Reject Inference Methods for Retail Credit Scoring

Summary

The document introduces reject inference in retail credit scoring: estimating default behavior for applicants who were denied credit and therefore lack observed repayment outcomes. It notes that parceling is one established approach and points to a bivariate probit model with sample selection. That model allows the relationship between applicant characteristics and default to differ between accepted and rejected populations, rather than assuming the accepted applicants represent everyone.

The response also names several alternative families covered in a referenced overview: fuzzy reclassification, iterative reclassification, and the three-groups approach, alongside multiple forms of parceling. It does not explain how to implement or compare these methods, provide equations beyond the sample-selection premise, or report validation results. The central modeling caveat is that rejected applicants’ outcomes are unobserved, so inference depends on assumptions about how the rejected and accepted populations differ. The document is a brief pointer to approaches, not a recommendation of one universally effective technique.

Key ideas

• Rejected applicants have no observed repayment outcomes, creating a selection problem for credit scoring.
• A bivariate probit model with sample selection can represent different default relationships for accepted and rejected applicants.
• Parceling is one approach, with fuzzy and iterative reclassification as alternatives.
• A three-groups method is another reject-inference approach named in the referenced overview.
• The document provides no comparative evidence establishing which method performs best.

Tags

# Reject inference


# Reject inference












What are the most effective techniques for reject inference in the context of retail credit scoring. Parcelling is something I use frequently... Any other approaches out there?

## Answer by goric (score 2)

https://quant.stackexchange.com/a/456

Have you looked at the bivariate probit model at all?

> Bivariate probit model with sample selection assumes that the distribution of the accepted applicant population is different from that of the rejected applicant population. That is, it is assumed that $P(default|X, rejected) \neq P(default|X, accepted)$ for some vector of explanatory variables X of the model predicting the default of companies.

I addition to that paper, there's an article that highlights different approaches available here: Theoretical approaches of reject inference.

It gives overviews of:

- Several different parceling methods

- Fuzzy reclassification

- Iterative reclassification

- Three-groups approach

among others.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.