This study presents a comprehensive evaluation of conventional and hybrid forecasting models for predicting seasonal rice production in Bangladesh, analyzing data from 1970 to 2023 across three major growing seasons: Aus, Amon, and Boro. The research implements sixteen distinct forecasting models, combining traditional approaches with innovative hybrid methodologies to enhance prediction accuracy. The conventional models include ARIMA, ETS, ANN, and TBATS, while the hybrid models merge these traditional approaches with advanced techniques such as WARIMA, ARNN, and THETA. The study employs seven performance metrics—MSE, RMSE, MAE, MPE, MAPE, MASE, and SMAPE—to evaluate forecasting accuracy. Results demonstrate that hybrid models significantly outperform conventional approaches, with ARIMA+ARNN achieving the lowest MSE (3.739646), followed by ARIMA+ANN (MSE: 3.859849) and ARIMA+WARIMA (MSE: 5.972818). In contrast, traditional models show higher error rates, with STL+ARIMA exhibiting the highest MSE at 562.6725. The seasonal analysis reveals that hybrid models maintain superior performance across all three growing seasons, providing more reliable predictions for each distinct agricultural cycle. These findings offer valuable insights for agricultural planning and policy formulation in Bangladesh, enabling more accurate seasonal rice production forecasts and supporting food security initiatives.

Keywords

Agricultural forecasting, Seasonal rice production, Hybrid time series models, Conventional time series models,Performance