Case Study 1: The GLM Revolution in Personal Auto Pricing

A study of how the generalized linear model moved from academic statistics to the engine of an entire line of business — and what it teaches about why models win, and what they leave for judgment. This case discusses a real, documented industry transformation. The specific carriers and the precise figures behind individual rate plans are proprietary, so this analysis stays at the level of the public, well-attested pattern and uses no fabricated statistics.

Background

For most of the twentieth century, personal-auto insurance was priced the way Chapter 11 describes the classical method: a manual rate for a class, adjusted by a handful of factors — driver age and gender, territory, vehicle, use, prior claims — each estimated more or less in isolation. The factors were few because the method demanded it: a one-way analysis can only look at one variable at a time, and the human labor of building and maintaining a rate manual put a hard ceiling on complexity. Two carriers looking at the same population would arrive at broadly similar prices, because they were all using broadly the same small set of factors estimated the same crude way.

Beginning in the 1990s and accelerating through the 2000s, that changed — first in the United Kingdom, where a competitive, lightly rate-regulated motor market and a cluster of statistically sophisticated actuaries pushed early, and then across the United States and much of the world. The change had a name: the generalized linear model. The GLM was not new mathematics — the statistical theory dates to the 1970s — but its application to insurance pricing at scale was transformative. For the first time, carriers could estimate dozens, then hundreds, of rating variables simultaneously, each one controlling for all the others, on the growing mountains of policy and claims data their computers could now hold. The crude rate manual gave way to the multivariate model.

The insurance / underwriting issue

The issue at the heart of the GLM revolution is the one §32.1 and §32.2 are built around: correlation among rating factors, and the money hiding in it. Under the old one-way method, the "young driver" factor silently absorbed the effects of the cars young drivers tend to own and the places they tend to live, because the method could not separate them. Every correlated pair of factors was either double-counted or under-counted, and the resulting prices were systematically off — not catastrophically, but consistently, risk by risk.

A GLM disentangles those effects. When all factors are estimated together, the young-driver relativity finally means "the effect of youth, holding car and territory constant," and the territory relativity means "the effect of place, holding driver and car constant." The prices that result are more accurate at the level of the individual risk. And here is the competitive engine that drove the whole industry to adopt the method: a carrier with a more accurate model can practice precise risk selection against competitors who cannot. If your model knows that a particular combination of characteristics is a better risk than the market price implies, you can charge that risk a little less, win it, and profit — while your competitor, pricing it on a cruder model, either loses the good risk or keeps it at an inadequate price. Do this across millions of risks and the cruder carrier is left, slowly, with the business the better-modeled carriers did not want: a textbook adverse selection (Chapter 1's term) inflicted not by applicants but by a competitor's superior math.

The discipline this enforced is exactly the combined ratio discipline of Chapter 3. A carrier whose model was a half-step behind did not fail dramatically; it watched its loss ratio drift unfavorably as the better risks were skimmed away and the worse risks accumulated, and it felt the pressure in the one number that tells the truth. The GLM revolution was, in this sense, a combined-ratio arms race — and a line of business that had been priced on a dozen factors came to be priced on hundreds.

What it shows

Three lessons stand out, and each maps onto a section of this chapter.

First, the model's power came from multivariate estimation, not from any single clever variable. The GLM did not win because it discovered a magic new rating factor; it won because it estimated all the factors correctly, together (§32.2). This is the unglamorous truth the chapter insists on: the gain was structural and statistical, not a trick.

Second, feature engineering quietly decided the winners. As the method matured, the competitive edge moved from "do you use a GLM?" — everyone did — to "what features have you engineered?" Carriers that built richer, smarter inputs (better territory definitions, engineered vehicle characteristics, interactions between age and vehicle) out-priced carriers using the same algorithm on cruder inputs. This is §32.5 playing out in the market: the model is rarely what decides the result; the inputs are.

Third, the regulators became central to the story, and the fairness questions sharpened. As models grew to hundreds of variables, regulators in many states pressed harder on which variables, not just on the bottom-line rate — because a model with hundreds of correlated inputs is exactly where a prohibited factor can hide as a proxy (Chapter 35's term). The credit-based insurance score (Chapter 8's term), which became a powerful GLM input, drew sustained regulatory and legislative scrutiny and is restricted or banned in several states. The very accuracy that made GLMs irresistible also made the fairness line — between pricing by risk and discriminating unfairly (Chapter 4's term) — harder to police, because the discrimination, if it occurred, was buried inside the math. That tension is the subject of Chapter 35; the GLM revolution is where it became unavoidable.

Outcome

The outcome is simply the modern market. Today, virtually all large personal-auto carriers price with multivariate predictive models; the classical one-way rate manual survives mainly in small, niche, or heavily regulated corners. The method spread from auto to homeowners, to small commercial, and into the selection and triage of larger commercial risks. GLMs remain the workhorse for the filed price because they are interpretable and defensible to regulators (§32.2, §32.3); gradient boosting and other machine learning increasingly sit alongside them for selection, triage, and the discovery of features a GLM is then given by hand. The actuary–underwriter–data-scientist triangle (§32.7) is, in large part, an institutional response to this transformation — a way to govern a price that no longer comes from a manual any one person can recompute.

Crucially, the revolution did not automate the underwriter out of personal auto so much as relocate the judgment. The high-volume, well-understood risks are now priced and often bound by model with no human in the loop (the straight-through processing of Chapter 20). The human work moved upstream — into feature engineering, model governance, and the validation that decides whether a model is trustworthy — and sideways, into the complex, non-standard, and novel risks where the data is too thin for a model to lead.

Lesson

The GLM revolution is the clearest real-world demonstration of the chapter's central claim: technology augments underwriters; it does not replace them — but it does replace the ones who refuse to understand it. A carrier that ignored multivariate modeling did not survive on the strength of its underwriters' intuition; it was selected against, risk by risk, by competitors whose models saw more clearly. And yet the model never removed the need for human judgment — it moved it to where the data ran out, made feature engineering a domain-expert's job, and made model governance and the documented override (§32.7) into core professional skills. The lesson for the underwriter reading this is not "models won, learn to code." It is "models won, learn to read them — well enough to build their inputs, judge their validation, and know the risks where your judgment still leads."

Discussion questions

1. The case argues that a carrier with a worse model suffers adverse selection inflicted by competitors, not applicants. Explain that mechanism step by step. How is it the same as, and different from, the classic adverse selection of Chapter 1?
2. "The gain came from multivariate estimation, not from a single clever variable." Why is this distinction important for how an underwriter should think about what a model is actually doing? (§32.2)
3. As GLMs grew to hundreds of variables, regulators pressed harder on which variables, not just the final rate. Why does model complexity make the fairness line harder to police, and which later chapter takes up that problem in full? (§32.5, Ch.35)
4. The case says the GLM revolution "relocated" rather than eliminated underwriting judgment. Name the three places the human work moved to, and connect each to a section of this chapter. (§32.5, §32.6, §32.7)
5. If you were the chief underwriting officer of a mid-size regional carrier that had fallen a step behind on modeling, what would you prioritize first — a better algorithm, better features, or better governance — and why? (§32.5, §32.6, §32.7)