Chapter 32 — Key Takeaways

A one-page card for the predictive models that price risk. You don't build them; you must read them well enough to use, question, and override them.

The core claims

  • A predictive model is an industrialized rate table. Its one essential gain over the classical one-way method is multivariate estimation — every factor estimated together, each one disentangled from the others. Everything else (speed, scale, third-party data) is secondary to that.
  • The GLM is the workhorse because it is interpretable. It splits the price into a frequency model (Poisson) and a severity model (gamma); the log link makes the factors multiply, reproducing a rate manual. You can read, argue, and defend every relativity.
  • The GBM out-predicts the GLM but you can't see inside it. It finds interactions automatically; it cannot be explained factor by factor. Rule of thumb: GLM where you must explain the price; GBM where you must rank the risk.
  • Neural networks change what data underwriting runs on (images, satellite tiles) — they extend the inspector's reach but fail in alien ways and demand human confirmation.
  • The inputs, not the algorithm, usually decide the result. Feature engineering is where predictive power lives, where the underwriter's domain knowledge enters, and where bias is stopped or let through.
  • A model is only as trustworthy as its out-of-sample validation. Lift and Gini measure separation, not price adequacy. Never judge a model on its training data.
  • The documented override is the most important artifact in model-era underwriting. Override only when you can name what the model lacks; write it down; log it so it feeds the next model.

The rule of thumb

Override the model only when you can name the specific fact it lacks — and a reasonable reviewer would agree that fact changes the answer. Three valid reasons: (1) the model is missing a material fact; (2) the model is out of its domain (novel/thin data); (3) the model is demonstrably wrong on this case. "I have a feeling" is not on the list. An undocumented override looks like caprice to an auditor and bias to a regulator — even when it's right.

The four questions to ask of any model

  1. Out-of-sample? Was lift/Gini measured on data the model never saw (ideally a later period)?
  2. Does the lift hold across segments? Great overall lift can hide flat or reverse lift in a key sub-book.
  3. Is it stable? Does the lift persist when the model is refit, or swing?
  4. What's the price level? Lift proves ranking, not adequacy — check the rate against your loss ratios.

Key terms

  • Generalized linear model (GLM) — predictors → outcome via a link function and error distribution, all effects estimated together; Poisson frequency × gamma severity = modeled pure premium. The interpretable pricing workhorse.
  • Gradient boosting machine (GBM) — an ensemble of many small trees, each correcting the last; highly accurate, finds interactions automatically, low interpretability; favored for selection.
  • Feature engineering — constructing/selecting the input variables a model learns from; the underwriter's seat at the table; where bias enters or is stopped.
  • Model validation / backtesting — testing performance on data the model didn't learn from; cardinal rule: never judge a model on its training data.
  • Lift / Gini — diagnostics of discriminating power (deciles best→worst; one summary number 0→1); both measure ranking, not price adequacy.
  • The pricing-model lifecycle — data → features → build → validate → file → deploy → monitor → refresh; the underwriter contributes features at the front and logged overrides at the back.

What you could defend to your manager

"The model scored Harbor Steel a 7, decline-leaning, on the inputs it had — two fires, an aging roof, the cat zone, the hot-work class. I overrode it to a 6, and here's the documented reason: the model was missing material facts — the signed roof-replacement contract and the new hot-work permit program that turn the loss history from a frequency signal into a management story we've conditioned. I logged the override so the next model gets a feature for post-loss corrective controls. The score sharpened the decision; it didn't make it. That's judgment and analytics working together — and I can defend every word of it to the committee and the regulator."