Insurance and the pricing of tail risk: why continuous ingestion follows

Pricing a promise against the improbable

An insurance premium is a price attached to a promise. The insurer promises to pay if a defined bad thing happens; the premium is what the promisee pays now for that contingent future payment. Underneath the promise sits an estimate: expected loss, plus a loading for capital, uncertainty and profit. Expected loss is a mean. Means are the easy part. What makes the pricing hard is that most of the risk in a well-run book does not live near the mean at all — it lives in the tail, in the rare, severe outcomes that barely move the average but consume the capital when they arrive.

Tail estimation is hard for two separable reasons, and it matters that they are separable. The first is that data are thin by construction — a one-in-two-hundred-year flood does not offer two hundred years of clean observation, and the worse the event, the fewer examples of it exist anywhere. The second is that the thing being estimated does not sit still. Coastlines develop. Buildings age and are replaced with different ones. Litigation norms shift, juries change their minds about what harm is worth, and construction costs move independent of any hazard at all. A premium is a claim about a distribution, and a distribution is not a fixed object waiting to be measured once and filed. It drifts under the feet of anyone trying to price it.

This is why actuaries distinguish frequency from severity, and hazard from exposure and vulnerability. How often does the event occur; how bad is it when it does; and — often the most volatile term of the three — how much value sits in harm's way, and how well is that value protected. A wildfire model can get the physics of fire spread exactly right and still misprice the year, if it does not know that vegetation has encroached on the transmission corridor since the model was last calibrated. The physics is the stable part. The world in front of the physics is the part that moves.

Where the discipline came from

The practice is older than the theory. Lloyd's marine syndicates in the eighteenth century spread hull risk across many underwriters precisely because no single one could absorb a total loss on a single voyage — an early, informal answer to the tail problem, solved by diversification rather than measurement. The formal answer came later, with credibility theory, associated with Whitney's early twentieth-century work and given its enduring shape by Hans Bühlmann in the 1960s. Credibility theory asks a precise question: how much weight should an actuary give to a policyholder's own thin experience versus the broad experience of the collective. The answer is a weighted blend, and the weight itself is a function of how much can be trusted from few observations. It is, in effect, an early formalism for combining sparse local signal with abundant background signal — a problem that reappears, unrecognised, a full century later.

Catastrophe modelling arrived as a commercial discipline in the late 1980s, with firms such as AIR Worldwide and RMS coupling stochastic event sets — thousands of simulated storms or earthquakes, statistically consistent with historical hazard — to engineering vulnerability curves describing how buildings actually fail. It was a genuine advance: a way to generate tail scenarios that had never occurred but were physically plausible. Hurricane Andrew, in 1992, tested it and found the gap. Insured losses ran to roughly $15.5 billion and eleven insurers became insolvent. The storm physics in the models was not badly wrong. The exposure inventory was: Florida's insured coastal exposure had roughly doubled since the last comparable landfall, and the models were still rating against the coast as it had been, not as it now stood. The corpus had stopped. The coastline had not.

The turn

Put the two failure modes side by side and a pattern appears that has nothing to do with meteorology. A model trained on a fixed historical record, complete to some cutoff, cannot register that the world has moved since the cutoff. That is not a flaw peculiar to catastrophe models. It is the defining property of a Large Language Model: a system that prices — predicts, completes, judges — from a frozen corpus, internally coherent and externally stale the moment the world drifts past the date the corpus stopped being collected. The actuary rating 2025 hurricane exposure on 1990–2015 loss experience is doing, procedurally, exactly what a corpus-bound model does when asked about anything that happened after its training data ends.

There is a second position, and insurance has that one too. A drone flight over a roof, a telematics feed running during a single trip, a loss adjuster standing in the wreckage: these are direct, current, and bounded. They see what is in front of them with real fidelity and see nothing else. This is the Large World Model position — accurate about the scene, blind to everything outside the scene's edges. The adjuster on site can tell you precisely what happened to this roof. Nothing in that vantage tells them that reconstruction costs rose 40 percent since the policy was written, or that a jury three counties over just set a verdict that will reprice every open casualty claim in the region.

What underwriting has always wanted, and rarely had the means to build, is neither of those. It is the position where every stream bearing on the hazard keeps running, and each observation entering the picture arrives as a belief with a source and an age attached, revisable when the source updates rather than trusted forever the moment it is filed. Sea-surface temperature anomalies. Construction cost indices. Court dockets. Satellite land-use change. Live sensor telemetry from the asset itself. Held together, dated, and open to correction, this is the underwriting position that credibility theory was gesturing at three centuries too early: weight the sparse against the abundant, and keep both current. This is the Large Universe Model. Insurance did not invent the idea. It has simply always needed it and never quite been able to afford it.

What continuous intake does not buy

State the misreading plainly so it can be set aside. The weak, seductive version of this argument says that enough live data makes tail risk predictable — that with sufficient streams, the hundred-year flood stops being a surprise. That version is false, and anyone selling it is selling something the insurance industry already knows is unsellable. Rare events stay rare because rarity is a property of the world, not of the observer's instrumentation. Deep uncertainty about novel perils does not dissolve because more sensors exist. No stream announces the year the flood arrives. What continuous intake buys is narrower and more useful: it keeps the observable terms of the loss curve current — exposure, vulnerability, cost — and it dates every assumption, so that staleness becomes a visible, checkable property of the model rather than an invisible one discovered after the claim arrives. It is a claim about knowing what you don't know, and when you last checked. Nothing more.

Objections that hold ground

Continuous data cannot manufacture tail observations. A thousand real-time streams add covariates, not hurricanes. You get overfitting on a handful of catastrophes and phantom drift signals.

This is correct where it bites. Frequency of genuinely rare events is not rescued by volume of ancillary data; no satellite feed produces an extra hundred-year storm to learn from. But the loss curve has terms besides frequency. Exposure and vulnerability are dense and observable at high cadence — land-use change and reconstruction indices move continuously and can be measured continuously. The right posture is parsimony in the hazard physics and breadth in everything that determines who stands in the hazard's way. That is a real constraint on the claim, not a rebuttal of it.

Insurance is a regulated pricing exercise. Rate filings, approved factors and anti-discrimination law bind what may be used and how fast a price can move. A continuously updating belief is often legally unusable.

Also correct, and often decisive in personal lines: US homeowners and auto rates in several states lag filed evidence by years, sometimes filed evidence is years stale before it is even approved. But the constraint sits on the boundary between belief and price, not on belief formation itself. Reinsurance, parametric cover, marine and large commercial property are negotiated rather than filed and already move on near-live signal. Reserving and capital allocation run on the belief regardless of what the filed rate says. The claim concerns what a system is permitted to observe, not what it is permitted to charge — and those turn out to be different questions with different answers by line of business.

Live pricing creates reflexivity. Once insureds see the signal, they optimise against it — gamed telematics scores, cosmetic roof repairs staged before an inspection flight — and the signal stops measuring what it measured.

Genuine, and it is why static rating factors sometimes beat cleverer live ones. But the fix is provenance, not retreat to a frozen corpus. A belief tagged with its source can be discounted the moment that source becomes strategically contaminated; an untagged number cannot be discounted at all, because no one remembers where it came from. Gaming is itself a detectable pattern at scale — a cluster of pre-inspection roof replacements is visible if anyone is watching for it. The answer to adversarial measurement is more streams with attribution, not fewer streams.

What this establishes, and no more

The argument establishes that pricing a moving distribution from sparse evidence rules out both a frozen corpus and a bounded scene, and that what remains — continuous observation of every relevant stream, held as dated, sourced, revisable belief — is not one option among several but the only position left standing once the first two are excluded. It does not establish that such a position predicts catastrophe, satisfies regulators, or resists a determined adversary. Those are separate fights, and insurance has been fighting them for three hundred years without fully winning any of them. What the concept fixes is narrower: the direction of travel. Wider coverage, better provenance, longer records. Scale, trust, and time — not a fourth kind of evidence, because there isn't one.