信用评级中多类别分类自变量的类合并方法研究

Study on Class Merging Method for Multi-class Categorical Variable in Credit Rating Modeling

信用评级建模中,当多个分类变量的类别较多时会给模型的估计和预测造成较大影响,因此需要对多类别的分类变量进行预处理。结合连续数据离散化的方法,提出基于Fisher精确检验、CACM准则和ACACM准则的有监督类合并方法。分别采用模拟数据和真实的小微企业信贷业务数据对所提出的方法进行分析,结果表明,对多类别的分类变量进行有效的类合并处理不仅有利于简化模型参数,而且有利于提高信用评级模型的分类效果。

The accuracy of parameter estimation and prediction is greatly influenced by multi-class categorical variable in credit rating models,so that it is necessary to preprocess these variables.Based on the idea of discretization,supervised class merging methods are proposed for multi-class categorical variables,which are under Fisher exact test,CACM criterion and ACACM criterion.In the third section,the proposed methods are applied to both simulated data and real credit business data of small and micro enterprises to quantify and compare their effect on the credit rating model.The results show that reasonable class merging for the multi-class categorical variable is beneficial,which not only simplifies the parameter estimation but also improves the predicting accuracy of credit rating model.