Inferring the Outcomes of Rejected Loans: An Application of Semisupervised Clustering

Abstract

Summary Rejection inference aims to reduce sample bias and to improve model performance in credit scoring. We propose a semisupervised clustering approach as a new rejection inference technique. K‐prototype clustering can deal with mixed types of numeric and categorical characteristics, which are common in consumer credit data. We identify homogeneous acceptances and rejections and assign labels to part of the rejections according to the label of acceptances. We test the performance of various rejection inference methods in logit, support vector machine and random‐forests models based on data sets of real consumer loans. The predictions of clustering rejection inference show advantages over other traditional rejection inference methods. Inferring the label of the rejection from semisupervised clustering is found to help to mitigate the sample bias problem and to improve the predictive accuracy.

Combating Sampling Bias: A Self-Training Method in Credit Risk Models

“…The final training set sizes among different methods in Table 2 also implies the need for fine-grained selection for inferred training set as discussed in other literature (Li et al 2020). Introducing more pseudo-labels into training set does not guarantee better classification performance.…”

“…They combine accepted applicants with their repayment and rejected applicants with estimated performance into inferred datasets and generate reject inference scoring models. Recent ML based works have proposed new models to assign labels to the rejects from the angle of semisupervised learning, such as semi-supervised SVMs, selflearning, and K-prototype clustering (Li et al 2017(Li et al , 2020Kozodoi et al 2019). One common practice used in the industry is to obtain external loan performance data from credit bureaus for rejected applicants, though it is relatively costly.…”

“…However, empirical results and reviews agreed that the common reject inference methods now do not outperform credit risk models with only accepted loan performances (Hand and Henley 1993;Ehrhardt et al 2021).Moreover, recent model-based methods like clustering have application limitations, such as having good performances on low dimensional data only according to theoretical findings. Datasets used in these studies are usually oversimplified as low dimension and relatively small in size as well (Li et al 2017(Li et al , 2020.…”

Social Network Analytics for Supervised Fraud Detection in Insurance

“…We would like to explore such methods in order to accommodate for the low ratio of labeled claims, which is an inherent problem we faced in our research. Such an approach was developed among others for rejected loan applications by Li, Hu, Li, Zhou, and Shen (2020) and could potentially help in finding the rare fraudulent claims in our case. Next, graph‐based anomaly detection methods would be an interesting approach to find the rare fraudulent claims in the network.…”

Characterizing Fairness Over the Set of Good Models Under Selective Labels

“…In contrast, a large literature advocates for various "reject inference" procedures that incorporate information from rejected applications in model construction. Reject inference procedures involve a class of methods for training a risk assessment on imputed outcomes for all applicants using augmentation, reweighing and extrapolation-based approaches [36,37]. There is considerable debate as to the performance improvements of reject inference over simply training a model on the observed population, but reject inference methods remain popular in credit scoring settings [38,39,32,33].…”