A category-aware and intrinsically explainable mamba for environment-driven crop price forecasting

Abstract To address the problems of strong nonlinearity, pronounced time-varying behavior, and the lack of interpretable modeling in agricultural price series under environmental influences, this study proposes an interpretable time series forecasting method based on Mamba. The method uses Mamba as the backbone, introduces an Environmental Selective State Space module to dynamically select salient environment related latent representations encoded from multidimensional meteorological inputs, and builds an interpretable branch for key feature selection to enhance the modeling of environmental variable contributions. Experiments are conducted on a price and meteorological fused dataset of Fruits, Vegetables, and Grains from 2022 to 2023, constructed from USDA AMS Market News and NCEI. The results show that the proposed method achieves MAE values of 0.1870, 0.0778, and 0.0219, RMSE values of 0.2515, 0.1056, and 0.0288, R 2 values of 0.9319, 0.9406, and 0.9588, and MAPE values of 5.0643%, 5.8601%, and 3.0459% on Fruits, Vegetables, and Grains, respectively. Overall, it outperforms several comparison models and shows good stability and predictive interpretability.