This paper presents a hybrid model for stock market forecasting that integrates news sentiment with historical time series data using Graph Neural Networks (GNNs). The authors construct a heterogeneous graph where nodes represent companies, news articles, and industries, with edges capturing relationships such as company-article associations and industry membership. Historical stock data is encoded via an LSTM, while news titles are embedded using a language model (e.g., BERT). GraphSAGE is employed to aggregate information from neighboring nodes, enabling the model to learn complex interactions between companies, news events, and industries. Two prediction targets are evaluated: a binary direction-of-change label (up/down) and a significance-based label indicating the magnitude of change.

Experiments on US equities and Bloomberg datasets demonstrate that the GNN model outperforms a baseline LSTM, achieving 53% accuracy on the direction-of-change task and a 4% precision gain on the significance-based task. Notably, the study finds that companies with more associated news articles yield higher prediction accuracy, and that news headlines contain stronger predictive signals than full articles, suggesting that concise summaries capture market-relevant information more effectively. The results highlight the value of integrating multimodal data and relational structure for stock market prediction, offering a promising direction for future research in financial forecasting.