Risk Signals Credit Lab

Credit scoring, model validation and portfolio stress testing in a self-contained case study.

This project takes synthetic customer-level credit data and walks through the full analytics chain: SQL exploration, default prediction with multiple models, probability calibration, SHAP explainability, Monte Carlo loss simulation and counterfactual analysis. It's not a production model. The goal is to show how these pieces connect in a realistic workflow.

What's in here

• SQL queries for segment-level portfolio checks (the kind of thing you'd run first in any credit review)
• Logistic regression baseline plus decision tree, random forest and gradient boosting comparisons
• Model validation: AUC, Brier score, calibration curves, risk band checks, threshold sensitivity
• SHAP explainability for global and individual-level feature attribution
• Monte Carlo portfolio loss simulation under baseline, mild and severe stress
• DiCE counterfactual explanations for high-risk customers
• Scenario-based time series view of PD under stress multipliers

Why I built it

I wanted a project that goes past the usual "train a classifier and report accuracy" pattern. In credit risk, the interesting questions start after you have a model. Is it calibrated? Do the risk bands make sense? What happens to portfolio losses under stress? Which feature changes would actually move a customer out of the high-risk band?

I structured it as a small model review case study where the predictions get checked, interpreted and translated into portfolio-level risk measures.

Project structure

data/raw/                       # Synthetic customer data
data/processed/                 # Model outputs, scores, stress results
sql/                            # Portfolio and segment queries
notebooks/                      # Executed analysis notebooks
reports/                        # Written summaries
figures/                        # SVG charts
src/risk_signals_credit_lab.py  # Main reproducible pipeline
tests/                          # Small smoke checks
LICENSE
requirements.txt

How to run

pip install -r requirements.txt
python src/risk_signals_credit_lab.py

This generates the data, runs the models, produces figures and writes the reports. The notebooks can then be run interactively for the sklearn/SHAP/DiCE extensions.

The main pipeline uses only numpy, pandas and the standard library. Notebooks pull in sklearn, shap and dice-ml for the richer analysis.

Tests

The tests are small smoke checks. They cover synthetic data generation, the feature split with a lightweight logistic model, and the validation plus SQL summary outputs.

pytest

Sample outputs

Stress testing results

Scenario	Expected loss	VaR 95	CVaR 95
Baseline	6,668,905	7,185,520	7,325,264
Mild stress	8,153,222	8,747,561	8,890,582
Severe stress	10,147,509	10,807,880	10,991,792

Full simulation output in data/processed/stress_test_summary.csv.

Limitations

• The data is synthetic, so the results shouldn't be read as real credit performance
• Stress multipliers are flat rather than macro-linked
• Risk bands are for interpretation only, not regulatory scorecards
• The Monte Carlo setup shows the workflow but isn't calibrated for capital estimation
• No reject inference, population stability or fairness analysis yet

License

MIT License.