Machine learning underwriting

Category: AI in insurance · Reviewed by Simon Temme, Account Executive · Last reviewed 2026-06-10

Machine learning underwriting is the use of supervised statistical-learning algorithms — including generalised linear models, gradient-boosting machines and neural networks — to predict claim frequency, severity, retention and lifetime value for the purpose of risk selection and pricing in United Kingdom general insurance. It is the technical core of modern personal-lines pricing and is increasingly used in commercial and specialty lines.

Category: AI in insurance · Aliases: ML underwriting, supervised learning underwriting · Established: GLM-based ML pricing in UK personal lines from the late 1990s; non-linear ML at scale from c.2015 · Related: AI underwriting, Algorithmic underwriting, Generalised linear model (GLM)

Definition

Machine learning underwriting refers to the disciplined construction of predictive models from historical insurance data, where an algorithm “learns” the relationship between explanatory variables (driver age, postcode, vehicle, building characteristics, exposure) and a target (claim frequency, claim cost, conversion). The output is a probability or expected value used to inform a technical price or an accept/decline decision.

Legal / Regulatory basis

• FCA Handbook SYSC 4 (general organisational requirements) and SYSC 7 (risk control): the firm’s senior management is responsible for adequate governance of any models that materially affect customer outcomes or capital.
• PRA Solvency II model documentation requirements; the PRA SS1/23 Model risk management principles for banks (May 2023) is widely referenced by insurance model-risk teams as directional best practice for tiering, validation and ongoing monitoring.
• FCA & PRA DP5/22 and FS2/23 confirm that “AI” — including conventional supervised ML — is already in scope of existing rules; no new ML-specific permission is required, but firms must evidence the governance they apply.
• FCA general insurance pricing rules in ICOBS 6B (the price-walking remedy) constrain how ML retention models may be used at renewal in personal lines.
• UK GDPR Articles 5, 22 and 35; Equality Act 2010 in respect of indirect discrimination via proxy variables.
• FCA Consumer Duty (PS22/9): ML pricing models must deliver fair value and be capable of being explained to retail customers at outcome level.

How it works in practice

A typical ML underwriting build follows a familiar lifecycle:

1. Data preparation — historical policy and claims records, joined to third-party enrichment, are cleaned and de-duplicated.
2. Train/validation/test split — usually a time-based split (e.g. train on years t-5 to t-1, validate on year t, test on year t+1) to mimic deployment.
3. Feature engineering — including interaction terms, monotonic constraints to encode underwriting intuition, and bias review of any geographic or demographic proxies.
4. Algorithm selection — most UK pricing teams maintain a GLM as the reference model and may layer a GBM “lift” or a constrained neural network.
5. Validation — out-of-sample lift, calibration, fairness slicing and stability under stress.
6. Documentation — a model risk file covering purpose, data lineage, assumptions, limitations and ongoing-monitoring plan, satisfying SYSC 4 governance.
7. Deployment — into a rating engine, with kill-switches, fallback rates and override logging.

The discipline of independent validation, periodic re-fit and challenger-model comparison is the practical content of “good governance” as set out in the AIPPF Final Report (2022) and the EIOPA AI Governance Principles (2021).

Common variations / Subsequent developments

• Frequency-severity decomposition for short-tail lines, vs. cost-of-claims modelling for long-tail.
• Tweedie-distribution GLMs that model frequency and severity jointly.
• Stacked ensembles combining a GLM base with a GBM residual model.
• Constrained neural networks that respect underwriting monotonicity constraints.
• Retention and lifetime-value models sitting alongside the risk model in commercial pricing engines, constrained at renewal by ICOBS 6B.

Example

A UK motor insurer rebuilds its private-car pricing model. The team trains a Tweedie GLM as the baseline, a LightGBM on the GLM residuals, and a small neural network on telematics traces for the policies with a black-box installed. Validation is on the most recent year. The firm’s model committee, chaired by an SMF-26 holder, signs off the model risk file. The firm’s Consumer Duty board paper records the fair-value assessment of the resulting price.

See also

• AI underwriting (batch 12 expansion)
• Algorithmic underwriting
• Generalised linear model (GLM)
• Gradient boosting machine (GBM)
• Explainable AI insurance

References

• FCA & PRA, DP5/22 — Artificial Intelligence and Machine Learning, October 2022.
• FCA & PRA, FS2/23 — Feedback Statement on AI and Machine Learning, October 2023.
• PRA, SS1/23 — Model risk management principles for banks, May 2023. https://www.bankofengland.co.uk
• Bank of England & FCA, AIPPF Final Report, October 2022.
• EIOPA, Artificial Intelligence Governance Principles, June 2021.
• FCA, PS21/5 — General insurance pricing practices, May 2021 (ICOBS 6B).
• FCA, PS22/9 — Consumer Duty, July 2022.
• UK GDPR; Equality Act 2010, https://www.legislation.gov.uk

This entry is part of the Apex Insurance Wiki. Last reviewed by Matt Bartlett on 2026-06-10. Next review: 2026-12-10.

Apex Insurance Brokers Limited. Authorised and regulated by the Financial Conduct Authority, FRN 724952. Registered in England and Wales, Companies House 07014570. This entry provides general information about UK insurance concepts and is not regulated advice. Consult your insurance broker on your specific position.