Seasonal-Cycle Aware Temporal Fusion Network for Climate –Driven Agricultural Commodity Price Forecasting

Agricultural commodity markets exhibit strong seasonal and cyclical behavior influenced by climatic variability, production cycles, and dynamic supply–demand interactions. Accurate forecasting of such markets is essential for policymakers, traders, and food security planning; however, traditional econometric and standalone deep learning models often fail to explicitly model recurring seasonal memory and climate-driven nonlinearities. Most existing approaches treat seasonality implicitly and rarely integrate cross-factor attention mechanisms that dynamically balance climatic and supply–demand signals. This study proposes a SeasonalCycle Aware Temporal Fusion Network (SCA-TFNet) for climate-driven agricultural commodity price forecasting. The framework introduces a multi-stream decomposition module for separating seasonal and cyclical components, a Seasonal Memory Gate to reinforce interannual recurrence patterns, and a cross-factor attention mechanism to adaptively weight climate and supply–demand influences. Experiments are conducted using FAO Food Price Index data, historical commodity price series, and climatic indicators such as rainfall and temperature anomalies. The model is evaluated using MAE, RMSE, and Directional Accuracy metrics. Results demonstrate that SCA-TFNet reduces forecasting error by up to 12–18% compared to ARIMA, standalone LSTM, and CNN–LSTM baselines, while improving directional consistency during seasonal transitions. The proposed framework offers a robust and scalable solution for modeling nonlinear seasonal dynamics in agricultural markets.