Industry-linked stock volatility prediction based on graph neural networks

Abstract

This paper develops a simple hybrid model that combines Graph Neural Networks (GNN) with Light Gradient Boosting Machine (LightGBM) to improve stock market volatility prediction by using links between industries. The model builds an industry correlation graph to extract relationship features through GNN, and these features are then used by LightGBM for volatility forecasting. Based on data from major U.S. market sectors, the proposed model increases R2R^2R2 by 6.5% compared with the baseline LightGBM model and shows lower prediction error during highly volatile periods. The findings show that using industry connections helps capture cross-sector risk transmission and improves both accuracy and stability. This approach can be applied to market monitoring and investment risk control. However, the current version uses a fixed correlation graph and daily data, which limits its ability to adapt to fast market changes. Future studies should focus on building adaptive graphs and combining real-time data sources for better short-term prediction