LLM-Augmented Enhanced Graph Transformer for Stock Movement Prediction

Predicting stock price movements remains challenging due to the complex interactions and dynamics of financial markets. Recent deep learning advances, particularly integrating numerical data with linguistic analysis via large language models (LLMs), have shown promise. This study proposes an LLM-Augmented Enhanced Graph Transformer that combines LLM-generated financial analyses, FinBERT semantic embeddings, and a Graph Transformer to predict daily stock movements for 260 selected S&P 500 stocks in 2024. We construct a static stock relationship graph based on the cosine similarity of aggregated textual embeddings, capturing long-term semantic dependencies while integrating numerical indicators. Experimental results show our approach outperforms traditional time-series models (e.g., LSTM, Transformer, Informer) and graph-based methods (e.g., GCN, GAT), demonstrating the effectiveness of multimodal fusion and graph-based attention. We also discuss computational constraints and the limitations of static graphs, highlighting future directions such as dynamic graph modeling and optimized text processing.