Interpretable hybrid credit scoring for thin-file and underbanked populations

RESUMO

We extend a residual-learning hybrid credit scoring framework (logistic regression scorecard plus gradient-boosting correction on its residuals, decomposed at each prediction into an interpretability ratio ρ(x) that measures the share attributable to the linear branch) along three axes: an East African empirical instantiation on Zindi Financial Inclusion in Africa data (Kenya, Rwanda, Tanzania, Uganda); a fairness audit granularity of the framework's three regions; and a thin-file segmentation analysis. On the Taiwan Credit Default benchmark retained for continuity, the calibrated hybrid attains AUC = 0.776 (ΔAUC = +0.057 vs. standalone logistic regression, +0.001 vs. XGBoost), reduces Brier Score by 23%, and concentrates the highest-default-rate borrowers (69.5%) in fully interpretable region. On Zindi, calibrated hybrid attains AUC 0.869 (ΔAUC +0.015 vs. LR, p < 0.001; -0.004 vs. XGBoost), cuts Brier Score from 0.158 to 0.085 (a 46% reduction), and replicates regional routing pattern. The fairness audit detects severe opaque ML-driven regional routing along three socioeconomic axes: rural respondents by 18 percentage points relative urban, primary-or-less-educated by 32 points relative secondary-and-above, and Ugandan by 22 points relative Kenyan, while gender shows essentially no routing disparity. The pipeline surfaces subgroup-routing violations that aggregate fairness metrics miss, in form directly usable by central-bank supervisors for digital credit.