credit-risk-pd-policy-simulation

A GitHub-ready Basel/IRB-style credit risk policy simulation project.

This repository scores applicants using:

• Proxy PD (prob_def) (because the included sample dataset has no observed default outcome)
• Policy cutoff -> loan_approval (1=approve, 0=decline)
• EAD/LGD/EL proxies for expected loss reporting

If you have a real dataset with an outcome (e.g., default_flag), use the supervised template notebook and extend the module to train/validate a true PD model.

Why “Proxy PD”? (Important)

The bundled dataset data/Bank_Loan_Approval.csv contains applicant features but no historical performance outcome (no default_flag, bad_flag, Personal.Loan, etc.). Without an outcome, you cannot do dev/val/test splits, AUC/KS, calibration, or backtesting. This repo therefore provides an explainable proxy PD calibrated to a target mean PD (default 5%) and a realistic decisioning layer.

Project Structure

credit-risk-pd-policy-sim/
├─ data/
├─ notebooks/
├─ src/irb_simulator/
├─ scripts/
├─ outputs/
├─ tests/
├─ LICENSE
├─ .gitignore
├─ requirements.txt
└─ pyproject.toml

Quickstart

Install

python -m venv .venv
# Windows: .venv\Scripts\activate
# macOS/Linux: source .venv/bin/activate
pip install -r requirements.txt

Score the dataset

python scripts/score_dataset.py --infile data/Bank_Loan_Approval.csv --outfile outputs/scored.csv --pd-cutoff 0.05

Make a report

python scripts/make_report.py --scored outputs/scored.csv --out outputs/report.md

IRB components implemented (proxies)

• PD: proxy PD (scorecard-style, calibrated mean)
• EAD: 30% of annual income (approved-only exposure)
• LGD: 60% base; 40% if CD/Securities; small mortgage uplift (bounded)
• EL: PD × EAD × LGD

Upgrade path (true IRB PD model)

Add an outcome column (e.g., default_flag) and run notebooks/02_supervised_pd_template.ipynb to do:

• dev/val/test split
• AUC/ROC
• calibration/backtesting (extend as needed)

License

MIT License (see LICENSE).