ISANE: An Interpretable Real-time News Event Evaluation Stock Prediction Model

Analyzing messages in the stock financial market helps investors to timely understand market sentiment, thereby reducing investment risks. Prior research typically employed pretrained semantic analysis models to obtain emotion vectors of news events for predicting stock market trends. However, these stock prediction models lack real-time analysis of news events and encounter inaccuracies in text semantics. Moreover, they usually function as black-box models, making it challenging for investors to comprehend prediction logic and enhance it. Therefore, we propose a news event assessment model named ISANE. This model represents news events as graph-structured data and utilizes Graph Attention Networks (GAT) to acquire semantic structural information from news events. Simultaneously, it employs BiLSTM networks to capture dynamic changes in market information. Ultimately, the model combines the semantic structural information of news events with historical market data to evaluate their impact on the stock market and enhance the model’s interpretability by visualizing the structural information of news text. Additionally, we use price fluctuations as labels for news events in model training, enabling the model to adapt more swiftly to market changes. Comparative experiments with existing state-of-the-art (SOTA) models demonstrate an 22% enhancement in our model’s prediction accuracy. However, further empirical testing and validation are necessary to ensure its viability in real-world stock markets.