基于两阶段思想处理拒绝推断的信用评分模型

本文首先从数据缺失机制的角度分析了信用评分模型的开发和应用中所存在的样本偏差问题,提出了可以用拒绝推断来处理此类问题;然后在曾经被应用于拒绝推断问题处理的Heckman两阶段模型的基础上,提出了用拟似然两阶段模型和广义偏线性模型这两种新的两阶段方法来处理信用评分模型中的拒绝推断问题。经过实证分析发现,应用这两种方法可以得到很理想的结果。另外根据本文的研究,人行征信这类外部数据是拒绝推断最有效的方法,如果此类数据缺乏,则用拟似然两阶段模型和广义偏线性模型是比较有效的拒绝推断方法。

Credit Score Model Based on Two-Step Reject Inference

We first explained that it's necessary to do reject inference in credit score model since the sample bias exists between the modeling sample and application population from the view of data missing. Secondly,on the base of Heckman two-step model,the other two methods named quasi-likelihood and generalized partial linear model are used to deal with reject inference in credit scoring models.The related empirical studies are done and the results demonstrates that the new two methods are effective if the most effective reject inference of credit bureau data is missing.