Stock price prediction model integrating autoencoder and bidirectional LSTM——Optimization based on attention mechanism

In recent years, financial time series forecasting methods based on deep neural network architectures have been proposed one after another, but there is significant heterogeneity in the performance of each model. This research systematically integrates the feature representation advantages of autoencoders (AE) and the time series modeling capabilities of bidirectional long short-term memory networks (BiLSTM) to construct a new framework for predicting stock market price trends. In terms of technical implementation, a data preprocessing mechanism based on stacked denoising autoencoders is first designed to effectively capture the non-stationary characteristics and implicit patterns of financial time series. Secondly, the bidirectional gated recurrent unit is innovatively embedded in the encoding-decoding architecture to achieve bidirectional time series dependency modeling of multi-dimensional market information, and then a parameter adaptive weighting mechanism is introduced to improve the contribution of prediction-related features through dynamic feature importance evaluation. The final prediction result is obtained through nonlinear mapping of the fully connected layer. Experimental results show that the hybrid model significantly outperforms other baseline models on two stock trading datasets.