Balancing Performance and Inclusion: A Novel Reject Inference Framework for Credit Scoring

Reject Inference (RI) is a critical process for mitigating sample bias in credit scoring by incorporating data from rejected applicants. To address this, we propose the Confident Inlier Extrapolation (CI-EX) framework, a semi-supervised method that iteratively labels the rejected pool by combining outlier detection with confidence-based selection. Our experiments on a large-scale real-world dataset, evaluated using AUC and the novel rea Under the Kickout (AUK) metric, reveal a fundamental trade-off between standard predictive accuracy and RI-specific performance. The results demonstrate that while CI-EX maintains competitive AUC, it consistently outperforms established literature benchmarks in AUK and Kickout metrics. This Pareto-optimal performance positions CI-EX as a robust solution for enhancing both the accuracy and the inclusive nature of credit decision-making systems.