Research Method of Futures Price Prediction Based on M2FiT Fusion Model

As global economic uncertainty rises, the futures market has become a crucial tool for investors and enterprises to manage risk. Accurate forecasting of futures prices provides essential decision support for investors and aids enterprises in optimizing production and inventory management. However, futures prices are influenced by multiple factors, and both traditional statistical models and machine learning methods face limitations in addressing nonlinear and non-stationary data. This paper proposes a fusion model, which integrates Mamba2's structured matrix transformation, FRETS' frequency-temporal learning, and iTransformer's sequence processing capabilities. By redesigning the FRETS module, the gradient issue in frequency channel learning is alleviated, improving the model's ability to capture global periodicity and trends in time series data. Experimental results show that the model outperforms others in forecasting various futures commodities, including copper, gold, rubber, and soybean oil, with MAE, MAPE, and RMSE reducing by 16.75% to 52.39%. Moreover, the forecasting results surpass those of existing methods. In addition, this paper presents an interpretability analysis using SHAP, which reveals the roles and synergistic mechanisms of each module in the fused model structure. The SHAP results indicate that Mamba2 and FRETS are complementary in their feature attention and output modeling, providing strong evidence for the rationality of the model design.Overall, the proposed model holds significant promise for futures prediction.