34 open-source Python libraries covering the full pricing stack. Ten of them we consider genuinely differentiated — tools that address hard problems in UK pricing where no adequate Python solution existed before. The full portfolio is below the flagship section.
All libraries are MIT-licensed, installable via pip, and built for Python 3.10+. The GitHub organisation is burning-cost.
Flagship Libraries
Ten libraries for the problems that matter most. Each addresses a specific hard problem — regulatory compliance, causal estimation, uncertainty quantification, smoothing, telematics scoring — where practitioners were previously doing things by hand or not doing them at all.
insurance-fairness [](https://github.com/burning-cost/insurance-fairness/actions/workflows/ci.yml)
Proxy discrimination auditing aligned to FCA Consumer Duty. Quantifies indirect discrimination risk from rating variables correlated with protected characteristics. Equality Act 2010 proportionality documentation built in.
FCA Consumer Duty
PRA SS1/23-compliant model validation reports. Bootstrap Gini CI, Poisson A/E CI, double-lift charts, renewal cohort test. HTML/JSON output structured for model risk committees and PRA review.
PRA SS1/23 · MRM
insurance-causal [](https://github.com/burning-cost/insurance-causal/actions/workflows/ci.yml)
Double machine learning for deconfounding rating factors. Standard GLM coefficients are biased wherever rating variables correlate with distribution channel or policyholder selection. DML removes that bias without a structural model. v0.4.0 adds causal forest heterogeneous treatment effects: GATES aggregates, CLAN segment profiling, and RATE/AUTOC/QINI targeting evaluation for segment-level price response.
Causal inference · DML
Whittaker-Henderson smoothing for experience rating tables. 1D, 2D, and Poisson variants with REML lambda selection and Bayesian credible intervals. Smooths age curves, NCD scales, and vehicle group relativities without parametric assumptions.
Non-parametric smoothing · REML
End-to-end pipeline from raw 1Hz GPS/accelerometer data to GLM-compatible risk scores. HMM driving state classification (cautious/normal/aggressive), Bühlmann-Straub credibility aggregation to driver level, Poisson GLM integration.
Telematics · HMM · UBI
insurance-conformal [](https://github.com/burning-cost/insurance-conformal/actions/workflows/ci.yml)
Distribution-free prediction intervals with finite-sample coverage guarantees. Five non-conformity scores tuned for Tweedie and Poisson claims — the default
pearson_weighted score produces 13.4% narrower intervals than parametric Tweedie at identical 90% coverage. Frequency-severity conformal intervals, online retrospective adjustment, Solvency II SCR bounds.Conformal prediction · Solvency II
insurance-credibility [](https://github.com/burning-cost/insurance-credibility/actions/workflows/ci.yml)
Bühlmann-Straub credibility in Python with mixed-model equivalence checks. Caps thin segments, stabilises NCD factors, blends a new model with an incumbent rate. The actuarial answer to the thin data problem.
Bühlmann-Straub · Thin data
Sarmanov copula joint frequency-severity modelling with GLM marginals. IFM estimation, analytical premium correction, Garrido conditional severity, dependence tests. Tests the independence assumption every pricing model makes.
Sarmanov copula · Joint modelling
EBM and Neural Additive Model for interpretable deep learning in insurance pricing. Shape functions per rating factor give the transparency of a GLM with the predictive power of a neural network. Factor table output for pricing committee review.
EBM · NAM · Interpretable ML
Exposure-weighted drift detection for deployed pricing models. PSI, CSI, actual-vs-expected ratios, Gini drift z-test. v0.8.0 adds SequentialTest for anytime-valid A/B testing of champion/challenger experiments via mSPRT, PITMonitor for calibration drift, and InterpretableDriftDetector for feature-interaction drift attribution.
Model monitoring · Drift detection
Search and Filter
Model Building
Tools for building and interpreting pricing models.
| Library | What it does | Install |
|---|---|---|
| [shap-relativities](https://github.com/burning-cost/shap-relativities) [](https://github.com/burning-cost/shap-relativities/actions/workflows/ci.yml) | SHAP-based rating relativities from GBM models — extract GLM-style multiplicative factors from CatBoost
_Benchmark: +2.85pp Gini lift over direct GLM; NCD relativity error 4.47% vs GLM's 9.44%_ | `pip install shap-relativities` | | [insurance-glm-tools](https://github.com/burning-cost/insurance-glm-tools) | GLM tooling — nested GLM embeddings, R2VF factor level clustering, territory banding, SKATER. Torch is optional: core clustering and banding works without it
_Benchmark: R2VF fused lasso clustering vs manual quintile banding benchmarked_ | Core: `pip install insurance-glm-tools`
Neural: `pip install insurance-glm-tools[neural]` | | [insurance-gam](https://github.com/burning-cost/insurance-gam) | **Flagship.** Interpretable GAMs — EBM tariffs, Actuarial NAM, Pairwise Interaction Networks, exact Shapley values
_Benchmark: EBM/NAM interpretable tariffs with exact Shapley values benchmarked on synthetic motor data_ | `pip install insurance-gam` | | [insurance-interactions](https://github.com/burning-cost/insurance-interactions) | Automated GLM interaction detection using CANN, NID scoring, and SHAP interaction values
_Benchmark: Production defaults recover both planted interactions; compact config less reliable_ | `pip install insurance-interactions` | | [insurance-frequency-severity](https://github.com/burning-cost/insurance-frequency-severity) | **Flagship.** Sarmanov copula joint frequency-severity — analytical premium correction, IFM estimation. Torch is optional: Sarmanov copulas and IFM estimation work without it
_Benchmark: Sarmanov copula joint modelling with analytical premium correction benchmarked_ | Core: `pip install insurance-frequency-severity`
Neural: `pip install insurance-frequency-severity[neural]` | | [insurance-spatial](https://github.com/burning-cost/insurance-spatial) | BYM2 spatial territory ratemaking — PyMC 5 ICAR, adjacency matrices, Moran's I diagnostics
_Benchmark: BYM2 vs raw vs manual banding benchmarked on synthetic territory data_ | `pip install insurance-spatial` | | [insurance-distill](https://github.com/burning-cost/insurance-distill) | GBM-to-GLM distillation — fits a surrogate Poisson/Gamma GLM to CatBoost predictions, exports multiplicative factor tables for Radar/Emblem
_Benchmark: 90-97% R² match between GBM predictions and distilled GLM factors_ | `pip install insurance-distill` |
_Benchmark: +2.85pp Gini lift over direct GLM; NCD relativity error 4.47% vs GLM's 9.44%_ | `pip install shap-relativities` | | [insurance-glm-tools](https://github.com/burning-cost/insurance-glm-tools) | GLM tooling — nested GLM embeddings, R2VF factor level clustering, territory banding, SKATER. Torch is optional: core clustering and banding works without it
_Benchmark: R2VF fused lasso clustering vs manual quintile banding benchmarked_ | Core: `pip install insurance-glm-tools`
Neural: `pip install insurance-glm-tools[neural]` | | [insurance-gam](https://github.com/burning-cost/insurance-gam) | **Flagship.** Interpretable GAMs — EBM tariffs, Actuarial NAM, Pairwise Interaction Networks, exact Shapley values
_Benchmark: EBM/NAM interpretable tariffs with exact Shapley values benchmarked on synthetic motor data_ | `pip install insurance-gam` | | [insurance-interactions](https://github.com/burning-cost/insurance-interactions) | Automated GLM interaction detection using CANN, NID scoring, and SHAP interaction values
_Benchmark: Production defaults recover both planted interactions; compact config less reliable_ | `pip install insurance-interactions` | | [insurance-frequency-severity](https://github.com/burning-cost/insurance-frequency-severity) | **Flagship.** Sarmanov copula joint frequency-severity — analytical premium correction, IFM estimation. Torch is optional: Sarmanov copulas and IFM estimation work without it
_Benchmark: Sarmanov copula joint modelling with analytical premium correction benchmarked_ | Core: `pip install insurance-frequency-severity`
Neural: `pip install insurance-frequency-severity[neural]` | | [insurance-spatial](https://github.com/burning-cost/insurance-spatial) | BYM2 spatial territory ratemaking — PyMC 5 ICAR, adjacency matrices, Moran's I diagnostics
_Benchmark: BYM2 vs raw vs manual banding benchmarked on synthetic territory data_ | `pip install insurance-spatial` | | [insurance-distill](https://github.com/burning-cost/insurance-distill) | GBM-to-GLM distillation — fits a surrogate Poisson/Gamma GLM to CatBoost predictions, exports multiplicative factor tables for Radar/Emblem
_Benchmark: 90-97% R² match between GBM predictions and distilled GLM factors_ | `pip install insurance-distill` |
Distributional & Tail Risk
Beyond point estimates: full distribution modelling and tail risk quantification.
| Library | What it does | Install |
|---|---|---|
| [insurance-distributional](https://github.com/burning-cost/insurance-distributional) | Distributional GBMs — Tweedie, Gamma, ZIP, NegBin objectives with per-risk volatility scoring
_Benchmark: GammaGBM +1.5% log-likelihood; prediction intervals calibrated_ | `pip install insurance-distributional` | | [insurance-distributional-glm](https://github.com/burning-cost/insurance-distributional-glm) | GAMLSS for Python — model all distribution parameters as functions of covariates, seven families
_Benchmark: GAMLSS sigma correlation 0.998 vs 0.000 for constant-phi_ | `pip install insurance-distributional-glm` | | [insurance-dispersion](https://github.com/burning-cost/insurance-dispersion) | Double GLM for joint mean-dispersion modelling — alternating IRLS, REML, actuarial factor tables
_Benchmark: Double GLM captures heteroscedasticity that constant-phi misses_ | `pip install insurance-dispersion` | | [insurance-quantile](https://github.com/burning-cost/insurance-quantile) | Quantile and expectile GBMs for tail risk, TVaR, and increased limit factors
_Benchmark: GBM lower TVaR bias on heavy tails; lognormal wins on pinball at small n_ | `pip install insurance-quantile` | | [insurance-severity](https://github.com/burning-cost/insurance-severity) | Spliced Pareto/Gamma severity, Deep Regression Networks, composite Lognormal-GPD, and EQRN extreme quantile neural networks for large loss modelling
_Benchmark: Composite reduces tail error 5.6% vs single lognormal; heavy-tail benchmark (Pareto α=1.5) shows 15–20% tail error reduction_ | `pip install insurance-severity` |
_Benchmark: GammaGBM +1.5% log-likelihood; prediction intervals calibrated_ | `pip install insurance-distributional` | | [insurance-distributional-glm](https://github.com/burning-cost/insurance-distributional-glm) | GAMLSS for Python — model all distribution parameters as functions of covariates, seven families
_Benchmark: GAMLSS sigma correlation 0.998 vs 0.000 for constant-phi_ | `pip install insurance-distributional-glm` | | [insurance-dispersion](https://github.com/burning-cost/insurance-dispersion) | Double GLM for joint mean-dispersion modelling — alternating IRLS, REML, actuarial factor tables
_Benchmark: Double GLM captures heteroscedasticity that constant-phi misses_ | `pip install insurance-dispersion` | | [insurance-quantile](https://github.com/burning-cost/insurance-quantile) | Quantile and expectile GBMs for tail risk, TVaR, and increased limit factors
_Benchmark: GBM lower TVaR bias on heavy tails; lognormal wins on pinball at small n_ | `pip install insurance-quantile` | | [insurance-severity](https://github.com/burning-cost/insurance-severity) | Spliced Pareto/Gamma severity, Deep Regression Networks, composite Lognormal-GPD, and EQRN extreme quantile neural networks for large loss modelling
_Benchmark: Composite reduces tail error 5.6% vs single lognormal; heavy-tail benchmark (Pareto α=1.5) shows 15–20% tail error reduction_ | `pip install insurance-severity` |
Credibility & Thin Data
When you don’t have enough data to trust a standard GLM.
| Library | What it does | Install |
|---|---|---|
| [insurance-credibility](https://github.com/burning-cost/insurance-credibility) [](https://github.com/burning-cost/insurance-credibility/actions/workflows/ci.yml) | **Flagship.** Credibility models for UK non-life pricing: Bühlmann-Straub, Bayesian experience rating, and individual experience rating (static, dynamic, surrogate, and deep attention credibility via `insurance_credibility.experience`)
_Benchmark: 6.8% MAE improvement on thin schemes vs raw experience_ | `pip install insurance-credibility` | | [insurance-multilevel](https://github.com/burning-cost/insurance-multilevel) | Two-stage CatBoost + REML random effects for high-cardinality group factors — ICC diagnostics
_Benchmark: 15.9% gamma deviance reduction; thin-group MAPE 63.6% vs one-hot 66.1%_ | `pip install insurance-multilevel` | | [bayesian-pricing](https://github.com/burning-cost/bayesian-pricing) | Hierarchical Bayesian models for thin-data pricing segments using PyMC 5 | `pip install bayesian-pricing` | | [insurance-thin-data](https://github.com/burning-cost/insurance-thin-data) | Pricing techniques for low-volume segments where standard GLM fitting is unreliable
_Benchmark: GLMTransfer narrows bootstrap 90% CI widths 30–60% vs standalone GLM_ | `pip install insurance-thin-data` | | [insurance-whittaker](https://github.com/burning-cost/insurance-whittaker) | **Flagship.** Whittaker-Henderson 1D/2D smoothing with REML lambda selection, Bayesian CIs, Poisson PIRLS
_Benchmark: 57.2% MSE reduction vs raw rates_ | `pip install insurance-whittaker` |
_Benchmark: 6.8% MAE improvement on thin schemes vs raw experience_ | `pip install insurance-credibility` | | [insurance-multilevel](https://github.com/burning-cost/insurance-multilevel) | Two-stage CatBoost + REML random effects for high-cardinality group factors — ICC diagnostics
_Benchmark: 15.9% gamma deviance reduction; thin-group MAPE 63.6% vs one-hot 66.1%_ | `pip install insurance-multilevel` | | [bayesian-pricing](https://github.com/burning-cost/bayesian-pricing) | Hierarchical Bayesian models for thin-data pricing segments using PyMC 5 | `pip install bayesian-pricing` | | [insurance-thin-data](https://github.com/burning-cost/insurance-thin-data) | Pricing techniques for low-volume segments where standard GLM fitting is unreliable
_Benchmark: GLMTransfer narrows bootstrap 90% CI widths 30–60% vs standalone GLM_ | `pip install insurance-thin-data` | | [insurance-whittaker](https://github.com/burning-cost/insurance-whittaker) | **Flagship.** Whittaker-Henderson 1D/2D smoothing with REML lambda selection, Bayesian CIs, Poisson PIRLS
_Benchmark: 57.2% MSE reduction vs raw rates_ | `pip install insurance-whittaker` |
Causal Inference
Separating what causes what from what correlates with what.
| Library | What it does | Install |
|---|---|---|
| [insurance-causal](https://github.com/burning-cost/insurance-causal) [](https://github.com/burning-cost/insurance-causal/actions/workflows/ci.yml) | **Flagship.** Double machine learning for deconfounding rating factors — CatBoost nuisance models, confounding bias reports. Includes price elasticity estimation and causal forest heterogeneous treatment effects via `insurance_causal.causal_forest` (GATES/BLP/CLAN inference, RATE/AUTOC/QINI targeting evaluation, HTE diagnostics). Use DML for portfolio-level average effects; causal forest for segment-level CATEs with n≥2,000 per group.
_Benchmark: DML removes nonlinear confounding bias at scale (n≥50k); honest: over-partials at small n_ | `pip install insurance-causal` | | [insurance-causal-policy](https://github.com/burning-cost/insurance-causal-policy) | Synthetic difference-in-differences for causal rate change evaluation — event study, HonestDiD sensitivity. v0.2.0 adds DoublyRobustSCEstimator (DRSC, arXiv:2503.11375): doubly robust — consistent if either parallel trends or SC weights hold
_Benchmark: SDID 98% CI coverage; naive before-after biased +3.8pp by market inflation. DRSC: 24% lower RMSE than SDID with few donors (N_co=6), equivalent with many (N_co=40)_ | `pip install insurance-causal-policy` |
_Benchmark: DML removes nonlinear confounding bias at scale (n≥50k); honest: over-partials at small n_ | `pip install insurance-causal` | | [insurance-causal-policy](https://github.com/burning-cost/insurance-causal-policy) | Synthetic difference-in-differences for causal rate change evaluation — event study, HonestDiD sensitivity. v0.2.0 adds DoublyRobustSCEstimator (DRSC, arXiv:2503.11375): doubly robust — consistent if either parallel trends or SC weights hold
_Benchmark: SDID 98% CI coverage; naive before-after biased +3.8pp by market inflation. DRSC: 24% lower RMSE than SDID with few donors (N_co=6), equivalent with many (N_co=40)_ | `pip install insurance-causal-policy` |
Fairness & Regulation
Proxy discrimination, Consumer Duty, and FCA evidence packs.
| Library | What it does | Install |
|---|---|---|
| [insurance-fairness](https://github.com/burning-cost/insurance-fairness) [](https://github.com/burning-cost/insurance-fairness/actions/workflows/ci.yml) | **Flagship.** Proxy discrimination auditing — FCA Consumer Duty, Equality Act 2010, bias metrics
_Benchmark: Proxy R²=0.78 catches postcode proxy that Spearman (r=0.06) misses entirely_ | `pip install insurance-fairness` | | [insurance-covariate-shift](https://github.com/burning-cost/insurance-covariate-shift) | Density ratio correction for book shifts — CatBoost/RuLSIF/KLIEP, LR-QR conformal, FCA SUP 15.3 diagnostics
_Benchmark: Density ratio correction for portfolio composition shift benchmarked_ | `pip install insurance-covariate-shift` |
_Benchmark: Proxy R²=0.78 catches postcode proxy that Spearman (r=0.06) misses entirely_ | `pip install insurance-fairness` | | [insurance-covariate-shift](https://github.com/burning-cost/insurance-covariate-shift) | Density ratio correction for book shifts — CatBoost/RuLSIF/KLIEP, LR-QR conformal, FCA SUP 15.3 diagnostics
_Benchmark: Density ratio correction for portfolio composition shift benchmarked_ | `pip install insurance-covariate-shift` |
Validation & Monitoring
Prediction intervals, conformal methods, and post-deployment monitoring.
| Library | What it does | Install |
|---|---|---|
| [insurance-cv](https://github.com/burning-cost/insurance-cv) [](https://github.com/burning-cost/insurance-cv/actions/workflows/ci.yml) | Temporal walk-forward cross-validation — respects policy time structure, IBNR buffers, sklearn-compatible scorers
_Benchmark: Walk-forward catches 10.5% optimism that k-fold hides_ | `pip install insurance-cv` | | [insurance-conformal](https://github.com/burning-cost/insurance-conformal) [](https://github.com/burning-cost/insurance-conformal/actions/workflows/ci.yml) | **Flagship.** Distribution-free prediction intervals for insurance GBMs and GLMs — 5 non-conformity scores tuned for Tweedie/Poisson claims (`pearson_weighted` default, `pearson`, `deviance`, `anscombe`, `raw`), frequency-severity conformal intervals (Graziadei et al.), online retrospective adjustment (RetroAdj), Solvency II SCR bounds. v0.6.0.
_Benchmark: pearson_weighted score 13.4% narrower intervals than parametric Tweedie at identical 90% coverage; 50k synthetic UK motor policies_ | `pip install insurance-conformal` | | [insurance-conformal-ts](https://github.com/burning-cost/insurance-conformal-ts) | Conformal prediction for non-exchangeable claims time series — ACI, EnbPI, SPCI, Poisson/NB scores
_Benchmark: Sequential conformal vs static prediction intervals on claims time series benchmarked_ | `pip install insurance-conformal-ts` | | [insurance-deploy](https://github.com/burning-cost/insurance-deploy) [](https://github.com/burning-cost/insurance-deploy/actions/workflows/ci.yml) | Champion/challenger pricing framework — shadow mode, SHA-256 routing, SQLite quote log, bootstrap LR test
_Benchmark: Champion/challenger routing, shadow mode, quote logging benchmarked_ | `pip install insurance-deploy` | | [insurance-monitoring](https://github.com/burning-cost/insurance-monitoring) [](https://github.com/burning-cost/insurance-monitoring/actions/workflows/ci.yml) | **Flagship.** Model drift detection — exposure-weighted PSI/CSI, actual-vs-expected ratios, Gini drift z-test. `SequentialTest` for anytime-valid champion/challenger A/B testing (mSPRT, Johari et al. 2022). `PITMonitor` for calibration drift via PIT e-process martingale (~3% FPR vs repeated Hosmer-Lemeshow's 46%). `InterpretableDriftDetector` for feature-interaction drift attribution with BH FDR control (zero false positives in benchmarks). Current: v0.8.0.
_Benchmark: Peeking t-test 25% FPR (5x nominal); mSPRT holds at 1%. PITMonitor FPR ~3% vs repeated H-L 46%. InterpretableDriftDetector: 2/5 drifting factors correctly attributed, zero false positives_ | `pip install insurance-monitoring` | | [insurance-governance](https://github.com/burning-cost/insurance-governance) | **Flagship.** Model governance — PRA SS1/23 validation reports, model risk management, risk tier scoring
_Benchmark: Automated suite catches age-band miscalibration manual checklists miss_ | `pip install insurance-governance` |
_Benchmark: Walk-forward catches 10.5% optimism that k-fold hides_ | `pip install insurance-cv` | | [insurance-conformal](https://github.com/burning-cost/insurance-conformal) [](https://github.com/burning-cost/insurance-conformal/actions/workflows/ci.yml) | **Flagship.** Distribution-free prediction intervals for insurance GBMs and GLMs — 5 non-conformity scores tuned for Tweedie/Poisson claims (`pearson_weighted` default, `pearson`, `deviance`, `anscombe`, `raw`), frequency-severity conformal intervals (Graziadei et al.), online retrospective adjustment (RetroAdj), Solvency II SCR bounds. v0.6.0.
_Benchmark: pearson_weighted score 13.4% narrower intervals than parametric Tweedie at identical 90% coverage; 50k synthetic UK motor policies_ | `pip install insurance-conformal` | | [insurance-conformal-ts](https://github.com/burning-cost/insurance-conformal-ts) | Conformal prediction for non-exchangeable claims time series — ACI, EnbPI, SPCI, Poisson/NB scores
_Benchmark: Sequential conformal vs static prediction intervals on claims time series benchmarked_ | `pip install insurance-conformal-ts` | | [insurance-deploy](https://github.com/burning-cost/insurance-deploy) [](https://github.com/burning-cost/insurance-deploy/actions/workflows/ci.yml) | Champion/challenger pricing framework — shadow mode, SHA-256 routing, SQLite quote log, bootstrap LR test
_Benchmark: Champion/challenger routing, shadow mode, quote logging benchmarked_ | `pip install insurance-deploy` | | [insurance-monitoring](https://github.com/burning-cost/insurance-monitoring) [](https://github.com/burning-cost/insurance-monitoring/actions/workflows/ci.yml) | **Flagship.** Model drift detection — exposure-weighted PSI/CSI, actual-vs-expected ratios, Gini drift z-test. `SequentialTest` for anytime-valid champion/challenger A/B testing (mSPRT, Johari et al. 2022). `PITMonitor` for calibration drift via PIT e-process martingale (~3% FPR vs repeated Hosmer-Lemeshow's 46%). `InterpretableDriftDetector` for feature-interaction drift attribution with BH FDR control (zero false positives in benchmarks). Current: v0.8.0.
_Benchmark: Peeking t-test 25% FPR (5x nominal); mSPRT holds at 1%. PITMonitor FPR ~3% vs repeated H-L 46%. InterpretableDriftDetector: 2/5 drifting factors correctly attributed, zero false positives_ | `pip install insurance-monitoring` | | [insurance-governance](https://github.com/burning-cost/insurance-governance) | **Flagship.** Model governance — PRA SS1/23 validation reports, model risk management, risk tier scoring
_Benchmark: Automated suite catches age-band miscalibration manual checklists miss_ | `pip install insurance-governance` |
Optimisation & Pricing Strategy
From rate change recommendations to live price experimentation.
| Library | What it does | Install |
|---|---|---|
| [insurance-optimise](https://github.com/burning-cost/insurance-optimise) [](https://github.com/burning-cost/insurance-optimise/actions/workflows/ci.yml) | Constrained portfolio rate optimisation — SLSQP with analytical Jacobians, FCA ENBP constraints, efficient frontier. Includes demand modelling (conversion/retention elasticity, price response curves) via `insurance_optimise.demand`
_Benchmark: Demand-curve pricing +143.8% profit lift over flat loading_ | `pip install insurance-optimise` |
_Benchmark: Demand-curve pricing +143.8% profit lift over flat loading_ | `pip install insurance-optimise` |
Time Series & Trends
Inflation, reporting delays, and dynamic pricing.
| Library | What it does | Install |
|---|---|---|
| [insurance-trend](https://github.com/burning-cost/insurance-trend) [](https://github.com/burning-cost/insurance-trend/actions/workflows/ci.yml) | Loss cost trend analysis — frequency/severity decomposition, ONS index integration, structural break detection
_Benchmark: 3.93pp MAPE improvement over naive OLS trend; structural break detection now confirmed on step-change DGP_ | `pip install insurance-trend` | | [insurance-dynamics](https://github.com/burning-cost/insurance-dynamics) | Dynamic pricing models — GAS score-driven filters, Bayesian changepoint detection (BOCPD/PELT)
_Benchmark: GAS Poisson +13% MAE improvement over GLM trend_ | `pip install insurance-dynamics` |
_Benchmark: 3.93pp MAPE improvement over naive OLS trend; structural break detection now confirmed on step-change DGP_ | `pip install insurance-trend` | | [insurance-dynamics](https://github.com/burning-cost/insurance-dynamics) | Dynamic pricing models — GAS score-driven filters, Bayesian changepoint detection (BOCPD/PELT)
_Benchmark: GAS Poisson +13% MAE improvement over GLM trend_ | `pip install insurance-dynamics` |
Other Libraries
| Library | What it does | Install |
|---|---|---|
| [insurance-survival](https://github.com/burning-cost/insurance-survival) | Survival models — cure models, customer lifetime value, lapse tables, MLflow wrapper
_Benchmark: Cure model recovers 34.1% cure fraction (true 35.0%); KM/Cox extrapolate to zero_ | `pip install insurance-survival` | | [insurance-telematics](https://github.com/burning-cost/insurance-telematics) | **Flagship.** HMM-based driving state classification and GLM-compatible risk scoring from raw telematics trip data
_Benchmark: HMM state features 3–8pp Gini improvement over raw trip averages_ | `pip install insurance-telematics` | | [insurance-synthetic](https://github.com/burning-cost/insurance-synthetic) | Vine copula synthetic portfolio generation — exposure-aware, preserves multivariate dependence
_Benchmark: Copula 64% better correlation preservation vs naive_ | `pip install insurance-synthetic` | | [insurance-datasets](https://github.com/burning-cost/insurance-datasets) | Synthetic UK motor portfolio with known DGP parameters — validate that your model recovers true relativities before using real data. Polars output supported via `polars=True` on `load_motor()` and `load_home()`
_Benchmark: GLM parameter recovery RMSE 0.069; OVB demo shows 24% NCD inflation when age omitted_ | `pip install insurance-datasets`
Polars: `pip install insurance-datasets[polars]` |
_Benchmark: Cure model recovers 34.1% cure fraction (true 35.0%); KM/Cox extrapolate to zero_ | `pip install insurance-survival` | | [insurance-telematics](https://github.com/burning-cost/insurance-telematics) | **Flagship.** HMM-based driving state classification and GLM-compatible risk scoring from raw telematics trip data
_Benchmark: HMM state features 3–8pp Gini improvement over raw trip averages_ | `pip install insurance-telematics` | | [insurance-synthetic](https://github.com/burning-cost/insurance-synthetic) | Vine copula synthetic portfolio generation — exposure-aware, preserves multivariate dependence
_Benchmark: Copula 64% better correlation preservation vs naive_ | `pip install insurance-synthetic` | | [insurance-datasets](https://github.com/burning-cost/insurance-datasets) | Synthetic UK motor portfolio with known DGP parameters — validate that your model recovers true relativities before using real data. Polars output supported via `polars=True` on `load_motor()` and `load_home()`
_Benchmark: GLM parameter recovery RMSE 0.069; OVB demo shows 24% NCD inflation when age omitted_ | `pip install insurance-datasets`
Polars: `pip install insurance-datasets[polars]` |