A novel method for corn futures price prediction integrating decomposition, denoising, feature selection and hybrid networks

In this study, a novel ensemble model for predicting corn futures, termed CBA-X-PVW, is introduced. Initially, the model employs the peacock optimization algorithm (POA) to identify the optimal parameter combination for variational mode decomposition (VMD). This step decomposes the corn futures price series into multiple intrinsic mode functions (IMFs). The noise component within the series is filtered out by setting a Pearson correlation threshold. Subsequently, wavelet threshold denoising (WTD) is applied to the reconstructed noise component, combined with the information and residual components, to obtain a denoised price series. Next, XGBoost is utilized for pre-training, selecting the most valuable features based on output weights to minimize feature redundancy. The final step involves inputting the selected features and the denoised price series into a hybrid network comprising a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism to achieve corn futures price prediction. We employ multiple evaluation metrics to compare the forecast accuracy of the proposed model with other benchmark models. Experimental results based on corn futures data from the Dalian Commodity Exchange indicate that CBA-X-PVW outperforms other models (with R2 of 0.9807 and RMSE of 12.96), with statistically confirmed by the D-M test. We also confirm its robustness under different time-sliding windows and training set splits, and additionally discover its long-term prediction ability. Through cross-domain tests with U.S. corn futures and corn starch futures, we conclude that the proposed model has good generalization performance and suggest some policy implication.