Abstract: Aiming at the prediction difficulties caused by the nonlinear and non-stationary characteristics of cotton futures prices and multi-factor coupling, a prediction model integrating multivariate variational mode decomposition (MVMD), Bayesian optimization (BO), and bidirectional Mamba was proposed, with the cotton futures settlement price as the prediction target variable. First, the orthogonal projection maximum information coefficient was used to screen key features; MVMD was adopted to jointly decompose the multi-dimensional feature sequences, extracting collaborative features and suppressing non-stationary noise. Then, the hyperparameters of the bidirectional Mamba model were optimized by the BO algorithm to avoid the blindness of manual parameter tuning.Finally, the optimized model was applied to the bidirectional temporal dependency prediction of each component. The results showed that in the multi-factor scenario, the comparative experiment indicated that the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) of the model were reduced to 58.05 yuan/t, 46.04 yuan/t, and 0.34%, respectively, and the coefficient of determination (R²) increased to 0.978, with prediction accuracy superior to that of the MVMD-BiMamba, BO-BiMamba, BiMamba, VMD-BO-BiMamba, Mamba, Transformer, and LSTM benchmark models.

Key words: multivariate variational mode decomposition (MVMD), Bayesian optimization (BO), bidirectional Mamba, settlement price, multi-factor, cotton futures price, prediction method