Vendor-to-vendor divergence in physical climate risk scores - materially different hazard assessments for the same asset - has become a recurring theme in risk practitioner coverage and supervisory climate reviews, and it deserves to be treated as the central validation problem of climate analytics rather than a footnote.
Physical climate risk vendor divergence is easy to demonstrate and uncomfortable to sit with: take one building, run it through two reputable climate analytics providers, and you will frequently receive risk assessments different enough to change the underwriting, lending or investment decision. Same asset, same peril, same time horizon - different answer.
For a profession that grew up validating models against observed loss experience, this is a new kind of problem. Neither vendor is "wrong" in a testable way, because the quantity in dispute - hazard frequency and intensity decades ahead under assumed emissions pathways - is not observable now. The divergence is not noise around a knowable truth; it is a disagreement between defensible modelling choices.
Where the divergence actually comes from
A physical climate risk score is a pipeline, and each stage contributes spread.
Climate model ensembles and downscaling. Vendors choose different global climate models, different ensemble weightings and different downscaling methods to get from coarse climate grids to asset-level resolution. Downscaling choices alone can flip a flood assessment: the same river, modelled at different resolutions with different digital elevation data, floods different postcodes.
Hazard modelling. Translating climate variables into peril intensity - rainfall into flood depth, wind fields into gust footprints, drought indices into subsidence and wildfire propensity - involves vendor-proprietary models with genuinely different structures.
Exposure and vulnerability. What does the vendor know about the building? Construction type, elevation, first-floor height, defences. Vulnerability curves - damage as a function of intensity - differ by provider and are frequently the largest single source of financial divergence.
Financial translation. Deductibles, indemnity assumptions, business interruption treatment. Two vendors agreeing on hazard can still disagree substantially on modelled loss.
The compounding of four stages of defensible-but-different choices is how "flood risk: high" and "flood risk: moderate" both arrive on your desk carrying the same address.
Why this is a validation problem, not a procurement problem
The instinctive response is to pick the "best" vendor and standardise. That response misreads the situation. Where the target is unobservable, vendor selection is not validation - it is delegation. Supervisory climate reviews on both sides of the Atlantic have pressed the same point: firms are expected to understand the models they rely on, including third-party ones, and increased actuarial involvement in climate work has made this the profession's problem to solve.
A defensible governance position looks like this:
Run more than one vendor on material exposure. Not everywhere - on the segments where the answer changes a decision. The divergence between vendors is itself information: it is a free, market-provided estimate of model uncertainty. A portfolio decision that is robust to the inter-vendor spread is defensible; one that flips between vendors is resting on a modelling choice you did not make and cannot explain.
Decompose before you compare. Vendor A vs Vendor B at the score level tells you nothing actionable. Decompose into hazard, vulnerability and financial stages, and the comparison becomes diagnostic: if hazard agrees and loss diverges, the argument is about vulnerability curves, which you can interrogate against engineering literature and claims experience.
Validate the observable components. Present-day hazard maps can be tested against observed events. Vulnerability curves can be tested against claims from past floods and storms. The projected component cannot be backtested - but it can be bounded, benchmarked across vendors and stress-tested for sensitivity to scenario and downscaling choices. Separate what can be validated from what can only be governed, and be explicit about which is which.
Document the choice as a modelling assumption. Whichever vendor's view enters pricing or capital, record it the way you would record any material assumption: alternatives considered, rationale, sensitivity, review trigger.
The pricing and fairness connection
Divergence has a consequence beyond the balance sheet. Where climate-adjusted scores flow into territorial rating, a customer's premium can depend on which vendor their insurer happens to license - a differential the insurer cannot substantiate if challenged. I have written separately about geographic ratemaking bias under climate risk; the fair-pricing argument there depends directly on this article's point: the projected component of a climate-adjusted rate rests on contested modelling choices, and should be defended - and bounded - accordingly.
What good looks like
The firms handling this well share a posture: they treat climate vendors the way a mature market treats catastrophe models - as competing scientific opinions to be understood, blended and challenged, not as utilities to be consumed. Multi-model thinking took the cat modelling world two decades and several unpleasant surprises to learn. Climate analytics can inherit the lesson without repeating the surprises.
Key Takeaways
- Physical climate risk scores for the same asset routinely diverge between vendors because hazard, vulnerability and financial modelling choices compound.
- Divergence is not resolvable by vendor selection: where the target is unobservable, choosing one vendor is delegation, not validation.
- Run multiple vendors on material segments and treat inter-vendor spread as a market-provided estimate of model uncertainty.
- Decompose comparisons into hazard, vulnerability and financial stages; validate observable components against events and claims, and govern the rest as assumptions.
- Vendor divergence flows into pricing: climate-adjusted territorial factors carry a differential that must be bounded and documented to be defensible.
Frequently Asked Questions
Why do climate risk vendors give different scores for the same property? Because a score compounds choices at four stages - climate model ensemble and downscaling, hazard modelling, exposure and vulnerability assumptions, and financial translation. Each stage admits multiple defensible approaches, and small differences compound into materially different asset-level assessments.
How should firms validate third-party climate risk models? Validate the observable parts - present-day hazard against event footprints, vulnerability curves against claims experience - and govern the unobservable projected parts: benchmark across vendors, quantify sensitivity to scenario and methodology choices, and document the selected view as a material modelling assumption with review triggers.
Is using a single climate data vendor acceptable? For immaterial exposure, often yes. For portfolios where the climate assessment changes underwriting, pricing or capital decisions, supervisory expectations increasingly imply understanding model uncertainty - and the practical route to that is running a second vendor on material segments and demonstrating decisions are robust to the spread.