大型电商平台消费贷产品风控模型的构建  

Construction and Application of Risk Control Model for Consumer Loan Products on Large E-commerce Platforms

作　　者：汪晓曦 马颖[2] WANG Xiao-xi;MA Ying(School of Management,Wuxi Institute of Technology,Wuxi 214121,Jiangsu,China;School of Management,Wuhan University of Technology,Wuhan 430070,Hubei,China)

机构地区：[1]无锡职业技术学院管理学院,江苏无锡214121 [2]武汉理工大学管理学院,武汉430070

出　　处：《武汉理工大学学报（社会科学版）》2021年第5期21-30,共10页Journal of Wuhan University of Technology:Social Sciences Edition

基　　金：2021年度无锡职业技术学院校级社科类课题(SK202102);2020年无锡职业技术学院“青蓝蓝工程”青年教师培养项目。

摘　　要：信用评分模型的构建及应用实施是学术界及工业界不断研究创新的重要课题。从机器学习模型的构建出发,从建模指标、模型构建、模型上线后预期表现几个角度说明了模型应用策略。在建模指标处理方面,采用原始变量woe变换作为XGBoost等模型的输入变量,并验证了模型效果提升;在模型构建上,创新采用多种算法单独建模,用XGBoost算法拟合各单模型输出结果,证明复合模型效果有明显提升;在模型上线后预期表现方面,提出了确定模型切分点的方法,即从授信额度、风险级别以及群体分布三个维度,预测即将进件群体的坏账率。本研究最后结合模型表现及业务模式给出实施建议,即对于部署难度高的复合模型可采用预授信的模式,而LR评分卡模型部署可延用申请-授信的模式。The construction and application of the credit scoring model is an important subject of continuous research and innovation in academia and industry.Starting from the construction of the machine learning model,this article explains the model application strategy from the perspectives of modeling indicators,model construction,and expected online model performance.In terms of modeling indicators processing,the original variable woe transformation is used as the input variable of XGBoost and other models,and it is verified the improvement of the model effect;in the model construction,a variety of algorithms were compared.Among those,XGBoost algorithm was used to fit the output results of each single model,which was proved that the effect of the composite model was significantly improved.The biggest innovation is that it proposes a method to determine the cut-off point of the model,and predicts the bad debt rate of the incoming group from the three dimensions of credit line,risk level and group distribution.In this paper,model selection and deployment are innovatively combined with model performance and business models to give implementation suggestions:for complex models that are difficult to deploy,the pre-credit model can be used;LR scorecard model deployment can be used to the application-credit model.

关 键 词：信用评分 风控模型 XGBoost 机器学习 模型应用策略 

分 类 号：F832.39[经济管理—金融学]