Financial institutions rely on credit scoring models to assess the risks associated with granting credit. In particular, application scorecards are commonly used as decision support mechanisms for customer acquisition and are developed based on approved applicants. Declined applicants are not included in the modeling exercise, which makes sense because their performance is not known. However, in the same way that some accepted applicants turn out to be bad risks and default, some declined applicants may well turn out to be good risks and profitable. Therefore, it can be argued that a model based solely on accepted applicants is not representative of the “through-the-door” population.
Reject Inference is the method of inferring the performance of those applicants who have been previously declined. Even though declined applicants may be too risky to approve, it can still be beneficial to consider them in model development to reduce sample bias and to be able to develop a model for predicting the behavior of the “through-the-door” population. This is also especially useful when the business objective is to find ways of increasing growth and revenue by identifying good applicants that would have been previously declined.
The idea is to combine application data of declined applicants and their inferred status via Reject Inference techniques such as fuzzy augmentation, parceling, and simple augmentation, with that of accepted applicants with known status. The inferred and known data can then be used to develop a potentially more representative application model.
For further information about Reject Inference capabilities that Angoss software KnowledgeSTUDIO offers, click here.