<p>Although the existing hybrid model based on decomposition can improve the prediction performance for financial data, it ignores the effective information of the error between the real value and the predicted value. In order to explore the value information of prediction errors, this paper constructs a hybrid secondary decomposition error-corrected model for four commodities prices such as gold, aluminum, whorl, and iron mine in the metal futures market. Firstly, VMD is used to decompose the original price, and then each component is predicted using 11 machine learning models. Secondly, the error sequence between the predicted results and the real price is decomposed using the CEEMDAN method and combined with the deep learning model for prediction. Finally, two predicted sequences are reconstructed to obtain the final one-step result, and based on the result of the one-step prediction, the two-step prediction is made using the sliding prediction method. The empirical results show that compared to the one-time decomposition model, the prediction accuracy of the secondary decomposition error correction model is improved by about 32%, 33%, and 22% respectively under single machine learning, ensemble machine learning, and deep learning models. In addition, the secondary decomposition error-corrected model using the GRU model has the best predictive performance and return on investment. The proposed VMD-GRU-CEEMDAN-GRU model can accurately predict the futures prices and provide valuable guidance for practitioners.</p>