Abstract: With the rapid development of financial technology, the financial industry has experienced or is undergoing major changes at many levels. In the field of financial risk management, due to the increasing complexity of the modern financial system, the limitations of traditional risk modeling methods have become increasingly prominent, while machine learning methods are good at capturing the complex nonlinear relationship between variables, have many inherent advantages over traditional economic analysis and prediction technologies, and therefore can better meet people’s modeling requirements and analysis and prediction demands for economy and finance, a typical complex and open giant system. So, we aim to give a more effective risk analysis system using machine learning methods.
This paper proposes a new systemic financial risk monitoring and early warning system based on machine learning techniques, selecting early warning indicators from eight levels: economic fundamentals, money supply, fiscal conditions, securities and interest rate markets, price indices, foreign exchange and exchange rate markets, leverage and banking system, and using five classical machine learning models and its integrated models to forecast systemic financial risk. In order to open the black box of machine learning, we deconstruct the machine learning early warning model using feature importance, and partial dependency plots(PDP). Feature importance is commonly used in tree models to analyze the importance of variables, while PDP is applied to different models, and its core idea is to examine the effect of different values of a feature on the output value of the model. The PDP method can not only identify the relative importance of variables, but also examine the non-linear effects of variables. Our sample interval is from January 2005 to December 2020, and all raw data are obtained from the Wind database.
The research results show that: (1)Compared with traditional linear models, machine learning models are good at capturing nonlinear relationships, and perform well both in and out of the sample. (2)Compared with Lasso model, SVM and other single model, integrated models have better prediction capabilities by improving the robustness of prediction results. (3)PDP model can effectively identify the nonlinearity and importance of features, thereby helping to open the black box of machine learning; among all the early warning variables, exchange rate, money supply, market interest rates and industrial product prices are key factors affecting systemic financial risks. Monitoring these key variables will help prevent systemic financial risks in an early stage. Our research work is helpful for promoting the application of artificial intelligence in the field of finance by providing a new technical framework for the systemic risk monitoring and early warning.

Key words: systemic financial risk, risk warning, machine learning, nonlinear modeling

摘要： 相对于传统的经济建模手段,机器学习方法具有良好的非线性建模性能等内在优点,从而显示出巨大的应用潜力。本文提出了一个基于机器学习技术的系统性金融风险监测预警体系,并从理论依据和实证检验双重角度给出了解释和验证。具体而言,从经济基本面、货币供给面、财政状况、证券和利率市场、价格指数、外汇和汇率市场、杠杆率及银行体系8个层面选取预测预警指标,并采用5种典型的机器学习模型及其集成模型对系统性金融风险进行预测,研究结果表明:(1)相对于传统的线性模型,善于捕捉非线性关系的机器学习模型,无论是在样本内还是在样本外均表现优异;(2)Lasso模型在向前一期预测时表现最好,SVM模型在向前多期预测时能力更强,集成模型则能兼顾样本内拟合效果和样本外预测能力,具有较强的稳健性;(3)PDP模型可以对特征的非线性性和重要性进行有效识别,从而有助于打开机器学习的黑箱;在所有预测预警变量中,汇率、货币供给量、市场利率以及工业品价格是影响系统性金融风险的关键因素,通过对这些重点变量的监测有助于对系统性金融风险进行早期防范。

关键词: 系统性金融风险, 风险预警, 机器学习, 非线性建模