Interpretable hybrid credit scoring for thin-file and underbanked populations: A fairness-aware residual learning framework with evidence from East African financial inclusion data

Battle Count: 69

Primarily focused on consumer credit scoring and financial inclusion rather than trading. Relevant to quantitative finance through credit risk modeling, regulatory compliance for digital lenders, and algorithmic fairness in financial decision-making. Limited direct applicability to trading strategies, portfolio optimization, or market prediction.

Moderate complexity: two-stage training (LR + XGBoost on residuals), adaptive weighting, Platt calibration, WoE encoding, fairness audit pipeline with bootstrap inference, TreeSHAP explanations. ~800 lines of Python code, runtime under 5 minutes on laptop CPU. Uses standard scikit-learn, XGBoost, and Python scientific libraries. Fairness audit module adds complexity with stratified bootstrap and DeLong test.

Code to be released publicly under permissive open-source license (~800 lines, end-to-end runtime under 5 minutes on laptop CPU). Both datasets (Zindi FSD and Taiwan Credit Default) are publicly available. Fixed random seeds specified. Hyperparameter grids documented. Bootstrap protocols (B=500 for fairness, B=2000 for AUC) clearly stated. Minor gap: exact repository URL not yet provided (camera-ready stage).