Proxy discrimination, protected characteristics, indirect use detection, and long-run fairness under feedback loops. Not the academic fairness literature — the specific problems UK insurers face under FCA supervision. 12 articles.
Applying fairness constraints, calibration corrections, and drift monitoring as sequential post-hoc steps is how most UK pricing teams work. It is also architecturally broken. E...
A companion post to our March coverage of arXiv:2603.17106. This one walks through the code: where exactly in a UK proxy-based fairness audit does the bias from Xin et al. enter...
The FCA is consulting in Q2 2026 on exactly this question. Every layer in a DA chain — insurer, MGA, broker, PCW — is expected to evidence outcomes, not just process. Here is wh...
insurance-conformal v1.3.1 adds ConditionalCoverageAssessor — a tool for detecting and decomposing conditional coverage failures in conformal prediction intervals. Here is the p...
Chouldechova (2017) proved that when group base rates differ, no classifier can simultaneously achieve calibration within groups, equal false positive rates, and equal false neg...
EquiPy (Fernandes Machado, Charpentier et al.) does distributional fairness correction via Wasserstein barycenters. insurance-fairness does proxy discrimination auditing for UK ...
Le, Denis and Hebiri (arXiv:2604.02017, April 2026) show that enforcing demographic parity over the full prediction distribution is both accuracy-costly and unnecessary. The rig...
Nayak's Calibrated Credit Intelligence (arXiv:2603.06733) is a credit paper, but UK insurers should read it. It addresses uncertainty quantification, fairness constraints, and t...
The Lindholm-Richman-Tsanakas-Wüthrich EJOR 2026 paper gives pricing teams something they did not previously have: an instance-level proxy discrimination score for every policy ...
The per-transition Lindholm correction for multi-state models is sound. But arXiv:2602.04791 has four gaps that matter before you build on it: no accuracy degradation figures, d...
Thibodeau et al. build a multi-firm market simulator and train an RL social planner to design fairness tax schedules. The collusion result stops the paper cold: a cartel that ex...
An FCA Research Note (December 2025) found a £28 unexplained ethnicity residual across six million motor policies. Your pricing team cannot measure it because you do not hold et...
Miao & Pesenti's KL discrimination-insensitive result is theoretically clean. Deploying it in a production GLM-based pricing system is not. The paper is silent on how to extract...
Most fairness sections in model validation reports say nothing. The KL discrimination-insensitive result from Miao & Pesenti (2026) gives you a precise, defensible claim: we app...
Miao & Pesenti (2026) derive the nearest fair probability measure in KL divergence. Their existence theorem is useful for FCA Consumer Duty attestation. Our insurance-fairness l...
insurance-fairness v1.2.0 adds PrivatizedFairPricer: discrimination-free pricing when the sensitive attribute is privatised via local differential privacy. Based on Zhang, Liu &...
The FCA's pure protection market study interim report landed in January 2026. The final report is due Q3 2026. For income protection pricing teams, the central question is wheth...
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 ...
The FCA has flagged travel insurance for mental health conditions and contents insurance for social renters as supervisory priorities. Here is what ML can genuinely help with, w...
Standard discrimination-free pricing applies to a single outcome. Income protection premiums are derived from a matrix of transition rates. Applying Lindholm to the aggregate pr...
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...
Solvency II requires age-based pricing. EU AI Act Article 21 lists age as a prohibited characteristic. Chouldechova's impossibility theorem shows you cannot satisfy both fairnes...
Age is a legally permitted rating factor under Solvency II. Age is also a protected characteristic under EU law. The EU AI Act imposes non-discrimination obligations on high-ris...
Vadlamani et al. (ICLR 2025, arXiv:2505.16115) formalise fairness at the prediction-set level. A model can be statistically valid at 90% coverage while covering elderly policyho...
Naively applying Wasserstein barycenter corrections sequentially across multiple protected attributes is miscalibrated: the ECDF for attribute k was fitted on the original predi...
Lindholm, Richman, Tsanakas and Wüthrich (EJOR, January 2026) give us a scalar proxy discrimination measure with a property none of the existing methods share: PD=0 if and only ...
Thibodeau et al. build a multi-firm market simulator and demonstrate the collusion pathology: a cartel that excludes every income group equally passes standard demand-fairness m...
Denuit, Michaelides & Trufin (arXiv:2603.16317) prove that autocalibration and group fairness are mathematically equivalent. A GBM that is well-calibrated overall but miscalibra...
Lim, Xu & Zhou (arXiv:2602.04791) show that multi-state models decompose into independent Poisson GLMs — one per transition — making every single-period fairness method directly...
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...
Miao & Pesenti (arXiv:2603.16720) derive the discrimination-insensitive pricing measure from first principles: find the nearest probability measure Q to the real-world P, in KL ...
Zhang, Liu and Shi (arXiv:2504.11775, 2025) extend discrimination-free pricing to the case where you only have a noisy privatised version of the protected attribute. The correct...
Every UK fairness audit that uses LSOA ethnicity percentages as its protected attribute is reporting a lower bound on bias, not the true figure. Here is what that means in pract...
The FCA's 2026 access priorities are not a compliance problem — they are an actuarial evidence problem. The question is whether the data exists to justify what your models are d...
Single-objective fairness constraints force a binary choice. NSGA-II finds the full tradeoff surface, so governance committees can make an explicit, documented decision about wh...
Miao & Pesenti (arXiv:2603.16720) derive discrimination-insensitive premiums by finding the probability measure nearest to the real-world measure in KL-divergence, subject to ze...
Denuit, Michaelides and Trufin (March 2026) unify autocalibration and non-discrimination into a single actuarial test. If your model fails it, you have a pricing problem and a r...
We ran insurance-fairness against ausprivauto0405 — a real Australian motor dataset with an explicit Gender field. Here is what FairnessAudit, MulticalibrationAudit, and Indirec...
A new paper (Xin, Hooker, Huang 2026) shows that BIFSG proxy race distorts regression-based fairness audits in two distinct mechanisms — and the direction of distortion is group...
NSGA-II finds the non-dominated pricing strategies across accuracy, group fairness, and counterfactual fairness simultaneously. TOPSIS turns that front into an auditable regulat...
A February 2026 paper from Lim, Xu, and Zhou cracks open the problem of fair pricing in multi-state insurance products — the ones that matter most for Consumer Duty obligations ...
insurance-fairness v0.6.3 ships DiscriminationInsensitiveReweighter. Here's why dropping the protected column doesn't work, how propensity-based reweighting does, and what the A...
EP25/2 (the FCA's evaluation of GIPP price-walking remedies) flags ongoing fair value supervision in motor and home. No single technical checklist exists for the pricing actuary...
Mutual information, proxy R-squared, and SHAP proxy scores all flag proxy discrimination but catch different things. A practical guide to interpreting conflicting signals in ins...
Under Consumer Duty and the Equality Act 2010, non-life insurers must test whether rating factors act as proxies for protected characteristics. Here is exactly how to run that t...
The FCA expects pricing teams to demonstrate their models don't proxy-discriminate under Consumer Duty. Most teams do this in Excel. Here is how to do it properly in Python, usi...
Most UK insurers don't hold ethnicity, religion, or disability status. Consumer Duty still requires evidence of fair outcomes. PrivatizedFairnessAudit solves this with local dif...
Fairlearn is excellent for classification fairness. It was not built for insurance pricing, the Equality Act 2010, or the FCA's specific concern: proxy discrimination in a multi...
Active Consumer Duty investigations in home and travel insurance. What a defensible pricing model actually requires under PRIN 2A, and what the FCA thematic review said about mo...
EquiPy is a technically excellent fairness correction tool built on optimal transport theory, from Arthur Charpentier's group at UQAM. insurance-fairness is an FCA-focused proxy...