Insurance Pricing Pipeline Demo

A single notebook that runs five libraries end to end on the same synthetic UK motor dataset: fit a CatBoost Poisson frequency model, extract actuarial factor tables via SHAP, audit the model for proxy discrimination under FCA Consumer Duty, monitor it for drift, and attach distribution-free conformal prediction intervals. The whole workflow runs in under three minutes on a free Colab GPU.

The dataset has a deliberate design: gender is correlated with region (urban areas skew female in the synthetic population) but has no causal path to claims. The fairness audit should detect this indirect channel — which is exactly the proxy discrimination pattern the FCA flags in Consumer Duty assessments — even though gender was never a model input.

Open in Colab

Click the badge above. The notebook installs its own dependencies via pip and runs top to bottom with no additional setup.

Libraries used

Step	Library	What it does
Frequency model	CatBoost	Poisson GBM with exposure offset
Factor tables	shap-relativities	Multiplicative rating relativities from SHAP values
Fairness audit	insurance-fairness	Proxy discrimination audit, D_proxy metric, Shapley attribution
Drift monitoring	insurance-monitoring	PSI/CSI per feature, A/E ratios, Murphy decomposition
Prediction intervals	insurance-conformal	Conformal intervals with finite-sample coverage guarantees

All five libraries are published on PyPI and available at github.com/burning-cost.

Notebook structure

1. Synthetic UK motor data — 5,000 policies, known Poisson DGP
2. CatBoost frequency model — standard rating factors, exposure as weight
3. SHAP rating relativities — factor tables in GLM format with confidence intervals
4. Proxy discrimination audit — detects indirect gender effect via region correlation
5. Model monitoring — PSI per feature, A/E with Poisson CIs, Murphy decomposition
6. Conformal prediction intervals — 90% coverage verified on held-out test set

Running locally

pip install catboost "shap-relativities[ml]" insurance-fairness insurance-monitoring insurance-conformal polars shap scikit-learn statsmodels
jupyter notebook insurance-pricing-pipeline.ipynb

Licence

MIT