Credit scoring AI retraining creates illusion of improvement, ECML PKDD paper finds

In a paper accepted to ECML PKDD 2026 (Research Track), Bruno Scarone and Ricardo Baeza-Yates systematically evaluate reject inference methods designed to mitigate survival bias in credit scoring. They uncover a structural failure mode: during natural retraining cycles, models can show improving accuracy while their recall collapses, creating an illusion of improvement. Practitioners believe the system is getting better, but in reality, the model's ability to correctly screen out defaulters (rejection quality) is deteriorating. This happens because standard evaluation metrics like accuracy are misleading under selection bias, as they fail to account for the unobserved outcomes of rejected applicants.

To break this feedback loop, the authors propose a controlled exploration strategy: the lender deliberately approves a small fraction (2-5%) of applicants who would normally be rejected and observes their true repayment behavior. This approach requires no statistical assumptions and is shown to be sufficient even at minimal exploration rates. Their experiments, conducted across two machine learning methods (likely gradient boosting and logistic regression) and three real-world credit datasets, confirm that accuracy and rejection quality give opposite recommendations on whether to explore. Accuracy favors no exploration, while rejection quality improves significantly with exploration. The findings suggest that standard evaluation protocols are inadequate for models trained under survival bias, and lenders must adopt exploration-based diagnostics to avoid worsening default rates.