Five questions to ask your catastrophe model before renewal

Catastrophe model output arrives at renewal looking like a measurement: an exceedance probability curve, a set of return-period losses, a number to four significant figures. It is better understood as an argument — assembled from exposure data, a hazard view, a vulnerability assumption and a financial module, any of which can be wrong independently.

These are the five questions we put to a modelled view before it is allowed to drive a structuring decision.

1. What proportion of the exposure is properly geocoded?

Everything downstream depends on this and it is the cheapest thing to fix. A portfolio geocoded largely to postcode centroid or, worse, to CRESTA zone will produce a plausible-looking curve that is systematically wrong about accumulation, because risks that are physically adjacent get smeared across a wide area and risks that are far apart get pulled together.

Ask for the geocoding match-level distribution, not a headline percentage. Street-level and better should dominate; if it does not, the first investment is in data, not in a second model.

2. How are secondary modifiers being treated?

Construction class, year built, number of storeys, roof geometry, basement presence, flood defences. Where these fields are blank, the model applies a default — usually a weighted average of the regional building stock, which may bear no relation to the portfolio being written.

The direction of the error is not predictable, which is what makes it dangerous. A specialist book of modern industrial risks defaulted to regional averages will be overstated; a book of older residential stock may be badly understated. Establish the completeness rate for the fields that matter to the peril in question and understand which way the defaults push.

3. Which loss perspective is being quoted?

Ground-up, gross of reinsurance, net of reinsurance, with or without loss adjustment expense, with or without demand surge, with or without storm-surge and post-event inflation loading. These produce materially different numbers from the same run, and it is remarkably common for two parties in a negotiation to be comparing figures on different bases without noticing.

State the perspective on every figure that appears in a submission. If a reinsurer’s technical price implies a different view of the same layer, the first thing to reconcile is not the hazard view — it is whether both sides are quoting the same thing.

4. What is the uncertainty around the return period being priced?

A one-in-two-hundred-year figure derived from an event set with thin representation at that return period carries wide confidence bounds, and those bounds do not appear on the summary page. Secondary uncertainty — the distribution of loss given an event, as distinct from the frequency of events — is frequently switched off or left at default because it slows the run down.

Ask what the secondary uncertainty setting was, and ask to see the curve rather than the three points someone extracted from it. Where a layer attaches close to a kink in the curve, small changes in assumption move the technical price a long way, and that is worth knowing before rather than after terms are agreed.

5. What does a second view say?

Vendor models embed genuine scientific disagreement — about event frequency, about clustering, about how vulnerability scales with hazard intensity. Where two established models diverge materially on the same portfolio, the divergence is information, not noise. It marks the part of the answer that is contested rather than settled.

A single-model view is defensible if the model is well-suited to the territory and peril and its limitations are stated. A single-model view presented as though it were the answer is not.

Using the answers

The purpose of this exercise is not to arrive at a better number. It is to know which parts of the structure are robust to the assumptions and which are not — and then to place the robust parts confidently and treat the rest with appropriate caution.

In practice, that usually means testing the attachment point against a range of views rather than a point estimate, and being explicit with reinsurers about where the uncertainty sits. Markets price transparency better than they price optimism. That principle carries into how a submission is built and presented — see placing facultative risk in a tight market.

Our analytics desk runs independent views for cedents ahead of renewal. Get in touch if that would be useful on your book.