Toward the next generation of stock movement prediction: GenAI-based multimodal stock movement prediction model

This study presents a novel multimodal approach to stock movement prediction that integrates technical indicators with analyst reports through an innovative architecture combining generative AI and graph neural networks. While traditional approaches often rely solely on quantitative data or simple sentiment analysis, our model uniquely leverages GPT-based summarization and BERT embeddings to complement conventional quantitative analysis with systematic qualitative insights from analyst reports, while using Graph Attention Networks (GAT) to model complex stock relationships. The model is unique in its dynamic modality integration mechanism, which quantifiably measures and adjusts the importance weights between different sources of information, providing unprecedented transparency into the model's decision-making process. This interpretable design allows investors to understand how the model prioritizes different sources of information under different market conditions. Extensive empirical analysis on the CSI 300 universe shows that our approach significantly outperforms traditional methods and benchmark indices across various portfolio configurations. The model delivers superior risk-adjusted returns with lower maximum drawdowns and higher information ratio, validating the effectiveness of our integrated approach. The robustness of this approach was further confirmed by achieving similar outperformance in the KOSPI 200 market.