Stock Prediction Model Based on Multi-feature Graph Attention Network

With increasing market uncertainty, stock prediction faces challenges such as poor data quality and low accuracy. Based on this, we construct a multi-feature graph attention network model, leveraging the LSTM model to capture sequential features of stocks, the BERT model to extract textual features, and meta-paths to build homogeneous graphs of stocks. The GAT model and attention mechanisms are employed for prediction. The experiments integrate multi-source data, including historical prices, market sentiment, and industry correlations. The results indicate that the proposed model surpasses single models in both accuracy and asset valuation, markedly enhancing predictive performance. This research provides a novel approach to stock prediction and holds practical value.