Explainable credit risk intelligence
An end-to-end data science system that turns Home Credit data into calibrated risk evidence—then makes every prediction traceable, governed and ready for human review.
The portfolio case
A credible lending model is more than an algorithm. RiskLens documents the data, split strategy, calibration, operating threshold, subgroup behavior, explanations and post-deployment drift.
Research decision support only. It does not autonomously approve or decline loans.
Frozen portfolio evidence
● model locked| Dimension | Measure | Result | Evidence |
|---|---|---|---|
| Discrimination | ROC-AUC | 0.7840 | One-time holdout |
| Minority-class quality | PR-AUC | 0.2732 | One-time holdout |
| Probability accuracy | Brier score | 0.0663 | Calibrated |
| Decision policy | Recall / precision | 42.57% / 27.28% | Threshold 0.1667 |
Evaluated once on 30,752 untouched applicants. Post-holdout tuning is prohibited.
Two assessment modes
RiskLens never pretends that a brand-new application has the same evidence as an applicant already represented in historical records.
Existing applicant
Scores an applicant from the Kaggle population using application data plus bureau, prior application, installment, credit-card and POS history.
New application
Accepts only information available at application time and uses a separately evaluated application-only model. It estimates risk; it does not grant a loan.
Reproducible workflow
Each stage is executable from the CLI, tested, versioned and represented by saved artifacts.
Contract checks and relational coverage across Home Credit tables.
Application attributes plus 196 leakage-safe historical aggregates.
Benchmarks, XGBoost selection, calibration and cost-sensitive policy.
SHAP, subgroup diagnostics, model card and permanently sealed holdout.
FastAPI, Streamlit, Docker, monitoring and evidence-grounded assistant.
System architecture
Raw Kaggle tables feed validated feature pipelines. Frozen artifacts power the authenticated API, while the dashboard presents decisions, evidence, monitoring and retrieval-grounded policy answers.
Responsible AI
The design surfaces limitations instead of hiding them behind a single performance number.
Global and local SHAP analysis with business-readable labels. Explanations describe model behavior—not causality.
Recall, false-positive rate, selection rate and calibration are reported by gender and age band for audit only.
Frozen references track prediction and feature PSI. Alerts trigger investigation, never silent post-holdout tuning.
Retrieval-grounded answers cite approved project evidence. Retrieval reached 100% hit@3 on its evaluation set.
Known limitation
Subgroup gaps are diagnostic, not proof of fairness or legal compliance. Any real deployment would require lender-specific data, validation, policy review and ongoing human oversight.
Review limitations ↗Production-minded engineering