ML methods applied to insurance pricing — GBMs, neural networks, embedding models, foundation models, and the recurring question of whether the complexity is justified by the lift. 6 articles.
A February 2026 paper provides the first statistically valid confidence intervals for global SHAP feature importance. We explain what changes for UK insurance pricing teams, whe...
insurance-glm-tools v0.2.0 ships RobustMMDGLM — a Gamma GLM that automatically downweights large losses and selects features via L1, based on Kang & Kang (2026). Replaces ad-hoc...
insurance-conformal v1.2.0 adds LCPModelSelector: locally adaptive conformal model and score selection that gives per-prediction tighter intervals while maintaining coverage gua...
Shankar & Cohen automate GAM structure search using NSGA-II evolutionary algorithms. The idea is legitimate; the problem is that EBM already does this better for insurance prici...
Pricing teams treat fairness as a single slider between accuracy and parity. NSGA-II reveals it is a landscape with multiple competing criteria. Here is what the Pareto front lo...
Kong, Liu & Yang prove that standard conformal coverage guarantees degrade unevenly when protected attributes are absent at test time. With post-ECJ gender prohibition and GDPR ...
Marginal 90% coverage can hide severe undercoverage for specific risk profiles. ConditionalCoverageAssessor — new in insurance-conformal v1.2.0 — quantifies it with CVI, decompo...
Laub/Pho/Wong's Actuarial Neural Additive Model has a genuine architectural insight in PWLCalibration monotonicity. It also depends on an unmaintained TensorFlow library. EBM is...
k-ary randomised response applies symmetric noise — wasteful and fairness-suboptimal. The Ghoukasian-Asoodeh optimal mechanism is asymmetric: minority groups get a higher correc...
A model can pass its age fairness audit and its gender fairness audit and still systematically overprice young women. This is fairness gerrymandering. We explain the CCdCov meas...
insurance-cv v0.3.0 adds SupportPointSplit (distributional train-test splitting via energy distance minimisation) and ChatterjeeSelector (nonlinear feature screening using Chatt...
When you launch a new product with no claims history, you borrow from a related portfolio. Transfer learning formalises this. But the most-cited deep learning method for domain ...
Text embeddings can compress 13,000 business categories into 24 useful features and add 2-3% Gini lift on commercial lines. For UK personal motor FNOL they do essentially nothin...
Tab-TRM sets the French MTPL benchmark at 23.589×10⁻² Poisson deviance, beating PIN ensemble by 0.3%. The linearisation result — Tab-TRM is approximately a state-space model — i...
Naively applying Wasserstein barycenter corrections sequentially across multiple protected attributes is miscalibrated: the ECDF for attribute k was fitted on the original predi...
The quantile premium principle maps a single number — your risk appetite parameter tau — to per-risk safety loadings. Zanzouri et al. (NAAJ 2025) shows QRNN outperforms tree-bas...
LLMs encode societal stereotypes. When you use one to generate rating features, those stereotypes enter your pricing model. The insurer is responsible. Here is the testing proto...
Three published frameworks use LLMs to generate tabular features and beat classical search tools on generic benchmarks. None has been tested on an actuarial dataset. We explain ...
Focal loss is a clever idea from computer vision that does not translate well to tabular insurance fraud data. AUC=0.63 from a three-stage focal loss neural network versus 0.75-...
FL solves the same variance-reduction problem as Bühlmann-Straub — but iteratively, with communication overhead, and without actuarial precedent. For UK personal lines, that tra...
Parametric Tweedie intervals over-cover low-risk policies and under-cover the high-risk tail. Conformal prediction fixes this with a finite-sample guarantee that does not rely o...
Marginal coverage guarantees say nothing about which policyholders are being undercovered. CVI decomposes conditional coverage into undercoverage risk and overcoverage cost. CC-...
insurance-distributional v0.3.0 ships ZeroInflatedTweedieGBM — the first open-source implementation of So & Valdez (2024) Scenario 2. When standard Tweedie gets structural zeros...
XGBoostLSS, LightGBMLSS, NGBoost, and PGBM can all output a full conditional distribution rather than a point prediction. The Chevalier & Côté benchmark (EAJ 2025) tested 11 alg...
LLMs can extract structured data from surveys, flag non-standard construction in loss adjuster notes, and rate categorical variables at scale. They cannot reliably assess claim ...
Avanzi, Richman, and Wüthrich reformulate individual claims reserving as a Markov Decision Process. We explain why it matters, what it actually does, and when a UK reserving act...
Laub, Pho and Wong's ANAM paper enforces smoothness and monotonicity architecturally, not as penalties. Here is what the mechanism actually is, why it matters more than the benc...
A practical comparison of CatBoost and XGBoost for UK personal lines insurance pricing — categorical handling, Tweedie support, and why we default to CatBoost.
Shared-trunk neural model for frequency-severity dependence in UK motor pricing. Explicit dependence testing where two-part GLMs assume independence - Python.