A Saliency-Fused Spatio-Temporal Graph Neural Network for Systemic Risk Early Warning in China’s Financial-Real Economy

Traditional graph neural networks (GNNs) often overlook two key phenomena when modeling the dynamic interactions between the financial and real economy sectors: node heterogeneity and the abrupt clustering of systemic risks. To address these challenges, we propose the Saliency-Fused Spatio-Temporal Graph Neural Network (SFST-GNN). Its key methodological design is a Saliency-Fused Spatial Attention Encoder that dynamically computes a saliency score for each node based on its temporal dynamics and integrates this score into the model in two complementary ways. First, during spatial attention computation, the encoder incorporates the saliency of neighboring nodes, enabling the model to prioritize risk signals from systemically important entities. Second, after information aggregation, it amplifies each node’s updated representation using its own saliency, simulating the "risk amplification effect" of critical nodes. In this manner, SFST-GNN can more precisely capture the asymmetric impacts of key nodes within the risk transmission network. Experiments conducted on a dataset of China’s financial-real economy system demonstrate that the model achieves superior performance in predicting risk events defined by statistical thresholds, across both 7-day and 15-day early-warning time windows. Furthermore, the dynamic saliency scores generated by the model successfully foreshadow the risk accumulation prior to the 2015 stock market crash, providing an intuitive, physically interpretable validation for the early-warning results.