity prices using machine learning models

Accurate forecasting of agricultural commodity prices is important for market planning, risk assessment, and agricultural policy support. This study develops a historical-price-based forecasting framework for nine agricultural commodities in Maharashtra, India, using mandi-level AGMARKNET data from 2001 to October 2025. Daily market records were cleaned and aggregated into monthly modal prices, and lag, rolling-window, and seasonal calendar features were used to generate one-step-ahead monthly forecasts for January 2021–October 2025. To avoid temporal leakage, outlier limits and preprocessing operations required for modelling were estimated from the available training history within each sequential forecasting step. Random Forest, Extra Trees, Histogram Gradient Boosting, and XGBoost models were compared with one-month Naive and 12-month Seasonal Naive benchmarks. The benchmark evaluation shows that the one-month Naive forecast achieved the lowest MAPE for all commodities, with MAPE values ranging from 2.30% for Paddy to 4.91% for Sunflower. The best machine learning models consistently improved over the Seasonal Naive benchmark but did not outperform the one-month Naive benchmark, indicating strong short-run price persistence in the monthly series. These findings provide a transparent and leakage-safe assessment of historical-price-only machine learning models for agricultural commodity price forecasting and highlight the need for future integration of exogenous market, production, weather, and policy variables.