Can a parametric climate insurance product prove it paid?

1. The instruments are already live

In November 2025 the Bureau of Fisheries and Aquatic Resources became policyholder for the Philippines' first parametric cover for small-scale fishers, alongside the Philippine Crop Insurance Corporation, Rare and Willis, developed through the Ocean Risk and Resilience Action Alliance with funding from the Government of Canada and the UK Government's Blue Planet Fund. It reaches 14,200 fishers across 24 coastal municipalities, pays up to USD 100 per policy cycle, and is offered as a benefit of fisher registration. Separately, in July 2025 the Department of Agriculture announced that PCIC is readying a parametric typhoon pilot for rice farmers with the Philippine Space Agency and the Philippine Rice Research Institute, targeting computation of compensation within three to five days of a storm exiting.

This is real public and donor money, moving now, to households among the most climate-exposed on earth. The question we think is under-asked is not whether these instruments are a good idea. It is whether anyone outside the issuer can check that they did what they said.

2. Three failure points, all of them record problems

A parametric product changes behaviour only if the household believes it will pay. That belief can break in three distinct places, and the three are usually held in three different systems by three different organisations, which is why nobody sees them together.

None of these is a modelling problem. All three are record problems, which means all three are machine-checkable once somebody writes the rules down.

3. The twelve rules

Each rule exists because a documented failure exists. Two of them, R6 and R7, are written directly from the last-mile findings above. A rule set that passes everything is decoration, so every one of the twelve has a negative test case in the repository built specifically to trip it, including a deliberately brochure-grade product encoding that nothing could ever settle or dispute.

The payability rules are the ones worth running first, because a payability report is a to-do list rather than a post-mortem. An honest "we do not know whether this cheque was banked" is a conformant record, and so is a declared legitimate name mismatch such as a married name. Silence about either is what fails.

4. The 8 km/h that decided a season

Encode a product's trigger and payout tiers, then replay that logic against observed events. Below, two encoded products against five Philippine landfalls. The interesting row is the fourth.

Severe Tropical Storm Nalgae, locally Paeng, made landfall at Virac in Catanduanes on 29 October 2022. Reported ten-minute maximum sustained winds at landfall were 110 km/h. PAGASA's current scale puts the boundary of a typhoon at 118 km/h. Nalgae arrived eight kilometres per hour short of being a typhoon. It killed over a hundred people, affected more than two million, and the Department of Agriculture reported roughly 67,000 tonnes of agricultural losses worth about PHP 1.3 billion, the bulk of it in rice.

The four catastrophic super-typhoons in the set all trigger comfortably, which is exactly why a product looks fine when tested against the storms everyone remembers. A farmer who lost a field to Nalgae and received nothing did not experience an index limitation. They experienced a broken promise, and told their barangay so.

Why this matters more than the modelling debate

Erik Lichtenberg and Eva Iglesias, in a reconsideration of the basis-risk literature published in the Journal of Development Economics in 2022, argue that the standard account has been too quick to treat basis risk as the explanation for low uptake, and that contract design deserves more of the blame. We think that is right, and that it sharpens rather than blunts the case for publishing results like the one above. If the problem is partly design, then design is a decision somebody made, and decisions can be disclosed. A trigger set at 118 km/h rather than 90 km/h is a choice about which losses the product declines to cover. That choice may be entirely defensible on pricing grounds. What is not defensible is that it is currently invisible to the household, to the cooperative selling it, and often to the donor paying the premium.

Berber Kramer, Maria Porter and Solomon Wassie, in a field experiment in Ethiopia published in the Journal of Risk and Insurance in 2025, find that basis risk and what a neighbour was seen to receive move perceived fairness and willingness to pay. That is the mechanism this work is aimed at. Households in these schemes do not price a product. They judge whether it treated them fairly, and they judge it collectively, in a barangay, over years.

So the disclosure we think should be routine, for any parametric product before sale, is: the historical events over the covered geography where the index did not reach the trigger; of those, the ones with documented losses of the kind the product exists to cover; the margin by which each one missed, because 8 km/h reads very differently from 80 km/h; and the measuring authority and basis, so the number can be independently recomputed. None of this requires proprietary data. Everything in the Nalgae result comes from public storm reporting and public agricultural loss figures.

5. The payout that fails before the storm

Almost all attention in adaptation finance verification points backwards: did the payout arrive, did the index behave. The failure that actually decides whether a household gets paid happens before the event, in the register, and it is the only kind you can still fix.

A great deal of climate cover in the Philippines is not bought. It is conferred. Cover for register-listed subsistence farmers and fisherfolk is premium-subsidised, and in the flagship parametric scheme for small-scale fishers the premium is donor-paid and the cover is offered as a benefit of registration. Enrolment is frequently done in bulk by a municipal agriculture office working from the register. Two consequences follow, and most verification designs miss both. First, the household often does not know it is covered, so there is no promise for it to disbelieve. Second, every decision that determines whether money can reach that household has already been made by the time a storm forms, and it was made by a data-entry process rather than by an underwriter.

Because the fisher product ties cover to registration, the register is not administrative background. It is the enrolment mechanism, the eligibility test and the payment instruction, all in one artefact that nobody treats as safety-critical. Rules P1 to P4 run over a register and a payout file and return a list of households who could not currently be paid, with a reason attached to each: missing barangay, no georeference, no or dormant instrument, unexplained name mismatch.

There is also a practical reason to lead with payability rather than settlement, and it is about who will agree to be audited. A settlement audit applied to a public insurer's records produces an externally auditable count of entitlements that were owed and never paid, which is a finding an institution has every reason to refuse. A payability audit produces a to-do list that makes the municipal office and the policyholder look competent when they act on it. Nobody in the chain wants to discover after a typhoon that four hundred registered fishers had dead cash cards.

Nathaniel Jensen, Andrew Mude and Christopher Barrett, writing in Food Policy in 2018 on index insurance in northern Kenya, show how basis risk and spatiotemporal adverse selection shape demand in a setting where households are making real purchase decisions. Their framework is the right one for a market with buyers. Our argument is about what happens to it when the purchase decision is removed: under bulk enrolment, demand elasticity stops being the operative variable for whether a household benefits, and administrative data quality takes its place. And when those households do eventually face a real decision, on top-up cover or on renewal after a subsidy lapses, what they will price is their memory of whether the free layer ever worked.

6. Two products, one headline, different promises

Where an index is measured decides who gets paid, and it is the field most often left unstated. Below are the two encoded products compared field by field. Four fields agree, including the headline number a buyer would compare. Two do not, and those two decide everything.

A product settled on wind at the storm centre pays every household in the scheme when the eye crosses the coast, and pays nothing to a household two hundred kilometres away that lost its season to the same storm. A product settled over the insured parcel does close to the opposite. Neither behaviour is wrong; they are different products for different risks. The problem is that on any brochure, product register or comparison table these two appear identical.

The methodological point is the transferable part: this divergence is found by reading the two encodings against each other, with no event data involved. That matters because the empirical route is expensive. Quantifying spatial basis risk properly needs per-location wind fields over decades, which most schemes will never commission. A structural check needs an afternoon and a text file. So rule R3 is a disclosure rule rather than a measurement rule: it requires every index to state its measurement basis and its spatial basis, and does not ask whether the choice was a good one, only that it be visible.

The groundwork here is not ours. Ghada Elabed, Marc Bellemare, Michael Carter and Catherine Guirkinger, writing in Agricultural Economics in 2013, set out multiscale index insurance precisely to attack the gap between an index measured at one scale and a loss suffered at another. Their work improves the instrument. Ours makes the instrument's scale choice legible to everyone downstream of the actuary. A multiscale contract is a better product; a stated spatial basis is what lets a buyer tell that it is a better product.

Nothing here needs new legislation. A regulator, public insurer or donor could require four fields on any parametric product they approve, fund or list: the index and the authority whose published measurement settles it; the measurement basis, since a ten-minute sustained wind and a gust are different quantities; the spatial basis, stated as a location rule rather than implied; and the comparator at the boundary, because "winds of 118 km/h" does not say whether 118 pays.

7. Verification cuts both ways

The honest counter-argument to everything above is that verification is exactly what has already killed several agricultural insurance schemes. Vietnam's national agricultural insurance programme, with premium support that at points reached 90 per cent for poor households, has repeatedly failed to reach expected participation, and the documented obstacles were largely administrative: cumbersome procedures, slow appraisal, loss verification requiring multiple signatures before a claim could move, and insurers withdrawing after poor loss experience.

In other words, the country most often cited for weak uptake despite heavy subsidy is one where the burden of verification was placed on the claimant, and where that burden is part of why it did not work. The resolution is that the word covers two mechanisms with opposite effects.

Parametric products are a large improvement on the first kind and have so far done nothing about the second. A product can settle in three days without a single farm visit and still leave nobody able to say whether the payment reached the person it was owed to.

Michael King and Anuj Pratap Singh, writing in Food Policy in 2020 on farmers' valuation of agricultural insurance in Vietnam, examine how households actually value these products rather than assuming they value them at actuarial worth. If valuation is behavioural and contextual, then the experience of the process is part of the product, and a scheme that pays eventually after four signatures is not a slower version of one that pays in three days. It is a different good. Extend that and assurance is not overhead on the product; it is a feature of it.

Three design constraints follow, and we hold ourselves to them rather than advertising them. Nothing new is asked of the household: no app, no smartphone, no identity document, no form. Checking is machine-checkable rather than clerical, because verification became a burden in the first place when it was performed by people one claim at a time. And an honest unknown conforms, because a standard that punished institutions for admitting gaps would teach them to hide gaps.

8. What a certification standard owes its own evidence

We wrote a rule saying that an encoding without stated provenance is an assertion rather than a certification. Then we assembled a small set of historical typhoon landfalls and sourced the wind speeds from whatever was reachable: one PAGASA primary document, two news reports of PAGASA bulletins, one figure from a different meteorological agency, and one from Wikipedia.

Every number was accurate. We checked them. That is not the point. A repository that says overstating your own evidence is worthless cannot cite an encyclopedia for a load-bearing physical measurement and describe the result as sourced. The inconsistency was ours, it was found by adversarial review before publication rather than after, and the fix was not to hide it. The event file now carries the measuring agency and a primary-or-secondary marker on every row, and the build report enumerates PAGASA's public archive to establish why four of five events rest on secondary sourcing: it publishes per-storm summaries for 2018 only.

The mechanism doing most of the work is small: a mandatory enumerated confidence field on every product encoding, which a validation rule fails if omitted.

Because the level is machine-readable, a downstream user can filter: someone comparing products for a regulator can require official encodings and ignore everything else, without having to trust our judgement. That is the difference between a caveat and a control.

Nathaniel Jensen, Christopher Barrett and Andrew Mude, in the American Journal of Agricultural Economics in 2016, measured the quality of an index insurance product rather than assuming it. The obvious extension, and the one we would like someone to perform on us, is that assurance tooling deserves the same treatment. A validator is a product with a false-negative rate and a coverage profile. Ensuring every rule has a case that trips it establishes only that each rule is live, not that twelve are sufficient. We do not know what fraction of real-world payout failures they would catch, because we have not yet run them over a real ledger.

This is a competitive argument rather than a confession. In assurance work, declared limits are the product: what makes a validator worth adopting by a regulator, donor or public insurer is a legible account of what it does not do. We hold the same position in adjacent work. In our industrial data crosswalks the headline finding is that four of seven standards examined cannot reject a mis-mapping at all, which means the reasoner check most projects rely on could never have failed. A check that cannot fail is not evidence, and the useful contribution is often to say so out loud.