| layout | page |
|---|---|
| title | About |
| description | Open-source Python libraries for UK personal lines pricing actuaries. The name comes from a basic actuarial concept — burning cost is the pure loss experience rate, used for experience rating. |
| permalink | /about/ |
Burning Cost is a set of open-source Python libraries for UK personal lines pricing actuaries. The name comes from a basic actuarial concept: burning cost is the pure loss experience rate — actual losses on a risk as a proportion of the subject premium — used for experience rating. We build tools for the problems where Emblem, Radar, and Akur8 stop — causal inference, proxy discrimination auditing, conformal prediction, model governance.
The name is also a philosophy. Simple, direct, no mystification. That is how we think about tooling.
Built by pricing practitioners who have worked across UK personal lines motor and home books.
34 Python libraries covering the full pricing workflow. See the full library index with pip install commands.
UK pricing teams have adopted GBMs (CatBoost is now the dominant choice for most new builds) but many are still taking GLM outputs to production because the GBM outputs are not in a form that rating engines, regulators, or pricing committees can work with. The tools here are about closing that gap — from raw data through to a signed-off rate change with an audit trail. All of it runs on Databricks.
Data & Validation
insurance-cv- temporal walk-forward cross-validation with IBNR buffers and sklearn-compatible scorersinsurance-datasets- synthetic UK motor data with a known data-generating process, for testing and teachinginsurance-synthetic- vine copula synthetic portfolio generation preserving multivariate dependence structureinsurance-conformal- distribution-free prediction intervals for insurance GBMs with finite-sample coverage guaranteesinsurance-monitoring- exposure-weighted PSI/CSI, actual-vs-expected ratios, and Gini drift z-tests for deployed modelsinsurance-governance- structured PRA SS1/23 model validation reports and model risk management, output as HTML and JSON
Model Building
insurance-credibility- Buhlmann-Straub credibility in Python with mixed-model equivalence checksbayesian-pricing- hierarchical Bayesian models for thin-data pricing segments using PyMC 5insurance-spatial- BYM2 spatial models for postcode-level territory ratemaking, borrowing strength from neighboursinsurance-multilevel- CatBoost combined with REML random effects for high-cardinality categorical groupsinsurance-trend- loss cost trend analysis with structural break detection and regime-aware projectionsinsurance-gam- interpretable GAM models: EBM tariffs, actuarial NAM, and pairwise interaction networks with exact Shapley valuesinsurance-interactions- automated GLM interaction detection using CANN, NID, and SHAP-based methodsinsurance-distill- GBM-to-GLM distillation: fits a surrogate GLM to CatBoost predictions and exports multiplicative factor tables for Radar/Emblem rating enginesinsurance-survival- shared frailty models for recurrent claims, cure models, competing risks, and retention modelling
Interpretation
shap-relativities- multiplicative rating factor tables from CatBoost models via SHAP, in the same format as exp(beta) from a GLMinsurance-causal- causal inference via double machine learning for deconfounding rating factors; includes price elasticity estimation (CausalForestDML, DR-Learner) viainsurance_causal.elasticity
Tail Risk & Distributions
insurance-distributional-glm- GAMLSS for Python: model ALL distribution parameters as functions of covariates, seven families, RS algorithminsurance-severity- severity modelling toolkit for UK non-life insurance: full predictive distributions per riskinsurance-quantile- quantile and expectile GBMs for tail risk, TVaR, and increased limit factorsinsurance-distributional- distributional GBMs with Tweedie, Gamma, ZIP, and negative binomial objectives
Commercial
insurance-optimise- constrained rate change optimisation with efficient frontier between loss ratio target and movement cap constraints; includes demand modelling (conversion/retention elasticity, price response curves) viainsurance_optimise.demand
Compliance & Governance
insurance-fairness- proxy discrimination auditing, optimal transport discrimination-free pricing, and FCA Consumer Duty documentation supportinsurance-causal-policy- synthetic difference-in-differences for causal rate change evaluation and FCA evidence packsinsurance-governance- model risk management: ModelCard, ModelInventory, and GovernanceReport generationinsurance-deploy- champion/challenger framework with shadow mode, rollback, and full audit trail
Infrastructure
burning-cost- the Burning Cost CLI; orchestration for pricing model pipelines
UK pricing teams have been building GBMs for years, mostly CatBoost. The models are better than the production GLMs. But many teams are still taking the GLM to production, because the GBM outputs are not in a form that a rating engine, regulator, or pricing committee can work with.
The issue is not technical skill. It is tooling. There is no standard Python library that extracts a multiplicative relativities table from a GBM. There is no standard library that does temporally-correct walk-forward cross-validation with IBNR buffers. There is no standard library that builds a constrained rate optimisation a pricing actuary can challenge. There is no standard library that generates a PRA SS1/23-compliant model validation report.
We wrote those libraries because we needed them. Then we kept going. Everything is built to run on Databricks — that is where UK pricing teams are working, and where our research demonstrates its best practice.
Our benchmarks use synthetic data with known parameters because that is the only way to measure bias — you need ground truth. But the libraries are designed for messy real-world data: fractional exposures from mid-term adjustments, IBNR-contaminated accident years, missing NCD values, vehicle group code changes across ABI revisions, and duplicate records from system migrations. Every API accepts exposure offsets as a first-class parameter. Every model handles missing values through CatBoost's native treatment rather than requiring imputation. If your portfolio does not look like np.random.default_rng(42), that is what these tools are built for.
Start a conversation on GitHub Discussions — that's where we discuss new features, answer questions, and take feedback. For everything else: pricing.frontier@gmail.com.
If you need help getting the libraries into production — adapting examples to your data schema, navigating model risk sign-off, or building a compliant audit trail — see Work with Us.