Credit Risk Scoring — Business Decision Thresholds
This project demonstrates how a machine learning model estimating probability of default (P(bad)) can be integrated into an actionable credit decision system (ACCEPT / REVIEW / REJECT).
Why this project matters
This project illustrates how machine learning outputs can be integrated into a real-world decision process. Beyond predictive performance, it focuses on decision thresholds, explainability, and operational constraints commonly found in regulated environments such as credit risk and financial services.
Dataset & Approach
- Dataset : credit-g (OpenML)
- Méthodologie : CRISP-DM
- Baseline : Régression logistique (interprétable, stable)
Decision Logic
Two thresholds : A (accept) and R (reject)
- ACCEPT if P(bad) < A
- REVIEW if A ≤ P(bad) < R
- REJECT if P(bad) ≥ R
Principe : the model estimates the risk, the business keeps the decision.
What’s inside
- End-to-end pipeline from data preprocessing to deployed decision service
- Interpretable baseline model (logistic regression) selected for stability and auditability
- Business decision layer translating risk scores into ACCEPT / REVIEW / REJECT outcomes
- Web API and decision-oriented UI
- Containerized deployment on Google Cloud Run with CI/CD (GitHub Actions, WIF)
Technical Stack
Python · scikit-learn · Flask API · Docker · Google Cloud Run · CI/CD
Intended audience
This project is intended for technical and non-technical stakeholders interested in how machine learning systems can support controlled, explainable business decisions rather than fully automated outcomes.