MA-LSTM: A stock prediction model based on multi-level attention-based LSTM network

With the advancement of deep learning and the flourishing of the economy and finance, intelligent finance has received high attention. Traditional time series models and machine learning models are no longer able to meet the requirements for prediction efficiency. While deep learning models still face technical issues such as gradient vanishing, difficulty in data feature extraction, and low efficiency. For these challenges, this paper proposes a Multi-level Multi-Head Attention Long Short-Term Memory network (MA-LSTM) stock price prediction model. In comparison to traditional models, MA-LSTM enhances data through attention mechanisms and utilizes Generative Adversarial Networks (GANs) for data augmentation, thereby improving prediction efficiency and endowing the model with better generalization capabilities. Finally, experimental research and interpretability analysis of model results are conducted using data from two stocks in the Shanghai and Shenzhen stock markets. The experimental results demonstrate that the MA-LSTM model is more efficient than traditional LSTM-based models.