GLMs remain core to actuarial pricing, despite machine learning gains

The real choice is control versus lift

Pricing teams keep asking the wrong question when they treat generalized linear models and gradient boosting machines like a knockout fight. The better question is which tool gives you enough predictive power without losing the ability to explain, defend and file a rate. Thomas Holmes, chief actuarial officer at Akur8, frames it that way, and it is the right lens for property and casualty pricing: the model has to work in the spreadsheet, in the review meeting and in front of the regulator.

That is why GLMs are still the backbone of actuarial pricing. They are interpretable, controllable and built for environments where every factor in the premium needs to be visible and defensible. GBMs can be stronger when the data gets huge and tangled, but they do not erase the need for structure, governance and business logic.

Why GLMs still dominate the parts of pricing that matter most

If you have ever had to explain a rate move across territories, age bands or coverage levels, you already know why GLMs hang on. A GLM lets actuaries see exactly how each variable contributes to the premium, then adjust the model variable by variable rather than hoping an opaque learner got the interaction right. That matters when you need to smooth trends across rating bands, enforce monotonicity or magnitude constraints, and apply expert judgment in sparse segments where the data simply does not give you enough signal.

GLMs also behave better when the real challenge is not squeezing out one more basis point of lift, but preserving logical relationships. They are especially useful for predictable extrapolation beyond the observed data, which is a real issue when you are pricing thin slices of exposure or maintaining consistency across nearby age, territory or coverage levels. In practice, that means GLMs are not just a legacy comfort blanket. They are the tool you reach for when the model has to be trusted as much as it has to be accurate.

The Casualty Actuarial Society has long treated that role as central, not incidental. CAS Monograph No. 5, *Generalized Linear Models for Insurance Rating*, is presented as a comprehensive guide to creating insurance rating plans with GLMs, with emphasis on model-building, data preparation, selection of model form, model refinement and validation. The 2025 revision of the monograph is a strong signal that GLMs remain core actuarial practice, not a museum piece.

Where GBMs actually earn their keep

GBMs come into their own when the data is too rich, too nonlinear and too interactive to specify by hand. Akur8’s May 21, 2026 analysis makes the point cleanly: GLMs are best when a model needs to be explained or adjusted, while GBMs shine on large, highly interacted datasets. That is not a subtle distinction. It is the difference between a pricing structure that an actuary can control and a pattern-recognition engine that can pull signal out of a pile of weak, overlapping effects.

Telematics is the clearest example. Driving data can include hundreds of features, from braking intensity and acceleration to trip timing and road type, and the interactions among those signals are exactly the sort of thing that turns a hand-built specification into a compromise. In that setting, GBMs can capture nonlinear relationships that would be painful, brittle or unrealistic to encode directly in a GLM. If the underwriting problem is basically a giant feature-engineering puzzle, the boosting model usually has the edge.

That edge becomes especially valuable in segmentation and competitive pricing. GBMs can uncover subtle pockets of risk and refine the view of the book in ways that are hard to get from a more structured model alone. The catch is that the better the model gets at prediction, the more the team has to work to keep the outputs governable and understandable.

Governance is no longer optional

The regulatory environment is a big reason this conversation has become more practical and less ideological. The National Association of Insurance Commissioners, based in Arlington, Virginia, has warned that AI systems used by insurers can create risks tied to inaccuracy, unfair discrimination, data vulnerability and lack of transparency or explainability. The NAIC’s guidance also emphasizes fairness, ethical use, accountability, compliance, transparency and a safe, secure, fair and robust system.

That matters because pricing is not a pure optimization exercise. Insurers have to explain rate movements to regulators, management and customers, and those audiences do not care that a model won a validation contest if they cannot trace the logic. A strong GBM can be useful, but in regulated insurance environments it still has to fit inside a governance framework that can stand up to review. A model that cannot be defended can become more expensive to use than a simpler model that leaves a little performance on the table.

Milliman’s take lines up with that reality. It has argued that data science and machine learning can improve underwriting and pricing, but explainability remains critical so intended users understand and accept model outcomes. That is the operating standard now: predictive methods are welcome, but only when they can be absorbed into a process that still looks and feels like insurance.

What pricing platforms need to support now

For software buyers, the practical conclusion is not to choose a camp. It is to make sure the platform supports both methods and lets the team move between them without tearing up the workflow. A modern pricing stack has to handle the compliance burden of GLM-based ratemaking and still give analysts room to test GBMs where the data justifies the extra complexity.

In real terms, that means the system should make it easy to:

• build and validate a GLM for filing and governance
• apply constraints that keep the model aligned with business rules
• test a GBM on dense, highly interacted data
• compare the two on both lift and explainability
• preserve actuarial judgment instead of burying it in automation

That is why model choice is increasingly a software architecture question, not just a statistical one. The best teams are not asking whether GLMs or GBMs are more modern. They are deciding where each model belongs in the pricing stack, and how much control they are willing to give up in exchange for signal.

The practical rule for actuarial pricing teams

If the problem is regulated, sparse or heavily file-driven, GLM should still be your first instinct. If the problem is data-rich, nonlinear and full of interaction effects that no human would want to specify by hand, GBM deserves a serious look. Most real pricing shops will need both, because the book does not live in one world or the other.

That is the actual story here: GLMs remain core because they solve the parts of pricing that insurance cannot afford to get wrong, while GBMs extend the toolkit where predictive lift is real and measurable. The winning posture is not allegiance to one model family. It is knowing exactly when explainability is the asset, when lift is the asset, and when the smartest move is to keep both on the bench.