Reject Inference In Credit Scoring With Robust Logistic Regression

Nowadays,banks developing credit scoring models tend to take training samples only from approved applicants and exclude those who are rejected in the loan approval process.However,in practice,credit scoring models are applied to all applicants,which will inevitably lead to sample bias and biased parameter estimation.Therefore,the reject inference must be considered and the rejected samples should be included in the establishment process of the scoring model to correct this bias.When the reject inference is made,there are applicants who are misclassified as bad customers in the rejected samples.The classic logistic regression model will be interfered by the misclassified samples,and the goodness of the model will deteriorate.In order to avoid the problem of sample bias and inference error,this paper designs a special reject inference mechanism,which first sets all the rejected samples as defaulting customers,and then uses the robust logistic model to deal with the problem of classification error in the rejected samples.The robust model is mainly based onγ-logistic regression with the idea of robust estimation,which can be directly applied to model the reject samples and has some tolerance to misclassification.By this mechanism,the weight of misclassified data in the rejected samples in the model is reduced by the estimation method of the minimum gamma divergence.The proposed mechanism does not need to assume or estimate the distribution of the rejected samples,which makes the model have better generalization performance than the previous models.The results of simulation and experiment also show with classification error in the rejected samples,the model with robust logistic regression is more accurate and performs better.