Abstract: With the accelerated integration of financial institutions such as banks and securities, the possibility of cross contagion of systemic financial risks is significantly increasing. How to effectively prevent and resolve systemic financial risks has become an important issue currently facing China. The key to solving this problem lies in precise measurement and effective warning of systemic financial risks. The existing research on systemic financial risk measurement is relatively mature, but there is still relatively little research on systemic financial risk warning. In limited early warning research, data processing is relatively rough, and risk classification often overlooks the peak and thick tail features of financial data, and early warning results are limited by the inherent defects of a single algorithm. The above greatly limits the effectiveness and applicability of early warning.
In view of this, this article further improves the research on systemic financial risk warning. We have constructed a relatively comprehensive early warning system that includes systemic financial risk measurement, extreme systemic risk warning, and analysis of the importance of warning indicator characteristics. Specifically, first, we introduce the isolated forest algorithm into the classification of systemic financial risks, which has better applicability in the classification of systemic financial risk sequences. Second, the rigor of assumptions and the singularity of models are two key factors that constrain the current research on systemic financial risk warning. This article attempts to break through the limitations of hypothesis conditions through the application of machine learning methods, and overcome the limitations of a single algorithm through the integration of classification algorithms, in order to better achieve early warning effects. We select 7 mainstream classification algorithms as base classifiers and combine them with fuzzy evaluation methods to construct an integrated classification algorithm based on multi algorithm voting. Third, this article provides guidance on how to prevent systemic financial risks from the perspective of the importance of indicator characteristics. Compared with previous studies, analyzing the influencing factors of systemic financial risks from the contribution of individual indicators to early warning results is a new perspective. In addition, we also analyze the impact of feature importance indicators on systemic financial risk from both static and dynamic perspectives.
The research conclusions of this article mainly include the following four aspects: Firstly, multiple methods including CoVaR, MES, and VaR can effectively characterize China’s systemic financial risks, and their changing trends are consistent. Secondly, the isolated forest algorithm can be well applied to the classification of systemic financial risks. The duration of extreme systemic financial risks is the longest during financial crises, followed by 2015 stock market disaster. Thirdly, the integrated classification algorithm constructed in the article is an effective technical means for systemic financial risk warning, and with the help of this algorithm, extreme systemic financial risks can be accurately warned. Data dimensionality reduction can further improve the accuracy of early warning and is more suitable for warning extreme tail events, while an increase in lag phases and classification numbers will significantly reduce the accuracy of early warning. Fourthly, systemic financial risks are closely related to the stock market, banking system, and external markets. The potential drivers of extreme systemic financial risks are fluctuations in stock prices at high levels, high credit risks for banks, and large-scale outflows of cross-border capital. At the same time, stock market fluctuations have a gradually decreasing positive impact on systemic financial risks, while the negative lag impact of interest rate spread changes on systemic financial risks has significantly weakened after the 2015 exchange rate reform.

Key words: systemic financial risk, integrated classification algorithm, risk warning, time-varying

摘要： 利用四种风险测度建立我国系统性金融风险测度体系,基于模糊评价方法与基分类器构建集成分类算法,结合孤立森林对风险的分类预警系统性金融风险,并深入分析预警指标的特征重要性及其对系统性金融风险的静态及动态影响。研究表明:集成分类算法是系统性金融风险预警的一种有效技术手段,数据降维能够进一步提升预警精度且更适用于对极端尾部事件的预警,而滞后期及分类数目的增加则会明显降低预警准确性。预警过程中,股票市场、外部市场及银行体系指标具有较高的特征重要性,这些指标在极端系统性金融风险爆发时期以股市的高位波动及利差收窄为主要特征,同时其对系统性金融风险还具有明显的时变影响。

关键词: 系统性金融风险, 集成分类算法, 风险预警, 时变性