关键词: 金融危机预警, 图神经网络, 信息溢出网络

Abstract: Establishing a scientific financial crisis early warning model is of great significance to prevent and dissolve major financial risks and maintain national financial security. Different from the existing early warning models based on early warning index system, this paper proposes a new graph neural networks (GNNs) early warning model based on information spillover networks. Firstly, this paper uses Elastic-Net-VAR model to construct three kinds of information (risk, volatility, and return rate) spillover networks among Chinese industries, and then uses gated graph sequence neural networks (GGNN) and graph isomorphism networks (GIN) to establish the financial crisis early warning models (including the dual signal early warning models and the three-signal early warning models) with the information spillover networks as the input variables. Finally, the early warning performance of the GNNs models is comprehensively compared with other five models. The empirical result shows that the risk spillover networks and the volatility spillover networks are more sensitive to crisis events than the return spillover networks. Industrial, consumer discretionary, and material industries are at the center of the networks, and are the main systemically important industries. The financial industry is more susceptible to risk spillover from the entity industries. The entity industries are mainly affected as the risk spillover from the upstream supply side of the industrial chain to the downstream demand side. In the dual signal early warning models, the GGNN model based on the risk spillover networks and support vector machine (SVM) model have the best early warning performance. In the three-signal early warning models, the GIN model based on the risk spillover networks has the best early warning performance. The research of this paper provides an operable new tool for regulatory authorities to prevent financial crisis scientifically.

Key words: Financial Crisis Early Warning, Graph Neural Networks, Information Spillover Networks