【作者】 周亮；

【导师】 李红权；

【作者基本信息】 湖南师范大学 ， 统计学， 2022， 博士

【摘要】 当今的世界呈现不确定性显著增强的趋势,经济金融系统备受各种“黑天鹅”和“灰犀牛”因素的冲击,防范化解系统性风险是保障经济社会高质量发展的“硬核”需求。在大数据、移动互联的新时代背景下,经济金融系统内部外部的关联性、互动性、复杂性和波动性都在提升,基于线性思维、有限信息源的经典建模分析方法日益捉襟见肘,不能有效捕捉并刻画系统性风险的复杂性本质。鉴于此,本文基于多源多频信息、机器学习非线性建模分析、时空多维联系的视角,全面、准确、动态地刻画系统性风险压力并寻求有效的监管之道,以期为防范化解金融风险提供新的、更有效的理论基础、分析框架以及具体的技术方法。本文首先从多源混频信息的角度构建了系统性风险压力的综合测度指标。具体而言,我们采用股票市场Co Va R月度数据、上市银行违约概率季度数据、房地产市场房价收入比月度数据及地方债务市场年度数据对各市场的风险进行刻画,利用混频动态因子模型探索了其隐藏的系统性风险压力指标SFR,SFR指标有效提取了股票市场、信贷市场、房地产市场及地方债市场的风险特征,能够较好地反映出2005年以来我国所面临的经济和金融风险状况。时差相关分析、线性回归及门限回归等模型均表明SFR相对单一指标对宏观经济的未来趋势具有更显著的预测能力,且预测能力具有明显的非线性性。系统性风险压力的监测预警方面,本文首先利用机器学习方法构建了系统性风险压力的非线性监测预警体系。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、杠杆率及银行体系选取预警指标,并采用5种典型的机器学习模型对SFR进行预测,实证结果发现能够捕捉到非线性关系的机器学习模型,无论在样本内还是样本外均表现出较强的预测能力。针对机器学习模型广受诟病的黑箱属性,我们采用特征重要性、SDT、PDP和SHAP等方法对机器学习预警模型进行解构,结果发现杠杆率对SFR的影响最大。进行监测预警时,有必要将不同的机器学习模型以及不同的模型解构结果融合使用,方可实现更为精确的预测及提前防范。其次,本文还从经济不确定性的视角探讨了其对系统性风险压力的预测能力。经济不确定性是当前宏观经济面临的重大挑战,对系统性风险有着不可忽视的影响。具体而言,我们利用弹性网络建模方法,基于预警指标体系及28个行业指数序列构造了衡量中国经济不确定性的指数,并检验了其对SFR的监测预警能力。结果发现,我们所构造的经济不确定性指数能够较好地反映中国经济的波动特征,在2009年、2015年以及2020年初达到了三个峰值,分别与次贷危机、2015年经济下滑及股灾,以及2020年初的新冠肺炎疫情相契合;并能够对系统性风险压力进行有效预测,经济不确定性指数对系统性风险压力具有至少一年的预测能力,且企业投资及市场情绪构成了其影响路径。本文最后从时间序列的角度进行了考察。具体而言,通过结构突变检验模型和马尔科夫区制转换等模型对系统性风险压力的时间序列特征进行分析,同时为了避免马尔科夫区制转换过于频繁,我们采用了多种方法对SFR原始序列进行降噪,并挑选出周期稳定性最好的降噪方法进行分析,结果发现:在2007年4月、2009年8月、2011年12月、2015年5月和2018年8月我国SFR出现了5个结构突变点,除2011年12月金融风险略有下降外,系统性风险压力越来越高;SFR可以被划分为风险不断累积的高风险区制,以及风险逐渐释放的低风险区制,我们现在正处于高风险区制,系统性风险压力仍在逐渐累积。总体而言,本文的研究为系统性风险压力的测度、识别及监测预警提供了新的、整体性的视角,同时也为系统性风险的防范和化解提供了经验支持和政策依据。更多还原

【Abstract】 Today’s world shows a trend of significantly increased uncertainty,the economic and financial system is subject to the impact of various "black swans" and "gray rhinoceros" factors,the prevention and resolution of systemic financial risks is to ensure high-quality economic and social development The "hard core" needs.In the context of the new era of big data and mobile Internet,the correlation,interaction,complexity and volatility of the internal and external economic and financial systems are increasing,and the classical modeling and analysis methods based on linear thinking and limited information sources are increasingly overstretched and cannot effectively capture and portray the complex nature of systemic risks.In view of this,this paper provides a comprehensive,accurate and dynamic portrayal of financial systemic risk stress and seeks effective regulatory approaches based on the perspective of multi-source and multi-frequency information,machine learning non-linear modeling analysis and spatio-temporal multi-dimensional linkage,with a view to providing a new and more effective theoretical foundation,analytical framework and specific technical methods for risk prevention and resolution.In this paper,we first construct a comprehensive measure of systemic risk stress from the perspective of multi-source mixed frequency information.Specifically,we use monthly data on Co Va R of stock market,quarterly data on default probability of listed banks,monthly data on house price and annual data on local debt market to portray the risk of each market,and explore its hidden systemic risk stress indicator SFR using a mixed-frequency dynamic factor model.SFR indicator effectively extracts the risk characteristics of stock market,credit market,real estate market and The SFR indicator effectively extracts the risk characteristics of the stock market,credit market,real estate market and local debt market,and can better reflect the economic and financial risk situation faced by China since 2005.The time difference correlation analysis,linear regression and threshold regression models all show that SFR has more significant predictive power for macroeconomic future trends than a single indicator,and the predictive power is significantly non-linear.As for the monitoring and early warning of systemic risk stress,this paper firstly uses machine learning methods to construct a nonlinear monitoring and warning system for systemic risk stress.Specifically,early warning indicators are selected from economic fundamentals,money supply,fiscal conditions,securities and interest rate markets,price indices,foreign exchange and exchange rate markets,leverage and the banking system,and five typical machine learning models are used to forecast SFR.The empirical results find that machine learning models that can capture nonlinear relationships show strong forecasting ability both in and out of sample.To address the widely criticized black box property of machine learning models,we deconstruct machine learning early warning models using feature importance,SDT,PDP,and SHAP,and find that leverage has the greatest impact on SFR.When monitoring and warning,it is necessary to integrate different machine learning models and different model deconstruction results to achieve more accurate forecasting and early prevention.Second,this paper also explores its ability to predict systemic risk pressures from the perspective of economic uncertainty.Economic uncertainty is a major challenge the current macroeconomy facing and has a non-negligible impact on systemic risk.Specifically,we construct an index to measure China’s economic uncertainty based on a system of early warning indicators and a series of 28 industry indices using an elastic network modeling approach,and test its ability to monitor and warn against SFR.We find that our constructed economic uncertainty index can better reflect the volatility characteristics of China’s economy,reaching three peaks in 2009,2015,and early 2020,which coincide with the subprime mortgage crisis,the economic downturn and stock market crash in 2015,and the new crown pneumonia epidemic in early 2020,respectively;and can effectively forecast systemic risk stress,with the economic uncertainty index has the ability to predict systemic risk stress for at least one year,and corporate investment and market sentiment constitute its impact path.Finally,this paper examines it from the perspective of time series.Specifically,the time series characteristics of systemic risk stress are analyzed through models such as structural mutation test model and Markov zone transition,while in order to avoid too frequent Markov zone transitions,we adopt various methods to noise reduction of the original SFR series and select the noise reduction method with the best period stability for analysis.,December 2011,May 2015 and August 2018,there are five structural abrupt change points in our SFR,except for December2011 when financial risk slightly decreased,the systemic financial risk is getting higher and higher;SFR can be divided into a high-risk regime where risks are accumulating,and a low-risk regime where risks are gradually released,and we are now in a high-risk regime where systemic financial risk The systemic financial risk is still accumulating.Overall,the research in this paper provides a new,holistic perspective on the measurement,identification,and monitoring and early warning of systemic risk stress,as well as empirical support and policy rationale for systemic risk prevention and mitigation.更多还原