The layer on which the insurance industry’s view of physical climate risk is produced: natural-catastrophe (NatCat) models, the data that feeds them, and the people who run them. This page treats it as a market-structure object rather than a technical one.
Opened 2026-10-14 from Sounding the Alarm, which is the only source behind it. Everything the memo supplies is in one paragraph; the reason it gets a page is that the paragraph is about the price and concentration of risk information, which is this wiki’s subject and not the memo’s.
What the source says
Four claims, in the memo’s own words:
- “Modelling capacity within the industry also varies greatly between firms.”
- “some insurers have inadequate and outdated Natural Catastrophe (NatCat) models and collectively rely on a small number of base case models.”
- “Internal models are very costly to develop and robust external data comes at a steep price. As a result, some insurers are not equipped to sufficiently model complex scenarios or measure aggregated risk across the firm.”
- “there is not enough attention paid to systemic risks that sit outside of the silos we tend to organise within” — and, later, “we see a broad failure to sufficiently grapple with systemic or concentration risk at the industry level.”
The memo places these next to each other and draws no connection between them. It also qualifies: “most firms have world-leading capabilities”, and “even with many firms exhibiting best practice in this area, exposure is building across the industry.”
⟨Background: the commercial catastrophe-model market is widely understood to be held by a small number of vendors. Specific names are not in the source and are not asserted here; the memo says “a small number” and does not say which. Any use of a vendor name in SINE’s own output needs its own source — see open task 14.⟩
Why the concentration is the harder finding
The memo’s more quotable claim is the haircut (below). The structural claim is the one that does more work, for three reasons.
1. The errors are correlated, not averaged
If many insurers derive their view of a peril from the same base-case model, their exposure estimates are not independent observations of the same world. They are one estimate, reproduced.
⟨Inference: the consequence falls on whoever reads the aggregate. A supervisor assembling exposure data across a market sees what looks like convergent independent judgement and is in fact a single model’s output, counted many times. Convergence is the main thing an aggregator has to go on, and here it is uninformative. This is the memo’s own “failure to grapple with systemic or concentration risk at the industry level” one level deeper than the memo takes it: the concentration is not only in who is exposed to which peril, it is in who computed the exposure.⟩
2. The layer that defines what counts as risk sits outside the regulated perimeter
Model vendors are not insurers. They are not prudentially supervised, do not hold capital against their output, and are not accountable to policyholders — while the thing they produce is the input to decisions about pricing, capital adequacy and withdrawal.
⟨Inference: this is a shape the wiki already carries. The adtech precedent transfers the finding that the rule-setting layer is not the deepest layer — in programmatic advertising, the layer that set the rules of the auction was not where the regulators were looking, and the same displacement shows up again in the AI stack, where four agencies locate the chokepoint on four different layers. Applied here: regulation of climate risk in insurance addresses insurers, and the view of climate risk is manufactured one layer up. The transfer is structural, not rhetorical — and it is why this observation generalises rather than remaining an industry anecdote.
Where the analogy stops: adtech’s rule-setting layer was also where the rent was. Nothing in this source says the model vendors earn a rent, only that their product is expensive. That is a different claim and it is not established here.⟩
3. The concentration is an affordability artifact
The memo’s claims (1), (2) and (3) are one causal chain it does not join up: internal models are very costly, robust external data “comes at a steep price”, so capability varies greatly, so firms that cannot build fall back on the shared base case.
⟨Inference: the structure of the risk-information layer is therefore a function of the price of data, which is SINE’s object and not the insurers’. Stated as a mechanism: where verified risk-relevant data is expensive, the number of independent risk views falls to the number of parties that can afford one — and the diversity of the whole market’s judgement collapses to the budget of its largest members. This is the most transferable sentence on the page. It is also an inference from a single anonymous memo, and marked as such.⟩
The second mechanism: the haircut
Distinct from concentration, and working on top of it:
“natural catastrophe modellers and actuaries are being asked to give a ‘haircut’ to future projections or ‘manage down the numbers’ in order to avoid causing alarm and to keep getting deals done. Modelling adjustments are frequently constrained by commercial considerations, particularly where higher estimates of future physical risk could result in reduced profitability or loss of market share.”
Scoped by the memo to “some (but not all) firms”, where “those raising these issues are coming under increasing pressure.”
Why the two mechanisms belong on one page. ⟨Inference: they multiply. A commercial adjustment applied to a firm’s own model stays inside that firm; applied to a shared base case, or to the assumptions a shared base case is run under, it propagates to everyone running it. Concentration converts a local incentive problem into an industry-wide one. Neither the memo nor anything else in this wiki establishes that this has happened — the memo does not say where in the chain the haircut is applied. It is the question the two findings jointly raise, and it is answerable: it asks whether the adjustment sits in the model or in the usage.⟩
The tension this page refuses to resolve
A shared base case is also how comparability is achieved. If every insurer ran its own model, exposure figures would be independent and mutually unintelligible; a supervisor could not add them up, a reinsurer could not price across them, and a regulator could not set a threshold that meant the same thing twice. Model diversity buys error-independence and pays in comparability. Model concentration buys comparability and pays in correlated error.
⟨Inference: this is the third instance in this wiki of one recurring shape — two opposite theories of how comparability is produced. The first two are explicit and institutional: KM-GBF prescribes indicators while SEEA EA prescribes structure, drawing on the same observations; and the eight decision-grade criteria, where Criterion C’s period-averaging and Criterion H’s transaction-level resolution pull against each other for the same reason — averaging is how comparability is manufactured, and it is what destroys resolution (see temporal resolution and the service case).
This is the first instance in which the comparability-producing artifact is a priced commercial product. In the other two the trade-off is argued in public by standard-setters; here it is settled by procurement, firm by firm, with nobody arguing it at all. That is the finding worth carrying out of this page: the trade-off is not being made badly, it is not being made.⟩
What a SINE-shaped instrument would have to do here
⟨Inference, and the weakest section on this page — it reasons from a single anonymous memo to a product shape, which is two steps further than the source supports.
The memo’s own problem, stated in its own terms, is that no party can see aggregated and concentration risk across the industry, because the firms that hold the pieces cannot disclose them: exposure data is commercially sensitive, and the memo’s authors could not even put their names to a qualitative description of it. That is a bargaining shape, and it is the one case in this wiki where the actors have already demonstrated the demand by building the cheap substitute — anonymity on Signal — and accepting its cost in credibility (see Inside Track).
So the instrument would have to produce an aggregate that a supervisor can act on from inputs no participant discloses — which is the secure-aggregation problem SINE’s own benchmarking line already addresses, applied to exposure rather than to a benchmark. What is new here is not the mechanism but the counterparty: the party who needs the aggregate is a prudential regulator, who can compel disclosure and therefore may not want a cryptographic alternative to compulsion at all.
Against the instrument, stated now rather than discovered later: the useful output would be a concentration statistic across the model layer, and the parties holding that information are the vendors, not the insurers. A mechanism among insurers measures how correlated their answers are, which is a proxy. Nothing here is a proposal; it is the shape to test if the insurance line is ever picked up.⟩
What is not established
- How many base-case models, which ones, and what share of the market runs them. The memo says “a small number” and nothing more.
- Whether the haircut practice is widespread. The memo says “some (but not all) firms”.
- Whether model vendors earn a rent, as distinct from charging a high price.
- Whether supervisors are already aware of the concentration. Nothing in this wiki speaks to it, and Solvency II internal-model approval is an obvious place where they might be.
These are open task 14.
Cross-references
Sounding the Alarm (2026) · Inside Track x (Re)Insurance · The insurance protection gap · Coordination vs. bargaining failure · Adtech as a regulatory precedent · AI stack market structure · Nature-related reporting frameworks · Decision-grade carbon-data criteria · Coffee — Minimum Viable Context
Sources
sources/insurance/2026-10-14-inside-track-reinsurance-sounding-the-alarm.pdf.
Single-source page.