摘要
在特征选择和特征衍生的基础上,分别基于特征扰动和XGBoost与Lightgbm的算法差异建立了四个单一模型;利用单一模型性能确定权重,构建了个人信用风险评估的线性组合模型。实证分析发现,有衍生特征的四个单一模型的AUC和KS均优于无衍生特征的四个单一模型,有衍生特征组合模型的AUC和KS均优于无衍生特征组合模型。实证结果表明,基于特征衍生的组合模型能显著提升个人信用风险评估的预测性能。
On the basis of feature selection and feature derivation,four single models are established based on feature perturbation and algorithm differences between XGBoost and Lightgbm.By using the performance of single model to determine the weight,a linear combination model of personal credit risk assessment is constructed.The empirical analysis has found that the AUC and KS of the four single models with derived features are better than those of the four single models without derived features,and the AUC model and KS model with derived features are better than those of the models without derived features.The empirical results show that the portfolio model based on feature derivation can significantly improve the prediction performance of individual credit risk assessment.
作者
黄宝凤
祁婷婷
Huang Baofeng;Qi Tingting(School of Economics,Nanjing University of Posts and Telecommunications,Nanjing 210023,Jiangsu,China)
出处
《征信》
北大核心
2021年第7期51-57,共7页
Credit Reference
基金
江苏省统计局重点项目(2019A004)
江苏省研究生创新创业项目(SJCX19_0301)。
关键词
个人信用风险评估
特征衍生
组合模型
personal credit risk assessment
feature derivation
portfolio model
作者简介
黄宝凤(1964-),男,江苏泰州人,教授,博士,主要研究方向为经济统计、数据挖掘;祁婷婷(1996-),女,江苏盐城人,硕士,主要研究方向为经济统计、数据挖掘。
