Long-term stock correlation prediction based on CNN-BiLSTM-Attention models

RNN, as a traditional prediction model, can effectively capture the short-term dependencies in the time series, but there are problems of disappearing gradient and explosion. With the proposal of LSTM model, the drawbacks of RNN are effectively solved. BiLSTM belongs to the evolution of LSTM, which is able to capture more features in the time series through bidirectional propagation, and is able to effectively predict the long-term time series. In recent years, factors such as COVID-19, international trade volatility and market changes have had a significant impact on the world economy, which will inevitably affect the volatility of stock prices. In order to reduce the investment risk of global investors, listed companies, financial institutions and consumers, accurate prediction of stock price movements is a good solution. Studies have shown that LSTM models have significant advantages in dealing with complex long-term data, while stock prices are affected by a variety of factors and show a nonlinear trend, which is difficult to be predicted by simple models. Therefore, this paper proposes a stock price prediction model combining Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory Network (BiLSTM) and Attention Mechanism (AM). The CNN can effectively extract the data features while the introduction of the Attention Mechanism can further capture the amount of features that have a greater impact, and finally the BiLSTM model produces accurate prediction results. In this study, the prediction results of the CNN-BiLSTM-Attention model are analysed and compared with those of the CNN and CNN-LSTM-Attention models, respectively. The comparison results show that the CNN-BiLSTM-Attention model has excellent prediction performance, compared with the CNN-LSTM-Attention model, the value of MAE decreases from 5.03 to 2.02, the value of RMSE decreases from 5.76 to 2.74, and the value of R2 increases from 0.9543 to 0.989 for the same dataset. 0.9896.