Catastrophe Model Output Is Not a Reinsurance Submission

A catastrophe model can produce a number with an unnerving amount of authority. It arrives as a clean exceedance curve, a PML at two return periods, and a chart tidy enough for a capacity meeting. Nobody feels the need to ask where it came from. That is the moment to slow down.

The catastrophe model is not a picture of the portfolio. It is a disciplined statement of what we do not know about the portfolio.

It substitutes a simulated history for one nobody has. It assumes things about hazard, vulnerability, and terms. Then it gives them a form we can challenge. The problem starts when we let the curve hide them. The distinction sharpens at a Canadian property renewal. Most treaties incept on 1 January.

The broker circulates the submission in late September. What leaves the cedent isn’t a model printout with a logo.

It is a data package.

Sums insured by CRESTA zone or forward sortation area. Occupancy, construction class, year built, deductible profile. Ten years of premium and loss history. Large-loss listings. Development triangles. Model output run on our own licence.

The package is an argument about actual exposure.

The curve is one exhibit in the argument.

A curve is not a portfolio, and a zone is not an address

CRESTA zones matter because loss comes from an event, not an average policy. Building, contents, business interruption, living expense: all separate, because one wildfire doesn’t make them behave alike. Insured value in the wrong geography isn’t conservative data. It is data that lost its own question. The easy way to make that mistake is treating an address as a location. A rooftop match and a postal-code centroid are not equivalent. The second puts a property on the wrong side of a floodplain.

Construction class and year built aren’t decoration either.

They are arguments of the vulnerability function.

A submission reporting the share of insured value with known construction tells the reinsurer where we are inferring rather than calculating.

I would make the data-quality page non-negotiable.

Nobody enjoys producing it.

That is partly why it works.

What the model actually says, and what it declines to say

The event set is the honest part of the apparatus. No insurer holds centuries of loss experience at a stable exposure base. Building stock, replacement cost, code vintage: all move too fast. So the model builds a catalogue of physically plausible events with annual rates, maps intensity to damage through vulnerability functions, and applies deductibles, limits, and reinsurance in a financial module.

Those modules shouldn’t collapse into one result.

Hazard comes from long geophysical records.

Vulnerability comes from claims and testing.

Financial terms are contractual and exactly knowable.

They carry different uncertainties.

If we cannot say whether a change in PML came from the model version, the exposure file, the view of demand surge, or a revised treaty assumption, we do not have a model governance problem.

We have a decision-right problem.

The labels also need their nouns back.

OEP asks about the largest single occurrence.

AEP asks about the annual total, several events included. AEP sits above OEP everywhere, and the gap is multi-event risk.

PML is not a self-explanatory loss number.

It is a point on a stated curve at a stated period. It stays incomplete until it says OEP or AEP, gross or net, expense in or out.

That is not pedantry.

Two firms can quote PMLs twice apart and both be right.

Different curves. Different periods. Different modules. Different netting.

The curve has not failed.

The conversation has.

The Canadian market makes this visible at renewal.

The submission presents OEP at two return periods, gross and net of quota share. Reinsurers rerun the portfolio on their own view, amplification and demand surge included. They routinely land a third away from our result. That range isn’t an embarrassment to smooth out beforehand.

It is the meeting.

“The model says the tower is sufficient” is a weak sentence if nobody can say which model, which data-completeness measure, which curve, and which treaty basis it contains.

The treaty is a different machine altogether

A reinsurance programme does not transfer a PML.

It transfers losses under a contract.

That sounds obvious until somebody uses a model output as though it carried its own attachment point, occurrence definition, and payment order.

It does not.

A quota share cedes a fixed share of premium and loss from the first dollar.

A surplus treaty cedes a variable share above a retained line.

A catastrophe excess layer responds only after the retained loss reaches attachment.

Those are different mechanisms.

The exposure report lets each one see the loss it covers.

Inuring language turns a clean model result into a wrong forecast.

Written net of quota share, a catastrophe programme sees the loss after the cession.

Written gross, it sees the full loss.

The order is a treaty clause, not a spreadsheet formatting choice.

A module applying the wrong order hasn’t made a modelling error.

It has modelled a different contract.

The same holds for the event definition.

Most property treaties use a seventy-two-hour clause for hail and wind.

The cedent normally elects the start of the window.

Two cells eighty hours apart become two occurrences.

Two retentions. Two sets of available limits.

A weather map cannot decide that.

The wording does.

Nor is a layer a permanent reservoir.

A reinstatement restores the limit after erosion, for premium proportional to the amount and the time left.

A cash-call clause makes a reinsurer advance funds before the underlying claims are paid.

Formal notification begins at half the retention, and the window can be a fortnight.

The loss curve helps set the purchase.

The treaty decides what happens when the roof claims arrive.

Which is why I prefer “submission” to “model pack.”

The first implies evidence, assumptions, and a counterparty deciding whether the contract makes sense.

The second suggests output for a machine to accept or reject.

The machine will not pay a claim.

Canada does not allow one generic answer here

Earthquake and flood are a useful corrective to imported catastrophe language.

Under the Minimum Capital Test we hold capital for earthquake from a modelled loss at a prescribed return period, with credit for reinsurance and a maintained earthquake reserve.

The earthquake guideline makes exposure data, model use, and recoverability prudential questions.

Flood has no equivalent prescribed Canadian charge.

That does not make flood less real.

It means we manage two correlated perils under different capital disciplines.

A US-style presentation treating the capital framework as background misses part of the Canadian decision.

Reinsurance credit is separate again.

Cessions to an unregistered reinsurer need collateral before they get capital credit.

Guideline B-3 adds a board-approved plan, counterparty diligence, and a limit on ceding substantially all risk.

So a treaty can be valid, placed, attractive, and still not deliver the balance-sheet relief in the deck.

There is a quiet operational consequence.

Model output, exposure store, treaty administration, capital calculation: they can’t merely reconcile at year-end.

It has to preserve the basis of each result.

Geocoding level. Coverage split. Model version. Return-period curve. Gross or net. Attachment. Exhaustion. Reinstatement. Collateral.

Those are not fields to flatten into “cat exposure.”

They are the conditions under which capacity exists.

Put the exposure view beside the treaty view and leave them there

The board asks for the PML, the programme limit, the retained exposure.

We make the answer resistant to false neatness.

Ask whether management gets an exposure view and a treaty view separately.

The exposure view names data gaps, concentration by peril and geography, coverage components, model version, and the range between our view and theirs.

The treaty view names quota-share assumptions, attachment, limit, hours clause, reinstatement, notice threshold, counterparty credit, and the basis for capital credit.

Put the two beside one another.

Don’t fuse them into a traffic-light score.

That separation does not slow the decision down.

It makes the disagreement actionable.

Underwriting can stop a new address breaching an accumulation zone.

A data team can repair a construction-class gap.

Reinsurance can test another attachment, or find facultative capacity.

Finance can test whether collateral supports the credit.

A single PML tells none of them whose job it is.

The position nobody wants to defend is the right one.

Uncertainty belongs in the submission.

The strongest reinsurance programme is not the one with the most confident loss curve.

It is the one whose exposure, assumptions, and contract stay visible when the curve is wrong.