Note
Credit-scoring models with XGBoost, SHAP and PSI stability checks
Two credit-scoring models developed and documented for the risk team: a payroll-data scoring model with decision thresholds and an early-delinquency model, each with SHAP explanations and PSI stability checks.
| Organisation | Banco de Bogotá |
|---|---|
| Sector | Banking, retail credit risk |
| Period | 10/2021 – 04/2022 · Bogotá |
| Role | Credit Risk Analyst (ML credit scoring) |
| Stack | Python · XGBoost · SHAP · scikit-learn |
Method
- Models. A scoring model on payroll data, with decision thresholds, and an early-delinquency model.
- Learner. Gradient-boosted trees (XGBoost).
- Explainability. SHAP values per feature and per decision, so the risk team could read why a score moved.
- Stability. Population stability index (PSI) between development and later windows, to catch drift in the inputs before it reached the score.
- Hand-over. Both models documented for the risk team. Developed, not deployed: I cannot verify what reached production after I left.
Confidentiality
Method only. No data, features or thresholds.