STSGAT: a Spatio-Temporal Sentiment Graph Attention Network for stock volatility forecasting

Stock volatility forecasting is important in financial markets, as it supports risk management, portfolio allocation, derivative pricing, and investment decision-making. However, traditional statistical and time series models often struggle to capture nonlinear dynamics, cross-asset relationships, and external information influencing market behaviour. Although machine learning and deep learning methods have improved prediction performance, many existing models still rely mainly on historical market data and insufficiently incorporate financial news and investor sentiment. To address these limitations, this study proposes a hybrid Spatio-Temporal Sentiment Graph Attention Network (STSGAT), combining Long Short-Term Memory (LSTM) and Graph Neural Networks (GNNs). LSTM captures temporal dependencies within each stock, while GNN models structural relationships and information propagation across stocks. VADER-based sentiment scores from financial news are incorporated to account for both market structure and news-driven emotional dynamics. Overall, this study develops an integrated STSGAT framework combining temporal learning, graph learning, and sentiment analysis. Empirical results show that the proposed model outperforms baseline models, providing more accurate guidance for financial investment strategies and risk-related decision-making.