The renewal that was already wrong
A property catastrophe underwriter renews a Gulf Coast wind book in November. The hazard curve behind the pricing was fitted using a vendor catastrophe model whose event set was last recalibrated on storms through 2021. The underwriter knows this — the model documentation says so on page one — and prices anyway, because it is the model the syndicate has approved for the year and reunderwriting from scratch on every submission is not something a book of 400 risks permits. Two named storms in the intervening seasons behaved outside the tail the model assigned them: rapid intensification within 24 hours of landfall, a pattern the historical fit treated as a low-probability outlier rather than a shifting mean. The claims from those storms are still working through the reinsurance layers when the renewal is bound.
By the time the loss triangles mature and the actuaries restate the curve, the book has already renewed twice more on the old assumption. The exposure registry was current — addresses, values, construction types all correct as of binding. The catastrophe model was not current, and could not be, because it was built once and held fixed. The underwriter was not negligent. The underwriter was pricing with a prior and calling it a posterior.
What actually happened, mechanically
Insurance pricing is Bayesian whether the underwriter uses the word or not. There is a prior — the hazard curve, the loss cost per hundred dollars of insured value, derived from history. There is supposed to be a likelihood — this season's actual claims flow, the reinsurance market's revised view of correlated exposure, the catastrophe model's own updated event set once the vendor issues one. The posterior is what should be charged next.
The failure was not a bad prior. Vendor catastrophe models are careful, expensive, defensible pieces of work. The failure was a missing likelihood term at the moment it was needed. Claims data from the two intervening seasons existed. It sat in a claims system, reconciled but not yet fed back into the pricing curve, because the model refresh cycle runs annually and the annual refresh had not yet occurred when the renewal bound. The exposure registry updated continuously; the hazard curve did not. Two different clocks, one book.
This is Bayesian updating's own structure, described precisely: a prior held with some strength, evidence arriving that bears on it, a likelihood ratio that should move the belief, and a posterior that becomes the next prior. Thomas Bayes' essay on the doctrine of chances, read to the Royal Society in 1763 after his death, solved exactly this — given observed successes, what can be said about the underlying rate. Laplace generalised it for planetary masses and birth ratios. The underwriter's hazard curve is an underlying rate. The two broken storm seasons are observed successes the rate has not yet absorbed.
Why "the model was current" is the wrong test
The instinct is to ask whether the catastrophe model was up to date. That is the wrong question, because it treats currency as a property of a file rather than a property of a process. The right question is whether the pricing system has a functioning likelihood channel at all — a route by which claims flow, revised hazard curves, updated exposure registries and reinsurance terms continuously reweight the prior, as opposed to a route where they accumulate in a warehouse until the next scheduled model release.
Four data streams matter here and each has a different natural rhythm. Claims flow arrives daily, sometimes hourly during an event. Catastrophe model versions arrive annually or after a major event forces an emergency recalibration. Exposure registries update on every endorsement, which for a large commercial book can mean daily. Reinsurance terms reset at treaty renewal, typically annually but occasionally mid-term after a major loss erodes retained capacity. A book priced with a Bayesian discipline treats all four as live likelihood inputs on their own schedules, reweighting the loss cost estimate as each one lands, rather than waiting for all four to align at a single annual pricing date.
Surely this is just actuarial best practice with a philosophical label attached. Insurers have run experience-rating and credibility weighting for a century. Nothing here is new.
That is fair, and it should be conceded rather than argued away. Credibility theory — the Bühlmann-Straub weighting of an individual risk's own experience against a broader class — is a working Bayesian mechanism, decades old, already embedded in commercial lines pricing. The point is not that underwriting has never updated. It is that the update has historically been episodic and scheduled — annual model refresh, annual treaty renewal, annual rate filing — rather than continuous. The two intervening storm seasons did not wait for the schedule. Claims arrived; the posterior they implied sat unabsorbed for a year. That gap between when evidence arrives and when it is permitted to move a price is exactly what a continuously running intake removes.
The three positions, and where underwriting currently sits
Bayes' rule needs two ingredients: a prior and a likelihood. Any system pricing risk sits somewhere on how it obtains the second, and there are three positions.
A hazard curve fixed at last model release is a prior with no fresh likelihood — a frozen corpus, in the vocabulary of the wider lineage this concept belongs to. This is the position of a Large Language Model: rich, carefully built, and stopped at a cutoff. It answers every submission with confidence calibrated to a world that no longer quite exists.
A pricing exercise that pulls live claims and current exposure for the duration of a single renewal — a bordereau review, a treaty negotiation — obtains a likelihood, but only for the episode. Once the renewal binds, the thread is dropped; the next submission starts the review again with whatever the standing model says. This is the Large World Model position: sensing while the scene is in front of it, losing the thread when the episode ends.
The third position is a book whose hazard curve, claims triangle, exposure registry and reinsurance terms are all treated as permanently open channels — where this month's loss experience becomes next month's prior without waiting for an annual cycle, and where every revision carries a record of which stream produced it, so a bad vendor release or a corrupted claims feed can be identified and unwound rather than silently baked into the rate. That is a Large Universe Model: not a product any insurer runs today, but the argued endpoint of the intake axis — streams that never stop, posteriors that genuinely become the next prior, provenance retained on every update.
| position | likelihood source | update rhythm | underwriting analogue |
|---|---|---|---|
| Large Language Model | none after cutoff | never | hazard curve fixed at last model release |
| Large World Model | sensed, scoped to episode | per renewal | live bordereau review during a single negotiation |
| Large Universe Model | continuous, provenance-tagged | permanent | claims, cat models, exposure and treaty terms all open channels |
No fourth position exists, because the only way to obtain likelihood is not at all, episodically, or continuously — there is no fourth manner of receiving evidence. After continuous intake with provenance, further progress is in calibration, in reweighting formulas, in how fast the underwriter trusts a signal — not in inventing a new channel through which evidence could arrive.
What the objections get right
The strongest objection to this whole picture is misspecification. A hazard curve is built on a hypothesis space — a set of assumed storm behaviours, a fitted distribution family. If rapid intensification within a day of landfall was never in that space at all, no amount of streaming claims data repairs it by updating within the wrong family; it just produces a confident, wrong posterior faster. This is correct and is the deepest limit on the whole approach. But it is a modelling failure, not an intake failure, and frozen models suffer it worse, because detecting misspecification requires watching the residual — the gap between what the curve predicted and what the claims triangle actually paid — accumulate over time. A model that never receives fresh claims cannot even notice it is wrong. Continuous intake does not fix a badly specified hazard family. It is the only thing that can reveal that the family needs rebuilding.
The second serious objection concerns tractability. Reweighting a hazard curve on every claims update, every exposure endorsement and every reinsurance reset is not free; approximations creep in, and a badly built weighting scheme can drift the loss cost estimate further from the true rate than an honest, stable, frozen model would. This is a real engineering risk, familiar from any recursive estimator: errors compound. The discipline against it is not to stop updating but to retain provenance — keep the raw claims record, the original model version, the exposure snapshot at each point — so the loss cost curve can be recomputed from source rather than trusted as an unauditable running total. A frozen model's stability is not evidence of correctness. It is stability purchased by refusing to look.
None of this makes the frozen hazard curve worthless. A well-built catastrophe model, honestly dated, remains a serious epistemic achievement and often prices better than a hastily assembled live feed with no discipline behind it. The claim is narrower and holds regardless: a prior cannot describe itself as conditioned on seasons it never saw. The renewal that broke did not fail because the model was bad. It failed because the price was spoken in the present tense about a past that had already moved on.