A hybrid neural framework for financial time series forecasting with CEEMDAN-CNN-BiLSTM-attention integration

Purpose Against the backdrop of policy intervention in China’s stock market, this paper proposes a prediction framework integrating frequency-domain multi-scale feature fusion with a multi-head attention mechanism for stock price forecasting. Design/methodology/approach The adaptive noise complete ensemble empirical mode decomposition (CEEMDAN) algorithm is combined with the k-means clustering algorithm to generate feature sequences across different frequency bands. Through Pearson correlation coefficient analysis, feature sequences with strong relevance to target labels are selected as model inputs. These preprocessed features are fed into a hybrid architecture comprising a convolutional neural network (CNN), bidirectional long short-term memory network (BiLSTM) and multi-head attention layers. Findings Extensive experiments on stock datasets from the Shanghai Stock Exchange (SHA) and Shenzhen Stock Exchange (SHE) demonstrate that the model achieves robust prediction accuracy and generalization capability and Welch’s t-test is performed to validate the results statistically. It achieves better prediction results in most datasets. Specifically, in the SHE: 000,021 dataset, the MSE is 0.906 and the MAE is 0.698, which are significantly lower than those of the baseline model. Originality/value This innovative approach integrating decomposition-aggregation algorithms with multi-head attention mechanisms offers a novel technical solution for stock market participants. The study concludes with a discussion on the model’s advantages, limitations and potential optimization directions.