Hybrid forecasting of agricultural commodity prices: Integrating machine learning, time series, and stochastic simulation models

This study examines a hybrid forecasting framework to evaluate the predictive performance of time series models (ARIMA, VAR), deep learning (LSTM), and stochastic simulations (GBM, FBM, BB) in forecasting agricultural commodity prices during global crises. Using daily data from 1985 to 2024, the analysis spans nine crisis periods, including the Global Financial Crisis, COVID-19, and the Russia–Ukraine conflict, and focuses on seven major agricultural commodities. Forecasting accuracy (MAPE, RMSE), risk (VaR), and return metrics are used to evaluate model performance. Results show that LSTM outperforms other models in capturing nonlinear dynamics during volatile episodes, whereas ARIMA provides stable results in shorter-term, low-volatility settings. GBM offers the best balance of forecast precision and risk-adjusted returns among stochastic models. In contrast, FBM captures memory effects but produces higher volatility. The findings highlight the importance of adaptive, context-specific forecasting models to enhance policy responses in food security, trade resilience, and agricultural risk management. Figure. The flow chart diagram process. • A hybrid LSTM–GBM–FBM–BB model forecasts agricultural spot and futures prices. • Analysis spans nine macroeconomic and geopolitical crises (1985–2024). • Hurst exponents >0.5 confirm long-term memory in price movements. • GBM yields the lowest MAPE and RMSE in volatile markets. • LSTM provides the best accuracy for soybean price forecasts.