Abstract: In recent years, due to the escalating risks of geopolitical conflicts and changes in trade policies across various countries, the risks and complexities in the agricultural product markets have intensified. Therefore, delving into the underlying mechanisms driving the price fluctuations of agricultural products and accurately predicting the futures prices of these products are fundamental and primary issues for ensuring the stability and smooth operation of China's agricultural supply chain. To address this challenge, this paper proposes a comprehensive interpretable forecasting framework for agricultural futures prices. Specifically, it first employs a multi-source feature fusion method to process the input features of agricultural futures prices, constructing several effective combined and statistical features. The Pearson correlation coefficient method and recursive feature elimination are then used to screen linear and nonlinear fusion features, respectively, thereby providing a reliable data foundation for subsequent predictive modeling. Additionally, a reinforcement learning (RL) enhanced Snow Ablation Optimizer (SAO) is utilized to intelligently and efficiently optimize the parameters of the Temporal Fusion Transformers (TFT), establishing an interpretable predictive model known as RLSAO-TFT. Empirical results demonstrate that the introduction of feature engineering enhances the robustness and accuracy of the agricultural futures price prediction model. The interpretable forecasting method for agricultural futures prices proposed in this paper contributes to ensuring the stable and healthy development of China's agricultural industry chain, thereby supporting the national strategies for rural revitalization and food security.

Key words: Agricultural product prices forecasting, Interpretable forecasting, Feature Engineering, Neural Network, Intelligent optimization.