Earlier this week my colleagues presented a webinar on dual score reject inference, which was very well attended and sparked several great questions. The central idea is to inject the historical pattern of reasonable accept and reject decisions, along with observed good and bad repayment behavior, to speed inference and improve model quality. Because the vast majority of these decisions today are guided by predictive scores and deterministic rules, this new approach holds great appeal.
I was happy to see such strong interest in a topic that, while decades old, still generates a healthy discussion on the art and science of its techniques. Reject inference is critical to effective risk modeling for credit decisions, and is a great example of how predictive analytics are intimately bound to the decisions they inform. Lenders obviously need to be cautious and measured when approving loans, but for the best future decisions, their prediction models must rise above the statistical bias inherent in their historical accept and reject choices. Reject inference is the technique used to derive reasonable estimates of consumer performance on the rejected and uncashed loans.
The new dual score inference technique utilizes a combination of binary outcome models — one that estimates accept versus reject probability, and another to estimate successful repayment given a decision to accept — to best project the repayment performance that cannot be directly observed. If this holds interest for you, be sure to check out the recording of the webinar.