【作者】 刘红霞；

【导师】 申峰；

【作者基本信息】 西南财经大学 ， 金融工程， 2022， 硕士

【摘要】 2008年全球金融危机后,我国的经济增长速度逐步减缓,同时在全球疫情的强烈影响下,国际政治和经济间的冲突凸显,中国经济下行的压力不断加大,国内经济仍处于复苏状态。在当前的疫情形势下,面临着严峻的生存压力与越来越多的不确定因素,许多公司因疫情期间管理不善等问题导致经营不确定性增加,从而遭受巨大损失,严重者甚至要申请破产。与此同时,公司面临的生存压力也会对银行产生一定程度的影响,如果公司出现资金周转不灵、偿债能力不足的情况,那么公司将无法如期偿还银行贷款,从而导致银行不良贷款率的增加,给整个国民经济的正常健康发展埋下隐患。所以,加强公司财务风险预警工作,并及时采取相应针对性的措施化解危机,提升公司经营绩效,是公司当务之急,也是股东和相关利益者首要之事。面对这样的情况,公司除了需要在经营管理环节下足功夫之外,还需要建立一套完善、高效、准确的财务预警系统。同时,银行也必须建立健全的公司财务预警体系,以确保能够在发放贷款过程中对公司财务状态做出科学的判断,从而有效地筛除贷款违约风险较高的公司,以此确保自身资产安全。梳理贷款违约财务预警的相关文献发现,过去大部分学者对贷款违约的预测都是基于机器学习模型针对静态建模而没有考虑随着时间推移宏观经济条件变化所引发的概念漂移问题。因此,本文基于实证角度从贷款违约财务预警出发,对我国上市公司银行贷款违约进行财务预警,构建基于时间窗口机制和自适应动态特征选择的贷款违约动态预测模型,从样本选择和特征选择两方面处理概念漂移问题。本文从CSMAR数据库收集到上市公司贷款逾期数据,将贷款逾期样本定义为贷款违约,共收集到2003年-2020年因贷款逾期的1531家上市公司作为贷款违约样本,同时运用得分倾向匹配法遵循行业相同、资产规模相近的原则匹配相对应的贷款正常的上市公司,建立初始预测样本集。在指标体系构建上,本文借鉴国内学者指标体系研究的五项基本原则,选用了涉及偿债能力、每股指标、发展能力、盈利能力、结构指标、风险水平、经营能力和现金流分析八个方面的37个备选财务指标构成实证所需初始数据集。正是由于贷款违约预测中的财务指标数据来自于不同年份且年份跨度较大,随着时间推移会发生不可预测的变化,可能会使原有的分类模型分类不准确,从而使预测模型无法正确判断财务状况,本文先进行静态基分类器构建择优选出最优基分类器,并将其用于构建基于时间窗口的样本动态更新和自适应的特征动态更新的动态预测模型,并用AUC、Accuracy、Recall、F值和G值对比评价静态与动态预测模型的性能。本文的实证研究结果表明,对贷款违约的财务预警来讲,随机森林静态预测模型相对于其他静态模型是综合评价最高、最稳定的基分类器;本文构建的基于时间窗口机制和自适应动态特征选择的动态预测模型比传统静态模型预测效果好,能有效解决预测过程中的概念漂移问题,其中,窗口宽度为5的时间窗口比全记忆时间窗口动态预测模型效果更好。最后,本文在研究结论的基础上,从公司和银行角度提出了相关政策建议。更多还原

【Abstract】 After the global financial crisis in 2008,China’s economic growth has gradually slowed down.At the same time,under the strong impact of the global epidemic,the conflicts between international politics and economy have become prominent,the downward pressure on China’s economy has continued to increase,and the domestic economy is still restarting.Chinese enterprises are facing a complex business environment.Many enterprises are under great pressure and increasing business uncertainty.Many enterprises gradually fall into financial crisis due to poor management during the epidemic,and even have to apply for bankruptcy.At the same time,the operating pressure faced by enterprises will also be transferred to the banking industry to a certain extent.Enterprises cannot repay loans on schedule due to financial difficulties,which will lead to the increase of bank non-performing loans and lay a huge hidden danger for the normal and healthy operation of the economic system.Thus,the top priority for shareholders and relevant stakeholders is to do a good job in the early warning of the company’s financial risks,take timely targeted measures to resolve the crisis and improve the company’s performance.In the face of such a situation,in addition to making great efforts in the operation and management,it is also very necessary for the company to establish a efficient,perfect and accurate financial early warning system.For banks,it also needs a perfect enterprise financial early warning system to help them make scientific decisions when issuing loans,screen out enterprises that may fall into loan default and protect the safety of banks’ own assets.Combing the relevant literature on financial early warning of loan default,it is found that in the past,most scholars’ prediction of loan default was based on machine learning model and aimed at static modeling,without considering the concept drift caused by the change of macroeconomic conditions over time.Therefore,based on the empirical perspective,starting from the financial early warning of loan default,this paper carries out financial early warning of bank loan default of Listed Companies in China,constructs a dynamic prediction model of loan default based on time window mechanism and adaptive dynamic feature selection,and deals with the problem of concept drift from two aspects of sample selection and feature selection.In this paper,the loan overdue data of listed companies are collected from CSMAR database,and the loan overdue sample is defined as loan default.A total of 1531 listed companies with overdue loans from 2003 to 2020 are collected as loan default samples.At the same time,the score tendency matching method is used to match the corresponding listed companies with normal loans in line with the principles of the same industry and similar asset scale,Establish the initial prediction sample set.In terms of the construction of the index system,this paper draws lessons from the five basic principles of domestic scholars’ index system research,and selects 37 alternative financial indicators related to solvency,structural indicators,operating capacity,profitability,cash flow analysis,risk level,development capacity and per share indicators to form the initial data set required for demonstration.It is precisely because the financial index data in the loan default prediction comes from different years and the year span is large,unpredictable changes will occur over time,which may make the classification of the original classification model inaccurate,so that the prediction model can not correctly judge the financial situation.Firstly,this paper constructs a static base classifier and selects the best base classifier,It is used to construct the dynamic prediction model of sample dynamic update based on time window and adaptive feature dynamic update,and the performance of static and dynamic prediction models is compared and evaluated by AUC,accuracy,recall,F value and g value.The empirical results of this paper show that the Stochastic Forest static prediction model is the highest and most stable base classifier compared with other static models for financial early warning of loan default;The dynamic prediction model based on time window mechanism and adaptive dynamic feature selection constructed in this paper has better prediction effect than the traditional static model,and can effectively solve the problem of concept drift in the prediction process.Among them,the time window with window width of 5 is better than the full memory time window dynamic prediction model.At last,this paper puts forward relevant policy suggestions from the perspective of companies and banks.更多还原