--- title: the model learns from a past that is leaving canonical_url: https://ensurance.app/guide/the-model-learns-from-a-past-that-is-leaving markdown_url: https://ensurance.app/guide/the-model-learns-from-a-past-that-is-leaving.md subtitle: "catastrophe models are calibrated to history; the condition of the ground under the exposure changes every year, and models often hold it still" category: nature-finance --- # the model learns from a past that is leaving *catastrophe models are calibrated to history; the condition of the ground under the exposure changes every year, and models often hold it still* A forest above a town can carry more fuel this summer than it did five summers ago. A marsh in front of a coastal street can be a little narrower. A floodplain upstream of an industrial park can be partly filled for a warehouse. None of that may be in the catastrophe model the reinsurer will use at the next renewal, or it is in there as a land-cover class set years ago. The forest does this work whether or not anyone pays. That is the shortest answer to the search phrase **catastrophe model limitations**: a cat model learns from a past that is leaving, and the condition of the ground under the exposure is one of the inputs that moves first. The model is still the best tool the market has. This post covers which gaps a reinsurer already manages, and the one that sits in a forest, a marsh, a floodplain or a reef. :::johnson **the condition of the ground is the input that moves first.** Hazard catalogs change slowly and exposure is counted every renewal. The fuel load, the wetland edge and the reef crest can change within a season or a few years, and models often hold them still. [See how a with-and-without run works →](/guide/how-an-intervention-lowers-the-loss?from=guide) ::: ## how a catastrophe model works A **catastrophe model** has four parts. A stochastic event set holds thousands of simulated storms, quakes, floods or fires, each with an annual rate. A hazard module turns each event into local intensity: wind speed, water depth, ground shaking, flame length. A vulnerability module turns intensity into damage for each building type, using damage functions fit to engineering tests and past claims. A financial module applies deductibles, limits and reinsurance terms. Run every event against the portfolio and the output is an **average annual loss** (AAL) and an exceedance probability curve: the chance that losses in a year pass any given amount. Treaty pricing, capital and retention all lean on those two outputs. The primary carrier's version of this argument is in [why better cat models aren't fixing your loss ratios](/guide/why-better-cat-models-arent-fixing-your-loss-ratios?from=guide). This post stays on the reinsurer's side: model risk. ## the limits a reinsurer actually manages Reinsurers manage model risk by blending vendor views, loading for non-modeled loss and uncertainty, and applying their own view of risk. Those adjustments act on the model's output. None of them changes the condition of the land. ### non-stationarity Every event set and hazard module is calibrated to history, and the damage curves to past claims: decades of storm tracks, river gauges and fire perimeters. That assumes the past describes the next ten years. When it does not, the hazard is **non-stationary**. Then a model fit to the record describes a climate and land surface that have moved on. Climate-conditioned catalogs help, but each is still a judgement about how far to lean away from the record. ### exposure growth versus hazard change The loss trend is real and it is large. The Swiss Re Institute puts the trend in insured nat-cat losses at 5 to 7% a year in real terms, implying about USD 148bn for 2026 (sigma 1/2026). Verisk's modeled global insured AAL was USD 59bn in 2012, USD 152bn in 2025 and USD 171bn in 2026. The honest reading is that most of that growth is exposure. Swiss Re attributes loss growth mainly to more value in harm's way: rising property values, urbanization and higher rebuild costs, with climate change amplifying some perils in some regions (Swiss Re, 2025). Verisk's figures are modeled, so growth includes exposure growth, and they are not directly comparable to Swiss Re's real-terms trend. A reinsurer has to separate three things the headline blends: more buildings, a better or different model, and a changing hazard. Only the third is non-stationarity, and it is the hardest to see in a single year's numbers. Land condition is a fourth, input-side change, and it is the one this post is about. ### secondary perils and local resolution Peak perils have long records and big footprints. Secondary perils turn on local detail: a narrow hail swath, one ridge's fuel, one clogged culvert. Verisk's 2026 breakdown puts severe convective storm at 40% of modeled global insured AAL, ahead of tropical cyclone at 27%, earthquake at 10%, flood at 7% and wildfire at 6%. Swiss Re counted a record 92% of 2025 insured losses from secondary perils, and the Los Angeles wildfires, about USD 40bn, were the largest insured wildfire loss on sigma records (sigma 1/2026). Wildfire and flood are partly ground-set (strongest for wildfire in dry forest and for frequent floods). Severe convective storm, the largest slice at 40%, is weather: hail and tornado do not depend on the condition of the land. A model that represents a hillside as one land-cover class, or a coastline as a fixed roughness value, cannot distinguish a thinned forest from an overgrown one or a healthy marsh from an eroding one. ### regulators adding stress tests Supervisors are adding scrutiny around the vendor model. The Bank of England's Prudential Regulation Authority ran its dynamic general insurance stress test, DyGIST, live in May 2026, with results due by year-end. California now allows wildfire catastrophe models in rate filings; its Department of Insurance finished its first model review in July 2025. ### the strongest counterargument A fair modeler would answer that vendors update constantly: new event sets after major seasons, revised vulnerability curves, climate-conditioned catalogs, finer grids. For peak perils the models are very good, and a well-run cat model prices a hurricane book better than anything else available. That is true. Updating fixes the event set and the damage curves. It does not, on its own, change how the model treats the condition of the land between the hazard and the building. ## the input models often hold still Here is the gap this series is about. In the standard portfolio view used for treaty pricing, the condition of the ground is usually a static or slowly refreshed layer. Fuel load is often a vegetation class. Wetland extent is a land-cover layer refreshed every few years. Floodplain connection usually follows whatever terrain model the vendor licensed. Reef crest health is often a single roughness value. Some wildfire and flood models do take fuel or terrain updates, and some give credit for mitigation; the gap is that the layer rarely moves year to year with what a land manager actually does. This is the common structure as we understand it, not an audit of any vendor; your vendor's documentation is the check. Each of those can change within a season or a few years, faster than a hazard catalog. Fuel builds between fires and drops after thinning or prescribed burns. Marshes erode or accrete. Floodplains get filled or reconnected. Reef crests bleach, break or recover. And each can be read: by remote sensing, by field plots, by the same parcel data a land manager already keeps. Beck et al. (2018) compared reefs present against absent and found that without them, annual expected flood damages would double and frequent-storm costs triple. That is not a measure of crest decline, but it shows how much of the risk a reef carries. The hazard is calibrated to history. The exposure is counted at renewal. The condition of the ground moves in between. Fuel reduction and wetland restoration are funded today, mostly from public budgets and grants, but rarely by the people whose loss they lower. ### what with-and-without work already shows The field has two modeled runs and one observed comparison. Using industry catastrophe models, Narayan et al. (2017) found Ocean County, New Jersey, salt marshes lowered average annual flood losses 16%; The Nature Conservancy and WTW (2021) found Northern Sierra ecological forestry lowered modeled AAL 40 to 60%. The Swiss Re Institute (2025) found about half the National Flood Insurance Program (NFIP) claim frequency in Florida's best-protected coastal-habitat areas in frequent storms. The modeled runs show the method can be run on industry tools; the claims comparison is observed, not modeled, and does not prove one project lowers one loss. The method and its limits: [how an intervention lowers the loss](/guide/how-an-intervention-lowers-the-loss?from=guide). ## reading the condition beside the price The first job is to read condition the same way every year for the same place. RealValue, the natural capital accounting behind ensurance, reads each parcel's ecosystem service value times its condition: what the land supplies, and how well it is holding up. It sits beside a modeled loss, not inside it. A condition reading is a reason to run the model again, not a result. That discipline is set out in [grade beside the price](/guide/grade-beside-the-price?from=guide). **Ensurance** is the second job: funding the condition so the input moves in the right direction before the season. A certificate tied to a named place, such as a forest above a town or a marsh in front of a street, carries the money to the work: thinning, a reconnected floodplain, a restored marsh edge, so the funder can see what was paid for and how the ground is doing, which a grant letter does not show. The forest or the marsh does the work whether or not anyone pays. Ensurance is how the people it protects pay for it. Browse named places on [specific ensurance](/specific?from=guide). ## what this does not touch Funding the ground moves the frequency and severity of mostly local, mostly secondary-peril losses: wildfire at the edge of a town, flood behind a marsh, wave and surge damage from frequent, lower-severity storms behind a reef. Surge in a major hurricane is peak peril and is untouched. It does not change a Category 5 at landfall, and it does nothing for earthquake. The peak-peril tail, where most reinsurance capital is held, is untouched. No reinsurer has booked an ensurance-funded loss reduction. The evidence above is the field's. What ensurance adds is the funding and the condition reading, so that the next with-and-without run has a real project to measure. What insurers and reinsurers can fund today is on [ensurance for insurers](/solutions/insurers?from=guide&topic=reinsurance). ## frequently asked questions ### what are the limitations of catastrophe models? They are calibrated to historical hazard, so they lag a changing climate and land surface (non-stationarity). Secondary-peril losses depend more on local detail than peak-peril losses. And the condition of the ground, such as fuel load, wetland extent, floodplain connection and reef health, is often a static or slowly refreshed layer, even though it can change within a season or a few years. ### is the rise in modeled catastrophe loss mostly climate change? Mostly not. Swiss Re attributes most loss growth to exposure: more value, more urbanization, higher rebuild costs, with climate change amplifying some perils and regions. Hazard change is real but is the hardest part to isolate from exposure growth and model updates. ### can land condition be put into a cat model? Yes, through a with-and-without run: the same events against the same buildings, once with the land as it is and once with the work done. Narayan et al. (2017) and TNC and WTW (2021) both did it. ### does this help with peak-peril reinsurance capital? No. It moves mostly local, secondary-peril losses. It does not touch hurricane landfall at the tail or earthquake. ### has a reinsurer booked a loss reduction from ensurance? No. The results cited here come from the field, not from ensurance. ## read next The model is not the only thing that has moved. Capital has too, and there is more of it than ever: [capital is not the shortage](/guide/capital-is-not-the-shortage?from=guide). ## sources [Swiss Re, sigma 1/2026: natural catastrophes in 2025, 2026](https://www.swissre.com/institute/research/sigma-research/sigma-2026-01-natcat-2025-wildfire-storm-risk.html) — Insured loss trend 5–7% a year in real terms, about USD 148bn implied for 2026; 92% of 2025 insured losses from secondary perils; Los Angeles wildfires about USD 40bn, largest insured wildfire loss on sigma records [Verisk, global modeled insured catastrophe losses, 2026](https://www.verisk.com/company/newsroom/global-insured-catastrophe-losses-now-expected-to-average-$171-billion-annually-verisk-finds/) — Modeled global insured AAL USD 171bn (2026), USD 152bn (2025), USD 59bn (2012); peril shares SCS 40%, TC 27%, EQ 10%, flood 7%, wildfire 6% [Risk & Insurance, Swiss Re on 2025 nat-cat losses, 2025](https://riskandinsurance.com/natural-catastrophe-losses-set-to-reach-145-billion-in-2025-with-1-in-10-risk-of-300-billion-peak-year/) — Loss growth mostly exposure; climate change amplifies some perils and regions [Bank of England, insurance stress testing, 2026](https://www.bankofengland.co.uk/stress-testing/insurers) — DyGIST 2026 dynamic general insurance stress test, live May 2026, results due by year-end [California Department of Insurance, 2024 Annual Report of the Commissioner, 2025](https://www.insurance.ca.gov/0400-news/0200-studies-reports/0700-commissioner-report/upload/2024-Annual-Report-of-the-Commissioner.pdf) — Wildfire catastrophe models allowed in rates; first model review finished July 2025 [Narayan et al., Scientific Reports 7:9463, 2017](https://doi.org/10.1038/s41598-017-09269-z) — Ocean County salt marshes cut average annual flood losses 16%, modeled with industry cat models [Beck et al., Nature Communications 9:2186, 2018](https://doi.org/10.1038/s41467-018-04568-z) — Without reefs, annual expected flood damages would double and frequent-storm costs triple [The Nature Conservancy and WTW, wildfire resilience insurance, 2021](https://wtwco.com/en-US/Insights/2021/07/wildfire-resilience-insurance-quantifying-the-risk-reduction-of-ecological-forestry-with-insurance) — Northern Sierra ecological forestry lowered modeled AAL 40–60% [Swiss Re Institute, coastal and flood protection from natural habitats, 2025](https://www.swissre.com/institute/research/topics-and-risk-dialogues/climate-and-natural-catastrophe-risk/coastal-and-flood-protection-natural-habitats.html) — About half the NFIP claim frequency in the best-protected Florida coastal-habitat areas, frequent storms, 2009–2022 ## the series 1. [what reinsurance cannot cede](/guide/what-reinsurance-cannot-cede?from=guide) 2. [secondary perils are not secondary](/guide/secondary-perils-are-not-secondary?from=guide) 3. [the loss moved down the tower](/guide/the-loss-moved-down-the-tower?from=guide) 4. [the model learns from a past that is leaving](/guide/the-model-learns-from-a-past-that-is-leaving?from=guide) 5. [capital is not the shortage](/guide/capital-is-not-the-shortage?from=guide) 6. [what a reinsurer can fund before the season](/guide/what-a-reinsurer-can-fund-before-the-season?from=guide)