Gradient boosting machine (GBM)

Category: AI in insurance · Reviewed by Taylor Watts, Broker · New Business · Last reviewed 2026-06-10

A gradient boosting machine (GBM) is an ensemble machine-learning algorithm that builds a strong predictive model by sequentially fitting small decision trees to the residuals of the model so far, optimising a chosen loss function by gradient descent in function space. In United Kingdom general insurance, GBMs have become the leading “lift” model on top of GLMs in personal lines and a growing tool in commercial lines triage.

Category: AI in insurance · Aliases: GBM, XGBoost, LightGBM, CatBoost · Established: Friedman (1999, 2001); XGBoost (Chen & Guestrin, 2016); insurance adoption from c.2017 · Related: Generalised linear model (GLM), Machine learning underwriting, Explainable AI insurance

Definition

A GBM is a non-parametric supervised model that captures complex non-linear interactions between explanatory variables and a target. The three implementations most used in UK pricing are:

• XGBoost (open source, Chen & Guestrin, 2016) — widely used historically;
• LightGBM (Microsoft, 2017) — faster on large datasets, leaf-wise tree growth;
• CatBoost (Yandex, 2017) — strong handling of categorical variables without explicit one-hot encoding.

All three can be configured for Poisson, Tweedie or gamma objectives, allowing pricing-team workflows that mirror existing GLM practice.

Legal / Regulatory basis

• FCA Handbook SYSC 4 — governance of models that materially affect customer outcomes; the GBM, like the GLM, is in scope.
• FCA Consumer Duty (PS22/9) — the consumer-understanding outcome means a firm must be able to explain why a price moved at outcome level, even when the underlying model is a GBM.
• FCA & PRA DP5/22 and FS2/23 — directly address the trade-off between predictive lift and explainability; firms must document how they manage it.
• UK GDPR Article 22 — automated decisions with legal or similarly significant effect, including some pricing decisions, require safeguards.
• PRA SS1/23 (directional) — encourages tiering, with higher-impact GBMs receiving more onerous validation.
• Equality Act 2010 — proxy-discrimination risk is higher in GBMs because they capture interactions that may correlate with protected characteristics; bias testing is required as part of governance.

How it works in practice

1. Reference baseline — a Poisson- or Tweedie-GLM is fitted first.
2. GBM build — the GBM is trained either standalone or on GLM residuals, with early stopping on a hold-out fold.
3. Hyperparameter tuning — depth, learning rate, regularisation; usually with cross-validation.
4. Explainability layer — Shapley-value (SHAP) summaries at portfolio level and per-quote SHAP plots for selected segments.
5. Bias testing — slicing performance by sensitive proxies (postcode-derived deprivation, age bands) to identify disparate impact.
6. Governance pack — a model-risk file that documents lift over GLM, calibration, fairness slicing and ongoing-monitoring plan, in line with the firm’s tiering policy.
7. Deployment — typically as a multiplier on the GLM technical price, capped at a maximum movement to retain stability.

Common variations / Subsequent developments

• Constrained GBMs with monotonic constraints to respect underwriting intuition (e.g. expected price never decreasing with previous claims).
• Stacked models combining GLM, GBM and neural-network outputs.
• GBM-driven referral engines in commercial lines that score new business for underwriter attention.
• GBM in claims-cost modelling for setting reserves on long-tail bodily-injury claims.
• Open-source MLOps tooling (MLflow, BentoML) enabling reproducible GBM lifecycle aligned with FS2/23 governance expectations.

Example

A UK home insurer overlays a LightGBM Tweedie model on its household pricing GLM. Live A/B testing shows a 3% gain in combined ratio at constant volume. SHAP outputs identify roof age × tree-canopy proximity as a key interaction. The pricing committee approves deployment with a +/-15% cap on the GBM multiplier, a quarterly bias review, and a documented Consumer Duty fair-value assessment.

See also

• Generalised linear model (GLM)
• Machine learning underwriting
• Explainable AI insurance
• AI underwriting (batch 12 expansion)

References

• Friedman, J.H., “Greedy Function Approximation: A Gradient Boosting Machine”, Annals of Statistics, 2001.
• Chen, T. & Guestrin, C., “XGBoost: A Scalable Tree Boosting System”, KDD, 2016.
• Ke, G. et al., “LightGBM: A Highly Efficient Gradient Boosting Decision Tree”, NeurIPS, 2017.
• Prokhorenkova, L. et al., “CatBoost: unbiased boosting with categorical features”, NeurIPS, 2018.
• FCA & PRA, DP5/22 and FS2/23 — AI and Machine Learning, October 2022 / 2023.
• PRA, SS1/23 — Model risk management principles for banks, May 2023.
• FCA, PS22/9 — Consumer Duty, July 2022.
• 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.