The CAT Modeling Gap
The CAT Modeling Gap in insurance isn’t that climate losses are rising. Everyone knows losses are rising. The gap is that loss is quietly migrating toward the perils we told ourselves we had handled.
Where the Loss Went While We Were Presenting
For years, we’ve talked about “climate risk” and “physical risk” as if it were a story about bigger hurricanes, larger wildfires, and more severe floods. Everything is bigger, stronger, more damaging. Those are the perils the insurance industry built its machinery around, and not by accident. They’re visible, they make headlines, they fill up social media posts, and they justify a whole vendor ecosystems and conference circuits. A peril that fills a keynote is a peril that gets funded. So the analytical attention followed the money and the spotlight, which mostly pointed at the same place.
The problem is that the spotlight is a lousy way to find what’s actually moving. We poured decades of capital, talent, and compute into understanding one slice of the risk landscape as deeply as physically possible, and while we were heads-down understanding that one slice, the rest of the map kept changing without us.
Every Core Competency Buys a Blind Spot
The industry has built genuinely extraordinary tools: hazard models, vulnerability curves, exposure databases, stochastic event catalogs, and reinsurance structures that price events that haven’t happened yet. The hazard-exposure-vulnerability concept is so institutionalized that it now has open reference implementations written in code.
But you don’t get that good at one thing for free. Every optimization buys a blind spot. The sharper your model of one peril, the easier it becomes to underweight everything just outside it, not consciously, just structurally.
So the industry’s best analytics cluster is around the risks it already understands, while a growing share of loss emerges from the places that got less of everything. Not no attention. Less. And in a capital-constrained business, “less” is where you get hurt.
We Answered the Wrong Question: On Time, Under Budget
Catastrophe models were built to answer one question: how bad can this get? Stress earnings, stress capital, stress the reinsurance tower under extreme-but-plausible scenarios. Worthy question. But it smuggles in an assumption that the landscape itself holds still, and the only thing worth worrying about is the tail.
That’s a different question from what happens when the whole landscape starts moving? The first assumes the future stays recognizable. The second assumes the map is the thing in motion.
For most of the modern insurance era, the gap between those two questions didn’t cost anyone money. The background climate was stable enough that history plus expert judgment carried the weight. That era is over. The IPCC’s Sixth Assessment Report spends most of its energy on shifts in frequency, intensity, and compounding, the baseline moving, not just the extremes stretching. We built the entire apparatus to answer the question that’s now the less important one.
World-Class at the Perils That are Bankable
Look at where the industry concentrated its firepower: major hurricanes, large earthquakes, catastrophic wildfires. The events that can take out a season’s earnings in an afternoon. Also, and very conveniently, the events where one more increment of model skill converts most directly into money. The incentives all pointed the same way, so the industry got exceptionally good at pricing them.
In some markets, too good, when a risk is deeply understood, capacity floods in, competition climbs, margins compress. The official history of cat modeling, as told by the big vendors’ corporate lore and the rating agencies, is a story of ever-sharper tools aimed at a stable set of high-headline perils. It’s a real achievement. It’s also a portrait of an industry optimizing toward the spotlight while the stage rotated.
Everyone got their gold stars, achievement awards, and atta Boys.
Volatility Doesn’t Read the Org Chart
Volatility doesn’t just make the perils we model bigger. It redistributes uncertainty across the whole portfolio, and the places where uncertainty is climbing fastest are not the places where we built the deepest tools. That’s the gap: the boundary between the well-modeled, high-confidence perils and the swelling share of loss coming from everything else.
The reinsurers have been ringing this bell for years. Swiss Re’s 2025 sigma report found secondary perils drove the majority of global insured losses for the third straight year. Munich Re keeps flagging the same thing.
For the third consecutive year is not a tail event. It’s a trend the industry keeps filing under “noise” because the alternative is admitting the category names are wrong.
Will the “Secondary” Perils Please Stand Up
Watch what’s actually showing up in the loss runs:
Severe convective storm (SCS)
Localized (pluvial) flooding
Infrastructure and systemic failure
Compound events and extreme precipitation clusters
We literally call these “secondary perils,” a naming convention that tells you exactly where the institutional attention went. They were cast as supporting actors around the primary signal. Now they’re carrying the picture. In recent years, SCS losses across North America and Europe have rivaled or surpassed headline hurricane seasons, and Swiss Re now calls them structurally elevated rather than episodic.
The label “secondary” isn’t a description anymore. It’s a legacy of where we were looking when the money moved.
If this makes you uncomfortable, good. Not because there are no models for these perils. It’s that decades of capital structure, governance, and institutional conviction were built around a different loss landscape. Having a model and having conviction are different things, especially when the empirical pattern is shifting under your feet, and your committees are still calibrated to the old one.
The Models Are Fine. That’s the Whole Problem.
Let’s be precise: the CAT models didn’t stop working. For the perils they were built for, most do exactly what they were designed to do, inside their assumptions and their data. The consequential dynamic is subtler and worse, loss is migrating toward exposures where uncertainty was always higher and the analytical investment was always lower.
Hurricane Otis, October 2023, is the cleanest example we have. Early on, forecasters had it peaking as a weak tropical storm and staying offshore; some guidance had it dissipating. Instead it ran one of the most extreme rapid-intensification events of the satellite era — +90 knots in 21 hours — and hit near Acapulco as a Category 5 with 165 mph winds, an intensity the region had never recorded.
No operational model caught the magnitude. NHC’s own word was “nightmare scenario.” Damages landed at $12–16 billion against initial industry insured-loss estimates of $2.5–6 billion. That gap isn’t only low Mexican insurance penetration, it’s the absence of a historical analog. Acapulco had never taken a Cat 5, so the stochastic catalogs had thin-to-no representation of the event, and the confidence interval around any estimate was, functionally, infinite.
The model wasn’t wrong. The problem had no precedent, and a catalog built on precedent has nothing to say about it.
Picture spending thirty years building the most accurate map of a coastline ever drawn while the population quietly moves inland. The map is flawless. It’s also answering a question fewer and fewer people are asking. That’s the position a lot of the ecosystem is drifting into: immaculate cartography of one risk category while the aggregate loss burden relocates to the parts of the map we left in low resolution.
The Market Is Already Voting With Its Capacity
Once you see the gap, the market’s behavior reads like a sinner in a confessional:
Carriers retrenching from whole geographies and lines
Reinsurance repricing and restructuring around the mid-sized perils
A scramble for parcel-level, asset-level, geospatial intelligence
Regulators asking harder questions about how risk is measured and justified
On the regulatory side, SFDR, CSRD, and climate stress testing out of the NGFS are all forcing institutions to show their assumptions at a granularity they’ve never had to before. And the market is answering with a wave of climate and location-intelligence platforms, GeoAI, Earth observation, parcel-level data, all competing on one thing: how well they close the distance between the model, the observation, and the underwriting decision.
That competition is the tell. It’s the industry pricing the gap in near real time.
Insurance Was Never a Climate Business (Sorry)
Strip the climate language, and here’s the machine underneath: insurance is a capital allocation business. Climate is one of several forces deciding where capital can flow, at what price, and on what terms. That’s it.
Capital has never minded risk; absorbing risk is the entire product. What capital cannot stand is unbounded uncertainty, especially uncertainty that seems to be growing. When loss moves into exposures that are harder to observe, explain, and model, capital does what capital always does: it gets selective. Capacity tightens, price moves, coverage shrinks, and the boardroom conversation flips from growth to defensibility.
None of this is a moral failure. It’s the system behaving exactly as designed, and that’s the point. You don’t change the outcome by appealing to it. You change it by changing what capital can see.
The standard-setters figured solved thi puzzle, to an extent. TCFD and the NGFS have spent years framing climate as a capital and financial-stability problem, not an environmental one. That framing is now bleeding straight into how insurers, banks, and asset owners price physical risk, which means the winners won’t be the ones with the best climate story. They’ll be the ones who make the uncertainty legible to capital faster than the competition.
Nobody Gets a Premium for Table Stakes
So here’s the implication, and it’s not the comfortable one. The next decade won’t belong to whoever has the “best” model in isolation. Model skill is becoming table stakes, and table stakes don’t earn a premium. It’ll belong to whoever closes the gap between model and reality, because that’s the gap capital is now paying to close.
The questions are shifting from probabilistic to concrete:
Who can continuously observe conditions on the ground, at scale?
Who can verify adaptation instead of assuming it?
Who can explain why this asset is materially different from the one next door?
Who can demonstrate a risk profile actually changed, rather than asserting it?
Those are evidence questions, and they live at the intersection of GeoAI, Earth observation, climate science, and financial-risk architecture. The plumbing already exists, the Open Geospatial Consortium standards are the interoperability layer that makes cross-platform data fusion actually work, and open initiatives like the Global Earthquake Model prove the transparent, standards-based version of this is buildable, not theoretical.
If loss is migrating toward the parts of the portfolio where uncertainty is highest, then the durable edge isn’t predicting risk. Everyone predicts risk. The edge is reducing uncertainty faster than everyone else — fusing models, measurements, and market behavior into something regulators and capital providers are willing to underwrite their reputations on.
The industry spent thirty years building better maps. That was the easy part, and it’s done. The hard part — the part nobody gets to skip now — is navigating a landscape that’s moving underneath the map while keeping capital in the system instead of watching it walk to the sidelines.
The firms that treat that as a modeling problem will keep producing beautiful maps of the wrong coastline. The ones that treat it as an evidence problem will own the next cycle.
The gap is already open. The only question is who’s building across it, and who’s still admiring their “perfect” historical model.
Call to Action
So I’ll ask the people who live in this: where are you actually seeing the gap widen first, convective storm, pluvial flood, systemic/infrastructure failure, or something the rest of us aren’t pricing yet? I want to hear where your loss runs are surprising you.
Sources
Swiss Re Institute, sigma 2025 and natural catastrophe research; Munich Re natural disaster review; IPCC Sixth Assessment Report; NOAA (Hurricane Otis); Oasis Loss Modelling Framework; EU SFDR and CSRD; NGFS; TCFD; Open Geospatial Consortium; Global Earthquake Model.
Clairvoyint helps expert teams turn analytical methodology into reusable, evidence-grounded agents.If your team relies on complex analysis that depends on expert judgment, fragmented tools, and defensible outputs, let’s talk.