基于大样本数据模型的汽车贷款违约预测研究  

Research on Auto Loan Default Prediction Based on Large Sample Data Model

作　　者：舒扬[1] 杨秋怡 

机构地区：[1]华中科技大学经济学院,武汉430074

出　　处：《管理评论》2017年第9期59-71,共13页Management Review

基　　金：中央高校基本科研业务费(2015AC007);华中科技大学研究生创新训练项目(2015650011)

摘　　要：本文运用国内某知名汽车金融公司2014年12月的47138条客户数据,首先运用ROC曲线检验逐步回归功效,再分别建立二值选择模型和计数模型对贷款客户违约状况进行预测,并运用遗传算法对不平衡样本进行一对一匹配,最终得到预测结果。结果表明现存违约评估体系不够有效,客户基本信息、区位、贷款信息、车型、信用状况、房产、贷款期间冲击事件等均会对违约状况产生相应影响。另外,我们得出匹配后的平衡样本预测准确率仍然很高,Logistic模型最适用于客户是否违约的预测,而负二项模型在违约时长的预测中效果更佳的结论。Using the data containing 47,138 customers in December 2014 from a well-known auto finance company in China,this paper first uses ROC curves to test the efficiency of Stepwise Regression,then respectively applies Binary Choice Model and Count Model to predict the default status of loan customers. Afterwards,we apply Genetic Algorithm to do one-to-one matching on unbalanced sample and finally obtain the predicted results. Based on the above analysis,we argue that the current default evaluation system is ineffective,and variables including customers＇ basic information,geographical zone,loan messages,car type,credit status,estate,impact events during loan period all have corresponding impacts to customers＇ default status. Furthermore,the paper finally concludes that balanced sample after matching still possesses superior prediction accuracy rate,that Logistic Model is the most suitable when companies intend to predict whether a customer will default,and that Negative Binomial Model has better efficiency if companies need to know the time length of a customer not paying back.

关 键 词：汽车贷款 违约预测 逐步回归 ROC曲线 二值选择模型 计数模型 遗传算法匹配 

分 类 号：F832.479[经济管理—金融学]