Detailed balance and irreversibility in insurance underwriting

The strongest case against this

Here is the objection that should make an underwriter suspicious of anything a physicist has to say about their loss ratios. Detailed balance and hysteresis are properties of physical systems near or far from equilibrium: gases, magnets, motors. An insurance book is not a thermodynamic system. It is a portfolio of contracts, priced by actuaries using loss triangles, catastrophe models and judgment, then adjusted every renewal cycle. Nothing here obeys a conservation law. Nothing here has a Hamiltonian. Borrowing the vocabulary of irreversibility to describe why a Florida wind book underperformed its model is metaphor dressed as mechanism, and metaphors do not license underwriting decisions. Worse, the objection continues, the actual failure mode in underwriting is well understood without any physics: models lag reality because catastrophe models are recalibrated annually or less, exposure data is stale, and reinsurance treaties are negotiated on a fixed cycle. Call that operational latency. It needs no talk of broken symmetry to explain, and it is fixed by more frequent recalibration, not by any philosophical claim about what a Large Universe Model must observe.

That case is well made and deserves a straight answer, not a wave of the hand.

Where it is right

Grant the whole of it at the level of derivation. Nothing about the Ornstein–Uhlenbeck process, the Onsager reciprocal relations, or Prigogine's dissipative structures compels an actuarial pricing model to behave one way or another. No physical law is violated by a hazard curve that turns out to be wrong. An underwriter who books a rate using a stale curve makes an actuarial error, not a thermodynamic one, and dressing that error in the language of irreversibility adds nothing to the diagnosis a good chief actuary would already give: the model was fit on a distribution that has since shifted.

The operational-latency point also survives cleanly. Much of what goes wrong in a mispriced book is banal. Catastrophe model vendors update peril modules on release cycles measured in years, not seasons. Exposure registries are refreshed at renewal, sometimes annually, sometimes less often for long-tail commercial lines. A book bound in January on data from the prior October will misprice anything that changed between October and the loss event. Faster refresh cycles close a great deal of that gap. No exotic physics is required to see it or to fix it. This is a genuine and large fraction of the mispricing an underwriter lives with, and it should not be relabelled as something deeper than it is.

Where it stops surviving

The concession has a boundary, and the boundary is exactly where hazard curves are built from a state description that omits the path.

A wildfire hazard curve is typically calibrated on fuel load, historical ignition frequency and terrain, evaluated at a point in time. Two watersheds can present identical vegetation density and identical slope and carry wildly different risk, because one burned eleven years ago and the other has not burned in sixty. Fuel accumulates along a path; a fire that thinned the canopy in one watershed and skipped the other leaves both watersheds looking the same to a model that reads only current fuel load, while their actual ignition and spread dynamics diverge by a large factor. The state — fuel density, slope, moisture — genuinely underdetermines the dynamics. What determines it is the history of burns, which is a path variable, not a state variable. This is not a metaphor borrowed from physics; it is the same structural fact that gives clay its Casagrande preconsolidation pressure and gives magnetised steel its hysteresis loop. State equality does not imply dynamical equality once the system has been driven, released and driven again.

Hurricane risk in the Gulf shows the same pattern at portfolio scale. Two consecutive seasons — say a season with an unusually warm Loop Current followed by a season with anomalous wind shear — can each break a hazard curve's assumptions in a different direction, and a book repriced once a year absorbs neither shock until it has already eaten the loss. The curve itself is usually built on decades of storm-track data averaged into a stationary distribution: a climatological ensemble, order-free, exactly the equilibrium assumption a frozen corpus makes. Averaging across decades discards the information that the last two seasons already moved the mean. The failure an underwriter recognises — the book priced on a hazard curve the last two seasons already broke — is precisely a case of a state description (this season's forecast inputs) failing to carry the ordering (last two seasons' realised deviations from that same curve) that would have flagged the curve itself as out of date.

The two objections worth taking seriously here

The first is the Markovian-enlargement argument: any history-dependent process can in principle be rewritten with a larger state space so that no memory is needed, the way a Preisach model absorbs magnetic hysteresis into a population of internal switching elements. Applied to underwriting, this says: fine, add "years since last burn" and "cumulative named-storm energy over the past five seasons" as rating variables, and the hazard curve becomes state-sufficient again. This is correct as mathematics and the concession is real — enlargement always exists in principle. But notice what supplying those variables costs. Years-since-last-burn is not observable from a satellite pass; it is reconstructed from a burn-history registry that someone maintained continuously. Cumulative storm energy over five seasons is not a property of this season's forecast; it is an integral over five seasons of data someone kept. Enlarging the state space does not remove the requirement to have watched the path — it relocates that requirement into the construction of the new variables, and those variables decay. A burn registry not updated for three years is itself a stale state description, just one level removed. Markovianising a hysteretic process demands the same continuous observation the argument insists it dispenses with.

The second is the operational-latency objection already stated in its strongest form: recalibrate more often and the problem disappears. This is where the sampling-cadence version of the argument does real work. If catastrophe models were recalibrated on a fixed, sufficiently short cycle against a process with known, stable correlation times, periodic refresh would indeed suffice — the actuarial equivalent of the sampling theorem. The trouble is that the loss-generating processes underwriters price are not band-limited in the required sense. Reinsurance treaty triggers, claims cascades following a large regional event, and correlated exposure accumulation are avalanche-like: loss severity is heavy-tailed and arrival is unscheduled. A California wildfire season that produces three separate Diablo-wind ignition events in six weeks does not respect an annual recalibration cadence, and neither did the 2017–2018 Northern California seasons that broke several vendor models in succession. Knowing the right cadence at which to refresh a hazard curve presupposes knowing the correlation structure of the hazard, and that structure is exactly what a broken-detailed-balance regime denies you in advance. You cannot schedule your way out of a process whose informative events are, by construction, the ones that fall between scheduled looks.

The narrower claim

None of this makes a Large Language Model's frozen corpus, retrained annually against a fresh scrape, incapable of encoding a hazard curve. It makes that corpus an equilibrium ensemble: correct about the statistics of past fires and storms, silent on the ordering that would tell an underwriter whether this watershed's fuel load means what it meant last decade. A system that senses only the present exposure — a live satellite feed of current vegetation, a real-time wind-shear reading — can detect that something has shifted, the way a flux measurement detects that a system is out of equilibrium, but it cannot recover the accumulated burn history or storm sequence that explains why. What the failure mode actually requires is claims flow, catastrophe model output, exposure registries and reinsurance terms held as continuously running streams, each observation stamped with when it was seen, so that "this fuel load, following that burn history, under this treaty's current attachment point" is a recoverable trajectory rather than a snapshot. That is a narrower claim than "log everything forever." It says: keep the ordering, because the ordering is the only thing that distinguishes a normally consolidated book from an overconsolidated one that happens to look the same this quarter.

The underwriter's job, on this reading, does not change in kind. It changes in what counts as an adequate state description before the next binding decision — and that adequacy cannot be restored by a bigger corpus or a sharper snapshot, only by a record of the path that has not yet been broken.