How Actuaries Price the “Unpriceable:” Catastrophe Modeling for Climate Risk
Inside the science, math, and judgment calls behind putting a number on the next hurricane, wildfire, or flood — and why that number is getting harder to trust.
The Problem That Broke Actuarial Science
On August 24, 1992, Hurricane Andrew made landfall just south of Miami as a Category 5 storm, carving a path of destruction through South Florida before continuing northwest into Louisiana. When the accounting was finished, Andrew had caused roughly $15.5 billion in insured losses in 1992 dollars, destroyed more than 25,000 homes, and pushed somewhere between eight and sixteen U.S. insurance companies into insolvency, depending on which source you consult (Insurance Information Institute [III], n.d.; National Association of Insurance Commissioners [NAIC], 2025; Verisk, 2017).
The financial damage was one thing. The intellectual damage was another.
For most of the twentieth century, property-casualty actuaries had priced catastrophe exposure the same way they priced most other coverages: by looking at historical losses, trending them forward for inflation and exposure growth, and loading for profit and risk. It was a methodology refined over decades and codified in actuarial literature. It worked beautifully for auto collision claims, workers’ compensation, homeowners’ non-catastrophe perils, and everything else with enough historical data to fit a credible distribution.
It did not work for Andrew.
Before the storm, loss estimates generated through conventional actuarial techniques — extrapolating from recent claim history — had catastrophically underestimated the exposure (III, n.d.; Repath, 2026). Florida’s most recent reference point for a Category 5 storm was the 1935 Labor Day Hurricane, which had tracked along a sparsely populated stretch of the Gulf Coast. In the intervening 57 years, South Florida had transformed from a winter retreat for the wealthy into a densely developed metropolitan region holding hundreds of billions of dollars in insured property. The historical record contained no event that remotely resembled a major hurricane striking modern Miami-Dade County, which meant the actuarial models had nothing to anchor on (Moody’s, 2022).
Andrew did not just produce the largest insured loss the U.S. insurance industry had ever seen at that time. It produced proof that the industry’s core pricing methodology was broken for low-frequency, high-severity perils. As the Insurance Information Institute later summarized, “Andrew helped to prove that past data is a poor gauge for future catastrophe exposure. Previous projections failed to recognize that science indicated unprecedented physical events were within the realm of reasonable possibility” (III, n.d., p. 3).
What emerged from the wreckage of Andrew — and, two years later, the 1994 Northridge earthquake, which delivered a similar lesson for seismic risk — was a new discipline. It borrowed from atmospheric science, structural engineering, probability theory, and computational simulation. It was called catastrophe modeling, and over the next three decades it would become the intellectual backbone of how the global insurance industry prices, capitalizes, and hedges against the risks that keep CEOs awake at night.
Today, that same discipline is facing its own existential test. Climate change is rendering the historical record — already a shaky foundation — even less reliable. Weather perils that were once considered secondary are producing losses that rival peak perils. Entire geographies are becoming uninsurable, at least by private carriers operating on commercial terms. And the catastrophe models themselves, built on the assumption of a reasonably stable climate system, are being asked to quantify a world that is rapidly departing from their calibration data.
This article is an inside look at how catastrophe modeling actually works — the modules, the math, the metrics — and how practitioners are trying to adapt a discipline built for a stationary climate to a world that is no longer stationary. It is written for anyone who wants to understand the science and judgment behind the numbers on their homeowners’ insurance bill, and for anyone curious about one of the hardest quantitative problems in modern finance: how to put a price on something that has never happened but almost certainly will.
Why Traditional Actuarial Methods Fail for Catastrophes
Before explaining what catastrophe models do, it is worth being precise about what they replaced and why.
Traditional actuarial ratemaking for property coverage relies on what practitioners call the loss cost method. An actuary gathers several years of loss experience for a given line of business, adjusts the losses to current cost levels, divides by exposure (typically house-years or $100 of insured value), and produces a pure premium. Layer on expenses, profit, and a risk load, and you have a rate.
This methodology rests on an assumption that is almost always reasonable for frequency-driven perils: the losses observed in the recent past are a statistically credible sample of the losses that will occur in the near future. For auto collision claims, where a typical insurer writes millions of exposure-years and claims arrive with Poisson-like regularity, this is close to self-evidently true. The law of large numbers does the work.
For catastrophe perils, the assumption collapses for three distinct reasons.
The first is the frequency problem. Major hurricanes striking specific coastal segments, magnitude-7-plus earthquakes on particular fault sections, and $10-billion-plus wildfire complexes are events with return periods measured in decades or centuries. As NAIC’s (n.d.-a) overview notes, “Because catastrophes are volatile and occur infrequently, even using 20-year averages may not adequately represent the true risk the company is exposed to” (para. 2). A 20-year historical window for Miami might contain zero major hurricane landfalls or it might contain three, and either sample would be essentially uninformative about the true underlying frequency.
The second is the exposure problem. Even when historical losses exist, the property and population at risk have usually changed dramatically. Florida’s population grew from 13.69 million in 1992 to 22.13 million by 2022, a 61% increase, making it the third most populous U.S. state (Moody’s, 2022). The number of buildings within Hurricane Andrew’s original wind footprint has risen roughly 40%, and modeling an Andrew-like storm against today’s exposure base produces radically different loss numbers than the actual 1992 experience (Moody’s, 2022). Historical losses, however carefully trended, do not capture this compositional shift.
The third is the correlation problem. Catastrophes produce massive simultaneous losses across thousands of policies. A single hurricane can generate tens of thousands of claims in a single afternoon. This is the opposite of the independence assumption that underlies most actuarial models, and it is precisely what causes insurer insolvencies. The company is not ruined by the expected annual loss; it is ruined by the tail event that wipes out ten years of premium in a single day.
The insurance industry understood all three problems in principle long before Andrew. What it lacked was a viable alternative methodology. You cannot simply declare that historical data is insufficient without specifying what should replace it. Catastrophe modeling, as it developed in the 1990s and matured in the 2000s, provided that alternative.
The Core Idea: Simulate What Has Not Happened Yet
The fundamental insight behind catastrophe modeling is simple to state and extraordinarily difficult to execute: instead of inferring future losses from historical losses, simulate future losses directly by combining physical science (what hazards are possible and how often they occur), engineering knowledge (how structures respond to those hazards), exposure data (what is actually at risk), and financial modeling (how losses flow through insurance contracts).
Karen Clark, who founded Applied Insurance Research — the first commercial catastrophe modeling firm, later acquired and rebranded as AIR Worldwide, now part of Verisk — had been building toward this idea since the late 1980s (AXA XL, 2022; Repath, 2026). The firm launched in 1987, five years before Andrew, and published a probabilistic hurricane model that had estimated potential Florida insured losses at levels the industry at the time considered implausible (AXA XL, 2022). Then Andrew happened, and the industry-wide narrative on probabilistic cat modeling shifted permanently. Risk Management Solutions (RMS), now part of Moody’s, followed shortly after, and by the mid-1990s, catastrophe modeling had become standard practice across the insurance and reinsurance industries (Repath, 2026). Today, the three major cat modeling vendors — Verisk, Moody’s RMS, and CoreLogic — collectively model more than 100 natural and man-made perils across over 100 countries (Repath, 2026).
At its core, a modern catastrophe model is a computational pipeline with four standard modules. The naming conventions vary slightly by vendor, but the function of each component is consistent across the industry (NAIC, 2025; Oasis Loss Modelling Framework [Oasis LMF], n.d.).
The Hazard Module
The hazard module is where the atmospheric and geophysical science lives. Its job is to generate a catalog of simulated events — typically tens of thousands to hundreds of thousands of them — that represent the full plausible range of what nature might produce. Each event in the catalog is parameterized by its physical characteristics (for a hurricane: landfall location, central pressure, forward speed, radius of maximum winds, track) and carries an annual probability of occurrence, also called its event rate (NAIC, 2025).
For a hurricane model, the hazard module typically starts with the historical best-track database maintained by the National Hurricane Center (in the U.S. context, this is the HurDat2 dataset maintained by NOAA). Statistical distributions are fit to the historical parameters — landfall frequency by coastal segment, the relationship between central pressure and maximum sustained wind, the distribution of forward speeds, and so on. Monte Carlo simulation is then used to draw from these distributions and generate a synthetic event set that is much larger than the historical record but statistically consistent with it (Oasis LMF, n.d.).
The required size of the stochastic event set depends on the peril and the geographic scope. For a simple peril in a small country, a few thousand events may suffice; for earthquake or windstorm in the United States, catalogs of 100,000 years of simulated activity or more are standard (Oasis LMF, n.d.; Verisk, 2013). The motivation for these huge catalogs is statistical: to estimate the 1-in-250-year loss with reasonable confidence, you need to have simulated enough years of activity that the tail of the distribution is well-populated.
Once an event is drawn from the catalog, the hazard module computes its local intensity footprint — the wind speeds, storm surge heights, earthquake shaking intensities, or flood depths at every location within the affected area. This is where physics-based modeling takes over from statistical modeling. Parametric wind field models such as the Holland profile translate central pressure and radius of maximum winds into a spatially distributed wind field. Storm surge models couple the wind and pressure fields to coastal bathymetry to compute inundation. Earthquake ground motion prediction equations translate magnitude, depth, and distance into local shaking intensity.
The output of the hazard module is, in effect, a massive table: for each of potentially hundreds of thousands of synthetic events, the local hazard intensity at every modeled location.
The Exposure Module
The exposure module holds the portfolio data — what is actually at risk. For a property insurer, this means location-level information on every insured structure: the street address or latitude/longitude, construction type, occupancy class, number of stories, year built, insured value, and insurance policy terms including deductibles, sublimits, and coinsurance (NAIC, 2025).
Exposure data quality is one of the quiet determinants of modeling accuracy. Two insurers with identical policy counts in the same ZIP code can have materially different modeled losses if one has captured precise building characteristics and the other is relying on ZIP-centroid geocoding and default construction assumptions. A substantial portion of the work done by catastrophe modeling teams at large insurers and reinsurers is not running the models themselves but rather cleaning, enriching, and validating the exposure data that feeds them.
The Vulnerability Module
The vulnerability module — sometimes called the engineering module or damage module — is where physical hazard becomes financial damage. It contains vulnerability functions, also called damage functions or fragility curves, that map hazard intensity to expected damage ratio (NAIC, 2025; Oasis LMF, n.d.).
A vulnerability function for, say, a one-story wood-frame single-family home in hurricane-prone Florida built to post-2001 building code standards might state that at a peak gust wind speed of 120 mph, the expected damage ratio is 15% of insured value, with a specified uncertainty distribution around that mean. At 150 mph, the expected damage ratio might be 45%. At 180 mph, it might exceed 80%. These functions are derived from a combination of post-event claims data, engineering judgment, laboratory testing of building components, and physics-based structural analysis.
Crucially, vulnerability functions differ by construction type, occupancy, age, and geography. A reinforced concrete commercial building responds very differently to wind than a manufactured home. Buildings constructed under modern Florida hurricane codes — enacted and progressively tightened after Andrew — respond differently than older structures (Claims Journal, 2022; III, n.d.). Vulnerability modelers continuously refine these functions based on claims analyses from major events, which is why major storms such as Hurricanes Katrina, Ike, and Ian have each produced updates to industry vulnerability models.
The Financial Module
The financial module takes the modeled physical damage and translates it into insured loss. This is where insurance policy terms get applied: deductibles (including the separate hurricane and wind deductibles common in catastrophe-exposed states), policy limits, sublimits for specific perils such as named windstorm or storm surge, coinsurance clauses, and attachment points and limits for reinsurance contracts (NAIC, 2025; Oasis LMF, n.d.).
A $500,000 home with a 2% hurricane deductible and a modeled gross damage of $50,000 produces a net insured loss of $40,000 after the $10,000 deductible is subtracted. Run this calculation across a portfolio of 200,000 policies for each of 100,000 simulated years, and you have the raw material for every output a catastrophe modeler cares about.
What Comes Out: The Exceedance Probability Curve
The headline output of a catastrophe model is the exceedance probability (EP) curve — a graph that shows, for every possible loss level, the annual probability that actual losses will exceed that level (NAIC, n.d.-a; Verisk, 2013).
There are two standard flavors of EP curve, and confusing them is one of the most common errors in cat modeling interpretation (Homer & Li, 2017).
The Occurrence Exceedance Probability (OEP) curve tracks the largest single event in each simulated year. If a simulated year contains a hurricane causing $3 billion in losses, a hurricane causing $1.5 billion, and a thunderstorm causing $200 million, the OEP for that year is based on the $3 billion event. OEP is the natural curve for sizing per-occurrence reinsurance programs and for regulatory capital frameworks that focus on single-event stress scenarios (Continuuiti, 2026; Homer & Li, 2017).
The Aggregate Exceedance Probability (AEP) curve tracks the sum of all events in each simulated year. Using the same year, the AEP would be based on $4.7 billion. AEP is the natural curve for annual budgeting, portfolio pricing, and aggregate reinsurance placements (Continuuiti, 2026; Homer & Li, 2017).
Both curves are typically expressed in terms of return periods — the reciprocal of the annual exceedance probability. A “1-in-100-year loss” is the loss level with a 1% annual probability of being exceeded. A “1-in-250-year loss” has a 0.4% annual probability. These return-period losses are sometimes called Probable Maximum Loss (PML) at the specified return period, though terminology varies (Homer & Li, 2017).
Two numeric outputs derived from the EP curve deserve special attention because they flow directly into pricing and capital decisions.
The Average Annual Loss (AAL) is the expected value of the annual loss distribution — equivalently, the area under the AEP curve, or the mean loss per simulated year averaged across the entire event catalog (Continuuiti, 2026; NAIC, n.d.-a; Verisk, 2013). In a 100,000-year simulation, AAL is simply total simulated losses divided by 100,000. AAL is the theoretical “pure premium” for catastrophe risk — the long-run expected cost per year. When an actuary prices a property policy, the AAL from the cat model is a direct input to the rate, typically forming the expected loss component of the cat load.
Consider a simplified example from Verisk’s (2013) tutorial on AAL: a modeler runs a 10-year hurricane catalog against a geographically dispersed property portfolio. The resulting annual losses might look like $0, $0, $5M, $10M, $15M, $20M, $25M, $40M, $75M, and $170M — seven years below the average, three years above, two years with no losses at all, and one catastrophic tail year. Averaging these produces an AAL of $36 million. Notice how profoundly the single $170M year shapes the AAL: strip it out and the average drops to roughly $18M. This is the signature of catastrophe risk — the mean is driven by the tail, not by the middle of the distribution.
The standard deviation and higher moments of the annual loss distribution drive the risk load — the additional premium charged beyond expected loss to compensate for the volatility of the risk. A portfolio with AAL of $36 million but a 1-in-100-year loss of $1 billion requires dramatically more capital — and therefore dramatically higher premium — than one with the same AAL but a 1-in-100-year loss of $150 million.
This relationship between AAL and tail risk is why catastrophe-exposed insurance is so expensive relative to its expected losses. The ratio of premium to AAL for Florida hurricane coverage can be three or four times the ratio for auto collision, not because insurers are being greedy but because the capital required to stand behind the promise of paying a 1-in-100-year loss is enormous, and that capital has a cost.
From Model Output to Policy Price
How does the output of a catastrophe model actually translate into the number on a policyholder’s premium bill? The mechanics depend on the line of business and regulatory environment, but the general architecture is consistent.
For a primary homeowners’ insurer in a catastrophe-exposed state, the process typically begins with running the insurer’s in-force portfolio through one or more catastrophe models to generate an AAL and EP curve at the total portfolio level. The AAL is then allocated down to individual policies using a geographic detail level consistent with the model’s resolution — typically something between ZIP code and street-address precision. This allocation produces a per-policy cat AAL that feeds into the rate calculation.
At the individual policy level, the cat AAL is often the single largest component of the pure premium in high-risk coastal or wildfire-exposed areas. For a home in Miami-Dade County, the modeled hurricane AAL can easily exceed the combined AAL from all non-catastrophe perils (fire, theft, water damage from plumbing failures, etc.) by a factor of two or three.
On top of the cat AAL, the insurer adds a cat risk load reflecting the cost of capital required to support the tail risk. This is typically calculated as a function of the marginal contribution of the policy (or geographic segment, or peril) to the tail of the portfolio’s EP curve. One common approach uses the concept of tail value at risk (TVaR) — the expected loss given that losses exceed a specified return period — as the basis for the capital allocation. The price of reinsurance purchased to offload some of this tail risk is often the most direct signal of the market cost of catastrophe capital, and many insurers explicitly pass through their reinsurance costs (or a loading reflective of them) into their primary rates.
Regulatory and rating agency frameworks add further structure. NAIC risk-based capital standards, S&P’s capital adequacy framework, A.M. Best’s BCAR model, and Solvency II’s standard formula or internal model requirements all incorporate catastrophe exposure explicitly, and all rely heavily on catastrophe model output as input (NAIC, 2025). An insurer that cannot demonstrate adequate capital against, typically, the 1-in-200-year or 1-in-250-year return-period loss will face rating downgrades or regulatory intervention, which means the cat model’s tail estimates directly constrain how much catastrophe business the insurer can write.
As the American Academy of Actuaries documented in its 2018 paper on the subject, catastrophe model output flows into ratemaking, mitigation credit quantification, reinsurance purchasing, capital management, and solvency assessment (NAIC, n.d.-a). The model is, quite literally, the engine of the business.
Climate Change Breaks the Historical Calibration
Catastrophe models were a huge advance over historical loss extrapolation, but they share a subtle dependency with the methodology they replaced: they are calibrated against historical data. The hazard module’s statistical distributions of hurricane frequencies, intensities, and tracks are fit to the historical record. The vulnerability functions are tuned to observed claims experience. The resulting event catalog is, in effect, a carefully constructed statistical portrait of the climate system as it existed over the past several decades.
This calibration works well if the climate system is stationary — if the underlying probability distributions are stable through time. If hurricane genesis frequencies, maximum intensities, precipitation rates, and tracks are drawn from fixed distributions, then a catalog fit to historical data is an accurate portrait of the future.
The climate system is no longer stationary.
The Intergovernmental Panel on Climate Change’s Sixth Assessment Report concluded definitively that “we now live in a non-stationary climate” and that anthropogenic climate change is unequivocally shifting the distributions of weather and climate extremes (Seneviratne et al., 2021, Chapter 11). For catastrophe modelers, this is not an abstract concern. It is a direct challenge to the statistical foundation of their discipline.
The practical consequences show up in several distinct ways:
Hurricanes are intensifying. Warmer sea surface temperatures provide more energy for tropical cyclones, and a warmer atmosphere holds more moisture (per the Clausius-Clapeyron relation, roughly 7% more water vapor per degree Celsius). The result, consistently documented in the peer-reviewed literature, is an observed increase in the proportion of tropical cyclones reaching major hurricane status and in the rainfall rates associated with these storms. Swiss Re and Munich Re have both attributed a substantial portion of the rising trend in insured losses from windstorm and flood to climate-driven intensification (Munich Re, 2025a, 2026; Swiss Re Institute, 2025).
So-called “secondary perils” are behaving like primary perils. The insurance industry historically categorized hurricanes and earthquakes as “peak perils” — high-severity, infrequent events that dominated tail risk — and wildfires, severe convective storms (thunderstorms, hail, tornadoes), floods, and droughts as “secondary perils.” In 2025, Munich Re reported that non-peak perils produced total insured losses of roughly $98 billion — on par with or exceeding losses from peak perils in many recent years (Munich Re, 2026). The distinction between primary and secondary has blurred, and models that were originally built with secondary perils as an afterthought are being asked to carry pricing weight they were not designed for.
Wildfire risk is expanding geographically. California’s 2025 Los Angeles fires produced an estimated $40 billion in insured losses, making them the costliest natural disasters of 2025 globally (Munich Re, 2026; Weiss, 2025). Ten of the 20 largest wildfires in California history have occurred in the last five years (Gaul, 2025). Wildfire models that were calibrated on decades of lower-frequency data are being stress-tested against observed event rates that exceed their historical distributions.
Flood losses are rising in both coastal and inland geographies. Climate-driven increases in extreme precipitation are producing pluvial (rainfall-driven) and fluvial (river) flood losses in regions that were historically considered low-flood-risk. Coastal flood risk is compounded by sea level rise, which adds a slow baseline increase on top of storm-driven surge.
The technical challenge this creates for catastrophe modelers is sometimes called the non-stationarity problem: how do you fit statistical distributions to historical data and project them into the future when the underlying system is changing? The phrase “stationarity is dead,” taken from a widely-cited 2008 paper by Milly et al. in the hydrology literature (cited in Cheng & AghaKouchak, 2016), has become a refrain in flood and precipitation modeling.
Guy Carpenter, a major reinsurance broker, estimated in an internal analysis that near-term climate change adds approximately 1% per annum to global average annual loss across the industry, with substantial uncertainty around that figure and with some individual perils increasing at higher rates (Turner, 2023). In actuarial terms, that is an enormous number compounded over multi-year horizons. And yet, as the author of that analysis noted, “these annual increases are rarely applied in practice by insurers; in a competitive market few want to be the first to move” (Turner, 2023, para. 4).
How the Industry Is Adapting
The catastrophe modeling industry has not been sitting still in the face of climate-driven non-stationarity. Several distinct adaptation strategies are playing out in parallel.
Climate-Conditioned Event Sets
The major model vendors now offer climate-conditioned or forward-looking versions of their hazard catalogs. Rather than calibrating solely to the long-term historical record, these catalogs are weighted toward — or adjusted to reflect — recent climatological conditions, with some versions projecting forward to specified future time horizons (2030, 2050, 2100) under specified emissions scenarios drawn from IPCC Shared Socioeconomic Pathways (SSPs).
Get Ignatius Wang’s stories in your inbox
Join Medium for free to get updates from this writer.
For hurricane modeling, this typically involves adjusting sea surface temperatures, atmospheric moisture, and vertical wind shear profiles in the physics-based storm generation models. For wildfire, it means adjusting fuel conditions, temperature, and precipitation distributions. For flood, it involves updating intensity-duration-frequency (IDF) curves to reflect projected changes in extreme precipitation.
The output is a parallel EP curve for the “climate-conditioned” view of risk that can be compared against the historical view. The difference between the two is the model’s best estimate of climate change’s impact on expected loss and tail risk.
Stochastic Year-Loss Tables and Blending
Modern workflows typically produce year-loss tables (YLTs) — records of every simulated year’s losses broken out by event — that allow for sophisticated post-processing. Actuaries can weight events differently, blend outputs from multiple vendor models, and perform sensitivity analysis on key assumptions (Institute and Faculty of Actuaries, n.d.).
The Institute and Faculty of Actuaries’ (n.d.) working paper on model blending notes that frequency blending of AALs across multiple cat models has emerged as an industry standard practice for reinsurance pricing, reflecting genuine epistemic uncertainty about which model’s view of the risk is most accurate. An insurer that relied on a single model would be exposed to a single vendor’s idiosyncratic assumptions; blending multiple views hedges that risk.
Open-Source and Alternative Frameworks
The Oasis Loss Modelling Framework, founded by a consortium of insurers frustrated with the proprietary “black box” nature of commercial vendor models, has provided an open-source platform that allows third-party hazard modelers (including academic groups and startup firms focused on specific perils or geographies) to plug into a standardized loss calculation engine (Oasis LMF, n.d.). The practical effect is a proliferation of alternative models for specific perils — Fathom for flood, KatRisk for climate-conditioned multi-peril — that insurers can use alongside the traditional vendors to triangulate their view of risk (Fathom, 2024; KatRisk, 2024).
Incorporating Climate Science More Deeply
At the research frontier, catastrophe modelers are working to incorporate higher-resolution climate model output, attribution science (which estimates the extent to which specific extreme events have been made more likely or more severe by climate change), and machine learning techniques for downscaling global climate model output to the local resolutions needed for insurance pricing.
World Weather Attribution, a research collaboration that publishes rapid attribution studies after major events, found that climate change made the catastrophic flash floods in Valencia, Spain in late 2024 approximately twice as likely to occur (Munich Re, 2025a). This kind of analysis is increasingly being integrated into cat model updates, allowing modelers to shift event frequencies in a scientifically defensible way rather than waiting for enough new events to accumulate in the historical record to shift the underlying distributions.
AI and Generative Modeling
The integration of artificial intelligence into catastrophe modeling has accelerated over the past three years, with applications spanning data quality enhancement (automated exposure cleaning, satellite-based property characteristic inference), hazard simulation (neural network surrogates for physics-based models, which can generate event footprints orders of magnitude faster than traditional approaches), and claims analytics (rapid post-event loss estimation from satellite imagery). The Mass Technology Leadership Council (2025) described this as a “game-changer” for the industry, noting that AI enhances data accuracy, reduces uncertainties, and adapts to evolving threats.
None of these adaptations, individually or collectively, fully solves the non-stationarity problem. Catastrophe modelers are upfront that climate-conditioned catalogs involve substantial assumptions and should be used as one input to decision-making rather than as a definitive forecast. But the discipline is, in aggregate, moving toward a more explicitly forward-looking posture than it had a decade ago.
Where the Math Meets the Market: The Insurance Crisis
None of this is abstract. The consequences of climate-driven increases in catastrophe risk — and of the modeling methodologies that quantify them — are playing out in real-time in the U.S. property insurance market.
In 2024, natural catastrophes produced approximately $140 billion in insured losses globally, the third-most-expensive year on record, with Munich Re’s chief climate scientist describing climate change as “showing its claws” (Munich Re, 2025a; Insurance Journal, 2025). The dominant U.S. events were Hurricanes Helene and Milton, which together produced insured losses in the tens of billions (Munich Re, 2025a).
2025 was lower in absolute terms — roughly $108 billion in insured losses according to Munich Re (2026) — but still the fifth consecutive year exceeding $100 billion, a threshold that was almost unheard of before 2017. Sheer luck, as Munich Re put it, spared the U.S. from hurricane landfalls in 2025, but the Los Angeles wildfires produced approximately $40 billion in insured losses, making them the costliest natural disasters of the year globally (Munich Re, 2026).
The market response has been dramatic. Florida and California led the nation in homeowners’ insurance non-renewal rates in 2024, with Florida at 3.35% and California at 3.18% (National Mortgage News, 2025). The California rate represented a 3.9-fold increase over 2018 levels; Louisiana, at 2.97%, was up 5.4-fold over the same period (National Mortgage News, 2025). The World Wildlife Fund reported that more than 1.9 million home insurance policies have been non-renewed in the U.S. since 2018, with the share of uninsured homes doubling from 5% in 2019 to 12% by 2025 (World Wildlife Fund [WWF], 2026).
The dynamics in Florida and California illustrate how regulatory structure interacts with actuarial modeling to produce different market outcomes. Florida, following years of insurer insolvencies and rapid growth of its state-run Citizens Property Insurance Corporation (which reached 1.4 million policies at its peak in 2023), implemented tort reform in 2023 aimed at reducing litigation-driven loss costs (New America, 2025; Newsweek, 2025). By the fourth quarter of 2024, Florida reported its first statewide decrease in average homeowners premiums in nearly a decade — down 0.7% (Newsweek, 2025). Eleven new property insurers entered the Florida market, and Citizens’ policy count had dropped below 850,000 by early 2025 through depopulation efforts (New America, 2025; Newsweek, 2025).
California’s trajectory has been harsher. The state’s Proposition 103, enacted in 1988, constrains insurers’ ability to pass through reinsurance costs and incorporate forward-looking catastrophe model output into rates (Newsweek, 2025). As one industry spokesperson put it, “Californians are paying artificially low home premiums that are not actuarially sound” (Newsweek, 2025, para. 10). The result was a private insurer retreat and explosive growth of the California FAIR Plan, the state-backed insurer of last resort, whose total exposure reached $529 billion by December 2024 — a 217% increase from September 2021 (Newsweek, 2025). When the January 2025 Los Angeles wildfires hit, the FAIR Plan faced over $4 billion in claims, exceeding its reserves and reinsurance, and required a $1 billion bailout that was partially funded through a surcharge on all California policyholders (Yale Law Journal, 2025).
In both states, the core tension is the same: catastrophe models are telling insurers that the actuarially sound premium is higher than what the market or the regulator is willing to accept. When regulators suppress the premium, insurers withdraw. When insurers withdraw, state-backed pools take on more exposure, often without the capital to match. The entire structure is being stress-tested against a climate that is departing from the historical baseline on which the whole system was originally calibrated.
Munich Re stated bluntly in its assessment that “there is a growing number of properties that are becoming uninsurable, unsaleable, and ultimately unusable because of their location” (International Institute for Management Development, 2025, para. 2). Yale E360’s reporting framed the broader pattern: “The risks associated with our hotter, wetter world are growing as natural disasters become more damaging” (Gaul, 2025, para. 3).
The Philosophical Problem: Pricing the Unpriceable
Underneath the technical sophistication and the market turbulence, catastrophe modeling faces a philosophical challenge that deserves explicit attention.
The discipline’s central premise is that low-frequency, high-severity risks can be priced probabilistically — that it is possible to assign meaningful probabilities to events that may never have occurred in the historical record but are physically plausible. This is fundamentally different from pricing a coin flip or an auto collision, where the probabilities can be estimated empirically to high precision.
For a 1-in-500-year hurricane affecting a specific coastal segment, nobody has 500 years of reliable loss data. Nobody has 500 years of reliable meteorological data. Nobody has 500 years of reliable building stock data. The 1-in-500-year loss is, in the end, an informed model output based on physics, statistics, and judgment — not a frequency observed empirically.
This is not a criticism of catastrophe modeling. It is the nature of the problem. But it has two important implications that are often underappreciated.
First, model uncertainty is substantial, and different models can produce meaningfully different answers for the same portfolio and peril. This is why sophisticated buyers of catastrophe risk (large insurers, reinsurers, capital market investors in catastrophe bonds) routinely run multiple vendor models and blend the outputs. It is also why reinsurance pricing often incorporates a “model uncertainty load” beyond the AAL — a recognition that the model output is itself an estimate subject to meaningful error.
Second, the rarer the event, the more dominant the role of assumption rather than data. A model’s estimate of the 1-in-10-year loss is generally close to what a long historical record would give you. Its estimate of the 1-in-1000-year loss is almost entirely driven by the tail assumptions in the underlying probability distributions, which are in turn driven by scientific judgment about the plausible upper bounds of the physical system. This is why two reputable models can produce 1-in-250-year losses that differ by 30% or more — not because one is wrong, but because the underlying science does not constrain the answer more tightly than that.
The Andrew-era promise of catastrophe modeling was that it would replace the inadequate historical-data methodology with something scientifically grounded. That promise has been largely kept. But it has been replaced with a more subtle promise: that sophisticated probabilistic modeling can produce estimates of tail risk that are useful for business decisions even when they cannot be empirically validated. That promise is harder to keep, and climate change makes it harder still.
What Actuaries Do Next
For practicing actuaries — and for the broader financial ecosystem that depends on their work — the path forward involves several concurrent adaptations.
The first is intellectual honesty about uncertainty. Catastrophe model output is frequently presented with spurious precision — point estimates of 1-in-100-year losses reported to three significant figures. Best practice is shifting toward explicit uncertainty bands and scenario-based communication. When the 1-in-100-year loss is reported as $2.4 billion, the honest accompanying disclosure is that the plausible range is $1.8–$3.2 billion, and that this range itself may understate the true uncertainty in a non-stationary climate.
The second is integration of forward-looking climate science. This means moving beyond backward-looking catalogs to explicit scenario analysis — running portfolios against climate-conditioned event sets under multiple emissions pathways and time horizons. The Task Force on Climate-related Financial Disclosures (TCFD) framework, now incorporated into jurisdictional disclosure requirements including the European Sustainability Reporting Standards (ESRS) and the International Sustainability Standards Board (ISSB) standards, has formalized this expectation for the largest financial institutions (Munich Re, 2025b).
The third is expanded use of parametric and alternative risk transfer. Parametric insurance — which pays out based on the occurrence of a defined event (e.g., a hurricane of specified intensity passing within a specified radius of a specified point) rather than measured losses — reduces the reliance on detailed loss adjustment and can expand insurability in geographies where traditional indemnity coverage is failing. Catastrophe bonds, which transfer peak peril risk directly to capital markets, have grown from a post-Andrew experiment to a mature asset class, with new issuance regularly exceeding $10 billion per year (Federal Reserve Bank of Chicago, 2018).
The fourth is deeper engagement with mitigation and resilience. Catastrophe models are increasingly being used not just to price existing risk but to quantify the loss-reduction benefits of mitigation investments — stronger building codes, retrofit programs, defensible space around structures in wildfire-exposed areas, managed retreat from the highest-risk coastlines. This creates a feedback loop in which the actuarial quantification of risk can inform public policy decisions that, over time, reshape the underlying risk itself.
The fifth, and perhaps the most consequential, is confronting the question of what happens when some risks become genuinely uninsurable on commercial terms. The private insurance market is extraordinarily effective at pooling and pricing risks up to a certain frequency-severity envelope, but catastrophe risk in the most exposed geographies is approaching and in some cases exceeding that envelope. The policy debate about what comes next — government backstops, universal coverage pools funded by broad-based surcharges, land-use reforms that reduce exposure at the source — is a debate that actuaries and catastrophe modelers will be asked to inform with rigorous analysis, even when the conclusions are uncomfortable.
Conclusion: Humility in the Face of Nature
Catastrophe modeling is, in many ways, one of the great intellectual achievements of modern actuarial science. It took a problem that had defeated decades of traditional methodology — pricing low-frequency, high-severity risks — and produced a tractable, scientifically grounded alternative. It has enabled an insurance industry that can stand behind trillions of dollars of property exposure in hurricane- and earthquake-prone regions that would otherwise be commercially uninsurable. It has given governments, investors, and policyholders a common language for discussing tail risk.
It has also, as every honest practitioner will acknowledge, never been tested against a world that is simultaneously warming, urbanizing, and departing from its historical statistical baseline. The Hurricane Andrew moment — the recognition that a methodology was inadequate to the problem — may be approaching again, not because catastrophe modeling is wrong but because the climate system is changing faster than our ability to re-calibrate.
The actuaries and catastrophe modelers who will lead the next chapter of this discipline are the ones who hold two ideas in their minds simultaneously: that the probabilistic framework is genuinely useful, and that its outputs are genuinely uncertain. They are the ones who can run a model, present its output to a decision-maker, and say honestly, “This is our best estimate. Here is why. Here is what we don’t know. Here is what could make it very wrong.”
Pricing the unpriceable is, in the end, an exercise in humility as much as in mathematics. The mathematics is necessary. The humility is what keeps the mathematics honest.
References
AXA XL. (2022, August 8). Remembering Hurricane Andrew’s lessons, 30 years later. https://axaxl.com/fast-fast-forward/articles/remembering-hurricane-andrews-lessons-30-years-later
Cheng, L., & AghaKouchak, A. (2016). Hydrologic extremes under climate change: Non-stationarity and uncertainty. In S.-Y. S. Wang, J.-H. Yoon, C. C. Funk, & R. R. Gillies (Eds.), Climate extremes and their implications for impact and risk assessment (pp. 15–26). Springer. https://link.springer.com/chapter/10.1007/978-981-10-2051-3_2
Claims Journal. (2022, August 29). Viewpoint: Can the Florida insurance market withstand a $100B repeat of Hurricane Andrew? https://www.claimsjournal.com/news/national/2022/08/29/312409.htm
Continuuiti. (2026, March 3). Exceedance probability: Definition, formula, and EP curves explained. https://continuuiti.com/blog/exceedance-probability/
Fathom. (2024, May 28). Flood risk and the modern actuary: Climate change and insurance. https://www.fathom.global/insight/flood-risk-and-the-modern-actuary/
Federal Reserve Bank of Chicago. (2018). Catastrophe bonds: A primer and retrospective (Chicago Fed Letter №405). https://www.chicagofed.org/publications/chicago-fed-letter/2018/405
Gaul, G. M. (2025, September 15). How climate risks are putting home insurance out of reach. Yale Environment 360. https://e360.yale.edu/features/climate-change-home-insurance
Homer, D., & Li, M. (2017). Notes on using property catastrophe model results. Casualty Actuarial Society E-Forum, Spring 2017–Volume 2. https://www.casact.org/sites/default/files/2021-02/2017_most-practical-paper_homer-li.pdf
Institute and Faculty of Actuaries. (n.d.). Catastrophe model blending techniques and governance. https://www.actuaries.org.uk/system/files/documents/pdf/catastrophe-model-blending.pdf
Insurance Information Institute. (n.d.). Hurricane Andrew and insurance: The enduring impact of an historic storm. https://www.iii.org/sites/default/files/paper_HurricaneAndrew_final.pdf
Insurance Journal. (2025, January 9). Insured losses from natural disasters hit $140B as climate change “shows its claws”. https://www.insurancejournal.com/news/international/2025/01/09/807524.htm
International Institute for Management Development. (2025, August 7). Climate change — Uninsurable areas. https://www.imd.org/ibyimd/sustainability/climate-change-the-emergence-of-uninsurable-areas-businesses-must-act-now-or-pay-later/
KatRisk. (2024). Catastrophe risk modeling. https://www.katrisk.com/
Mass Technology Leadership Council. (2025, February 21). Gen AI in catastrophe modeling: Enhancing climate resilience in insurance. https://www.mtlc.co/gen-ai-in-catastrophe-modeling-enhancing-climate-resilience-in-insurance/
Moody’s. (2022). 30 years on from Hurricane Andrew: Managing a storm of Florida’s own making. https://www.moodys.com/web/en/us/insights/insurance/30-years-on-from-hurricane-andrew-managing-a-storm-of-floridas-own-making.html
Munich Re. (2025a, January 9). Climate change presses on: Natural disaster losses in 2024. https://www.munichre.com/
Munich Re. (2025b). Climate ambition and reporting. https://www.munichre.com/en/company/sustainability/climate-ambition-and-reporting.html
Munich Re. (2026, January 13). Climate change presses on: Devastating wildfires and intense thunderstorms exacerbate losses for insurers. https://www.munichre.com/en/company/media-relations/media-information-and-corporate-news/media-information/2026/natural-disaster-figures-2025.html
National Association of Insurance Commissioners. (n.d.-a). Catastrophe models (property). https://content.naic.org/insurance-topics/catastrophe-models-(property)
National Association of Insurance Commissioners. (n.d.-b). Catastrophe models (property) — CIPR topics. https://content.naic.org/cipr-topics/catastrophe-models-property
National Association of Insurance Commissioners. (2025, March). Catastrophe modeling primer. https://content.naic.org/sites/default/files/committees-pending-action-cat-mod-primer.pdf
National Mortgage News. (2025, October 29). Home insurance crisis deepens in Florida, California. https://www.nationalmortgagenews.com/news/home-insurance-crisis-deepens-in-florida-california
New America. (2025). Climate change, housing, and homeowners insurance in Florida: Lessons for California. https://www.newamerica.org/future-land-housing/briefs/insurance-in-florida-lessons-for-california/
Newsweek. (2025, April 7). Florida vs. California: How two states are facing the insurance crisis. https://www.newsweek.com/florida-california-how-two-states-facing-insurance-crisis-2055256
Oasis Loss Modelling Framework. (n.d.). Introduction to catastrophe modelling. https://oasislmf.org/application/files/6917/1624/0852/Intro_toCatModelling.pdf
Repath. (2026, February 23). Catastrophe risk modeling: Why cat models fail infrastructure investors. https://repath.earth/catastrophe-risk-modeling/
Seneviratne, S. I., Zhang, X., Adnan, M., Badi, W., Dereczynski, C., Di Luca, A., Ghosh, S., Iskandar, I., Kossin, J., Lewis, S., Otto, F., Pinto, I., Satoh, M., Vicente-Serrano, S. M., Wehner, M., & Zhou, B. (2021). Weather and climate extreme events in a changing climate. In V. Masson-Delmotte, P. Zhai, A. Pirani, S. L. Connors, C. Péan, S. Berger, N. Caud, Y. Chen, L. Goldfarb, M. I. Gomis, M. Huang, K. Leitzell, E. Lonnoy, J. B. R. Matthews, T. K. Maycock, T. Waterfield, O. Yelekçi, R. Yu, & B. Zhou (Eds.), Climate change 2021: The physical science basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (pp. 1513–1766). Cambridge University Press. https://www.ipcc.ch/report/ar6/wg1/chapter/chapter-11/
Swiss Re Institute. (2025). Sigma 1/2025: Natural catastrophes: Insured losses on trend to USD 145 billion in 2025. https://www.swissre.com/institute/research/sigma-research/sigma-2025-01-natural-catastrophes-trend.html
Turner, J. (2023). Climate change physical risk in catastrophe modelling. Journal of Catastrophe Risk and Resilience, 1(1). https://journalofcrr.com/comment/01-01-turner/
Verisk. (2013, March 25). Modeling fundamentals: What is AAL? https://www.verisk.com/blog/modeling-fundamentals--what-is-aal-/
Verisk. (2017). Dealing with disasters: A lesson from Hurricane Andrew. https://www.verisk.com/verisk-review/fall-2017/dealing-with-disasters-a-lesson-from-hurricane-andrew/
Weiss. (2025). Uninsurable: Florida’s home insurance collapse signals national trend. Weiss Ratings.
World Wildlife Fund. (2026, January 15). Escalating extreme weather risks push US insurance system to breaking point. https://www.worldwildlife.org/news/press-releases/escalating-extreme-weather-risks-push-us-insurance-system-to-breaking-point/
Yale Law Journal. (2025, December 3). The uninsurable future: The climate threat to property insurance, and how to stop it. https://yalelawjournal.org/essay/the-uninsurable-future-the-climate-threat-to-property-insurance-and-how-to-stop-it
