Temporal-Graph Deep Networks for Stock Market Forecasting and Volatility Analysis

Abstract

This paper presents a novel Temporal-Graph Deep Network (T-GDN) designed for comprehensive stock market forecasting and volatility analysis. The proposed framework integrates temporal convolutional mechanisms with graph-based relational modeling to capture both sequential dependencies and inter-asset correlations in dynamic financial environments. Unlike conventional recurrent neural or transformer architectures that primarily model univariate time dependencies, T-GDN treats the financial market as an evolving graph in which nodes represent assets and edges encode dynamic dependencies such as sectoral relations, co-movements, and sentiment similarity. The model incorporates a hierarchical temporal-spatial encoder that combines graph attention layers and dilated causal convolutions, allowing multiscale information flow across time and entity dimensions. To enhance interpretability, a volatility-aware attention module highlights market segments contributing to high predictive uncertainty. Empirical evaluation on benchmark datasets including S&P 500 and CSI 300 demonstrates that T-GDN outperforms existing deep baselines-LSTM, GAT, and Temporal Fusion Transformer-by margins of 6.8% in RMSE and 9.3% in volatility-tracking F1 score. Further ablation analyses verify that graph-temporal fusion substantially improves stability during abrupt market transitions such as economic announcements and policy shifts. The findings suggest that the proposed framework not only advances accuracy in stock prediction but also enhances model interpretability for practical financial decision-making.