Problem
A credit scoring model is trained only on loans that the previous model approved. In microlending the approval rate is 10–30%, so most applications (about 80% in the talk) have no outcome label. The approved sample is biased: a feature that was strong on all applications can look weak on the approved ones. A new model trained on this sample may then rank applicants worse than expected once it scores all of them.
Example from the article: at the MFO DOZARPLATY, a payday lender, the strongest feature of the old model lost almost all its information value after training on issued loans, while previously weak features moved to the top.
Solution in brief
The talk reviews ten reject inference methods and says which five the team applied. The preferred one corrects the new model with the total probability formula. It combines a default model on approved loans, an approval model and a constant default rate of rejects taken from credit bureau data.
Questions this case answers
- How to correct selection bias in a scoring model trained only on accepted applicants when rejected applicants have no observed outcome
- How to estimate the default probability of the whole applicant flow by combining an approved-branch model with a constant rate for rejects using the total probability formula
- How to correct selection bias in a credit scoring model trained only on approved applicants
- How to choose a reject inference method when the approval rate is low
Tools and methods
- Loans issued beyond the cut-off — approve some applications below the cut-off to get an unbiased validation sample.
- Credit bureau analysis — check what rejected applicants did at other lenders after about 3 months.
- Assign all rejects as bad — train a model that replicates underwriter decisions.
- Bayesian correction — combine , and a bureau-based .
- Parceling, fuzzy augmentation, hard and soft cut-off — pseudo-labeling of rejects, not applied.
Google or a one-time email code. No password.