Balisa50/credit-risk-scorecard

Basel II-compliant credit risk scorecard for West African microfinance. WoE/IV, logistic regression scorecard, Gini/KS/PSI validation, stress testing. Python + Next.js. # Credit Risk Scorecard Basel II-compliant credit scorecard for West African microfinance. Logistic regression with WoE encoding, built on 12,000 synthetic loans calibrated to microfinance risk profiles in the region. Why build this: most credit risk tooling assumes the data and risk distribution of mature Western markets. West African microfinance has different risk drivers - mobile money usage, agricultural income seasonality, informal employment - so the feature importance looks different. ## Pipeline 1. **Data cleaning** - handle informal income fields, missing collateral data, outlier capping at 1st/99th percentile 2. **WoE/IV feature selection** - bins continuous variables, calculates Information Value per feature. Drops anything below IV 0.02 3. **Logistic regression** - fit on WoE-transformed features, convert coefficients to Basel II integer scorecard points (PDO = 20, base score = 600) 4. **Validation** - Gini 0.29, KS 0.23 on the time-based holdout. PSI 0.008 across validation windows (stable) 5. **Stress testing** - shift default rate +50% (economic stress), +100% (severe), shift feature distributions, re-score the book ## Stack - Python - pandas, scikit-learn, scipy, matplotlib - Next.js + Recharts - scorecard UI and stress test visualiser ## Running ```bash pip install -r requirements.txt jupyter notebook notebooks/ # Dashboard cd dashboard npm install && npm run dev ``` ## Results All figures below are the test-set values written by `pipeline/run_pipeline.py` into `public/data/pipeline_results.json`, which is also what the dashboard renders. | Metric | Train | Test | Industry threshold | |--------|-------|------|--------------------| | Gini coefficient | 0.35 | **0.29** | > 0.4 | | KS statistic | 0.27 | **0.23** | > 0.3 | | AUC-ROC | 0.68 | **0.65** | - | | PSI (population stability) | - | **0.008** | < 0.1 | Discrimination falls short of the usual thresholds, and that is a property of the synthetic data rather than the modelling. No feat …