Incorporating related stock and text for stock price movement prediction based on information fusion

Investors have long been concerned with the analysis of textual information related to target stocks when making stock investments. We collected and extracted stock news headlines from the Chinese stock market and utilized a Large Language Model (LLM) to identify stocks related to the target stock and the relationships among them. Based on these relationships, we constructed a related stock relationship graph. Considering the dynamic changes in the relationship weights between stocks, we employed Graph Attention Networks (GAT) to build a feature fusion model for integrating the features of related stock news. Factors influencing prediction were considered, including different methods of text concatenation, the identification of related stocks, the relationship modeling between related stocks, and the fusion of related stock features. A comparative analysis of the Long Short-Term Memory (LSTM) model, the Bidirectional Long Short-Term Memory (Bi-LSTM) model, and the Long Short-Term Memory with Attention (LSTM-Attention) model revealed that both the news headlines of the target stock and the related stocks, along with their relationships, impact stock price movement prediction. Conversely, ignoring feature fusion and textual feature extraction can negatively affect prediction accuracy.