This study proposes an ensemble soybean futures price forecasting framework that combines ex-ante feature selection with ex-post interpretability to address the lack of causal verification in feature selection and insufficient interpretability. TreeSHAP and transfer entropy are employed together for two-stage feature selection to provide dual validation of statistical significance and information causality. Three functionally complementary sub-model groups, including gated recurrent units, random forest, and temporal convolutional network are designed, with an evolutionary algorithm combined with reinforcement learning to dynamically optimize ensemble weights. A collaborative interpretability framework combining global attribution and local explanation is established. Empirical research based on monthly data from 2000 to 2023 shows that the mean absolute percentage error on the test set is lower than that of all comparative methods. Interpretability analysis reveals the multi-dimensional driving mechanisms from three levels of strategy, tactics, and market trading conditions, providing decision-making support for government reserve regulation, enterprise supply chain management, and institutional investment allocation.