【作者】 张楠；

【导师】 杨旭；

【作者基本信息】 北京交通大学 ， 应用统计硕士（专业学位）， 2023， 硕士

【摘要】 互联网消费金融是通过为客户提供小额贷款,满足人们各个场景下的信贷需求。自2013年以来,互联网消费金融迅速发展,市场规模不断扩大,但由于我国征信体系尚不健全,在面对数百万用户申请贷款时,互联网消费金融平台难以捕捉客户信息,因此不能识别及量化客户信用风险,2022年互联网消费金融行业平均不良贷款率达3%,面对较高的不良率,互联网消费金融公司的风险控制显得尤为重要。因此,互联网消费金融公司需要在互联网高速发展的背景下,面对互联网产生的海量多维数据,建立一套高效准确的互联网消费金融风控模型从而控制信用风险。目前互联网消费金融公司量化信用风险的风控模型主要是基于比较传统的逻辑回归算法构建的风控模型,但是面对数百万客户申请贷款时,在互联网高速发展的背景下产生的高维数据,逻辑回归算法构建的风控模型不仅运行速率较低,而且模型的预测能力下降,不能很好的识别互联网消费金融公司的信用风险。本文的研究重点则针对互联网消费金融公司面对的信用风险评估问题,不仅应用因果推断理论从海量维度中筛选变量,而且选择了机器学习中的LightGBM算法,针对互联网消费金融传统模型中只采用通过样本建模,未考虑拒绝样本产生的样本选择偏差问题,创新性地对Heckman两步法进行改进,构建基于Heckman两步法进行样本选择偏差修正的互联网消费金融风控预测模型,在选择模型评价指标时选取了KS、AUC指标来评价。本文基于某消费金融公司客户信贷数据,分析了互联网消费金融的特征及信用风险,搭建全流程的大数据风控模型进行实证研究,发现大数据背景下利用Heckman两步法进行样本选择偏差修正和LightGBM算法构建的互联网消费金融风控模型KS、AUC都有提升,能有效量化借款人的信用风险,实现对借款人违约概率的预测及控制,从而降低互联网消费金融信用风险。本文研究成果对互联网消费金融信用风险控制提供了一定的思路和选择,具有一定的参考意义。更多还原

【Abstract】 Internet consumer finance is to provide customers with small loans to meet people’s credit needs in various scenarios.Since 2013,Internet consumer finance has developed rapidly,the market scale has been expanding,but because China’s credit system is not yet perfect,in the face of millions of users to apply for loans,Internet consumer finance platform is difficult to capture customer information,so can not identify and quantify customer credit risk,in 2022 Internet consumer finance industry average non-performing loan rate of 3%,in the face of a higher defect rate,Internet consumer finance company risk control is particularly important.Therefore,Internet consumer finance companies need to establish an efficient and accurate Internet consumer finance risk control model to control credit risk in the face of the massive multi-dimensional data generated by the Internet under the background of the rapid development of the Internet.At present,the risk control model of quantitative credit risk of Internet consumer finance companies is mainly based on the traditional risk control model constructed by the logistic regression algorithm,but in the face of the high-dimensional data generated in the context of the rapid development of the Internet when millions of customers apply for loans,the risk control model constructed by the logistic regression algorithm not only has a low operating rate,but also the predictive ability of the model is reduced,which cannot well identify the credit risk of Internet consumer finance companies.The research focus of this paper is to focus on the credit risk assessment problem faced by Internet consumer finance companies,not only applying the causal inference theory to screen variables from massive dimensions,but also selecting the LightGBM algorithm in machine learning,and innovatively improving the Heckman two-step method for the traditional model of Internet consumer finance that only uses sample modeling and does not consider the sample selection bias caused by rejecting samples.An Internet consumer finance risk control prediction model based on Heckman’s two-step method for sample selection bias correction was constructed,and KS and AUC indicators were selected for evaluation when selecting model evaluation indicators.Based on the customer credit data of a consumer finance company,this paper analyzes the characteristics and credit risk of Internet consumer finance,builds a fullprocess big data risk control model for empirical research,and finds that the sample selection bias correction using Heckman two-step method under the background of big data and the Internet consumer finance risk control model KS and AUC constructed by LightGBM algorithm are improved,which can effectively quantify the credit risk of borrowers and realize the prediction and control of borrowers’ default probabilities.So as to reduce the credit risk of Internet consumer finance.The research results of this paper provide certain ideas and choices for the credit risk control of Internet consumer finance,and have certain reference significance.更多还原