客户分组对商业银行个人信用评分模型的提升作用研究  

Research on the Promotion Effect of Customer Grouping on Personal Credit Scoring Model of Commercial Banks

作　　者：张亚京 赵志冲 Zhang Yajing;Zhao Zhichong(Postdoctoral Workstation of Credit Reference Center,the People’s Bank of China,Beijing 100031,China;Postdoctoral Station of the Institute of Finance,the People’s Bank of China,Beijing 100033,China;School of Management Science and Engineering,Dongbei University of Finance and Economics,Dalian 116025,Liaoning,China)

机构地区：[1]中国人民银行征信中心博士后工作站,北京100031 [2]中国人民银行金融研究所博士后流动站,北京100033 [3]东北财经大学管理科学与工程学院,辽宁大连116025

出　　处：《征信》2021年第12期67-71,共5页Credit Reference

基　　金：国家自然科学基金重点项目(71731003);国家自然科学基金项目(72071026);中国博士后科学基金资助项目(2020M680804)。

摘　　要：分组模型是指根据借款人的行为特征分出不同的客群,是信用评分模型开发中的重要一环,可以提升信用评分模型的精度。采用模糊C均值聚类和CART决策树两种方法对全部借款人进行分组,并对分组后的每个客群进行WOE数值转换和逻辑回归信用评分模型的构建,通过对比发现分组后信用评分模型的KS和AUC均有提升,其中模糊C均值聚类作为无监督学习方法也取得较好的模型性能。The grouping model,which refers to the classification of different customer groups based on the behavioral characteristics of borrowers,is an important part of credit scoring model development and can improve the accuracy of credit scoring models.Fuzzy C-means clustering and CART decision tree methods are used to group all borrowers,and WOE numerical conversion and logistic regression credit scoring model are constructed for each grouped customer group.By comparison,it is found that the KS and AUC of the credit scoring model are improved after grouping,and fuzzy C-means clustering as an unsupervised learning method also achieves better model performance.

关 键 词：分组模型 信用评分模型 模糊聚类 CART决策树 

分 类 号：F820.4[经济管理—财政学] F224