You have been in your cat modelling role for six months. You are presenting a model run to a senior property underwriter, fifteen years in the market, started at Lloyd's before the Equitas run-off. You show the OEP table. The 1-in-100 number is higher than he expected.
He leans back and says something you will hear variants of for the rest of your career: "These models. They didn't exist when I started. We wrote earthquake and hurricane on gut feel and market knowledge, and we made money. Now there's a machine telling me what a risk is worth, and half the time it's wrong."
He is not entirely wrong. The models have been wrong, sometimes catastrophically. But the reason those models exist at all, and the reason they are now central to how the market works, is rooted in a specific sequence of events that began in South Florida on 24 August 1992, and understanding that history tells you something important about what the models are for and what they are not for.
Before 1992, the insurance and reinsurance market managed catastrophe risk primarily through experience, instinct, and market convention. Underwriters at Lloyd's and the London company market priced windstorm and earthquake risks based on a combination of historical loss records, treaty terms negotiated over years of market cycles, and the collective wisdom of senior practitioners who had lived through previous major events. This approach had worked tolerably well for decades.
The tools available were almost entirely deterministic. A deterministic approach to catastrophe risk means estimating the loss from a specific defined scenario, "what would happen to our portfolio if a repeat of the 1906 San Francisco earthquake occurred?" or "what is our exposure if a Category 4 hurricane makes landfall at Miami?" These scenario estimates were based on mapped hazard zones, engineering judgment, and historical analogy. They told underwriters roughly how bad a specific named event might be. What they could not tell them was how likely that event was, how often they should expect to exceed any given loss threshold, or what the full distribution of possible annual outcomes looked like.
Critically, the market's mental model of catastrophe risk was anchored to events that had actually occurred. The largest hurricane losses on record before 1992 were from Hugo (1989, USD 5 billion insured) and Frederic (1979, USD 1.7 billion insured). The market implicitly treated these as rough upper bounds, not because anyone consciously believed a larger event was impossible, but because the human tendency to anchor on recent experience is powerful and the incentive to price for events outside that experience is limited when your competitors are not pricing for them either.
When every underwriter in the market prices catastrophe risk using the same historical data and the same loss experience, they all underprice the same tail risks simultaneously. No individual underwriter has a financial incentive to price more conservatively than the market, doing so loses business to competitors who price lower. This collective underpricing of tail risk is not incompetence; it is a structural feature of competitive insurance markets. It is also exactly what cat modelling was designed to correct, by making the full probability distribution of possible losses visible, regardless of whether those events had ever occurred.
Hurricane Andrew made landfall near Homestead, Florida at approximately 4:52 AM local time on 24 August 1992. At landfall, the storm carried sustained winds of 165 mph, making it a Category 5 system at its most intense, though it had weakened slightly to the high end of Category 4 by landfall. Andrew tracked due west across the southern tip of the Florida peninsula, through Homestead and Florida City, and into the Gulf of Mexico before making a second landfall in Louisiana two days later.
The Florida damage was unlike anything the modern insurance industry had experienced. Homestead was effectively levelled, structurally destroyed across a wide swath of the city. The mobile home communities that had expanded across South Florida's suburban fringe in the 1970s and 1980s were wiped out almost entirely. Entire suburbs of Miami-Dade County lost their roofs. The combination of extreme wind speeds, inadequate building codes, and decades of unchecked suburban expansion into hurricane-exposed territory produced losses that bore no resemblance to any previous event.
When the loss estimates began arriving in the days after Andrew, the market's first reaction was disbelief. Initial industry estimates were in the USD 7–8 billion range, large, but manageable. As the full scale of the damage became clear over the following weeks, the estimates climbed: USD 10 billion. USD 15 billion. USD 20 billion. The final insured loss settled at approximately USD 27 billion, more than five times the previous largest insured hurricane loss in history, and more than the entire Florida homeowners insurance market had collected in premiums over the preceding decade.
Eleven property insurers became insolvent as a direct result of Andrew losses. Dozens more faced capital crises. The Florida insurance market itself faced a genuine threat of collapse, reinsurers who had been providing capacity to Florida homeowners carriers simultaneously refused to renew at any price, leaving the market without the capital base to continue operating. The state of Florida was forced to create the Florida Hurricane Catastrophe Fund, a state-backed reinsurance facility, as an emergency intervention to prevent market failure. Andrew did not merely test the industry's capital adequacy. It broke the industry's basic assumption that its pricing had been adequate.
Andrew did not merely prompt the insurance industry to collect more historical data. The lesson of Andrew was precisely that more historical data would not have helped, Hugo and the prior historical record were the data the market had, and they had not produced an adequate estimate of tail risk. What was needed was a fundamentally different approach: one that could estimate the probability of events beyond the range of the historical record by modelling the physics of the hazard rather than simply averaging past losses.
Within months of Andrew, three firms emerged to build probabilistic catastrophe models. Each approached the problem with a slightly different emphasis and team composition, establishing a competitive dynamic that has characterised the vendor model market ever since.
The three-vendor competitive structure that emerged post-Andrew has defined the cat modelling market for three decades. While the specific ownership and branding has changed, EQECAT absorbed into CoreLogic, RMS acquired by Moody's, AIR by Verisk, the fundamental dynamic has remained: a small number of vendors building and maintaining complex proprietary models, with the insurance and reinsurance market dependent on their outputs for pricing, capital, and regulatory purposes. The implications of this concentration, for model transparency, for diversity of approaches, and for the risk of correlated modelling errors, remain actively discussed in the industry today.
The fundamental methodological contribution of the first generation of cat models was the shift from deterministic scenarios to stochastic catalogues. This is the most important conceptual development in the history of the discipline, and understanding it is essential for understanding what is a cat model and what it produces.
The stochastic approach answers a question the deterministic approach cannot: not "what happens in this scenario?" but "what is the probability distribution of all possible annual losses?" This distribution, usually expressed as an Exceedance Probability (EP) curve, is what enables the modern insurance market to price risk, set reinsurance programmes, allocate capital, and satisfy regulatory requirements. None of these functions were possible with scenario-based deterministic assessment alone.
The shift from deterministic to stochastic is also a shift in what questions underwriters can ask. Before cat models, an underwriter could ask: "what is my exposure to a Category 5 at Miami?" After cat models, the question becomes: "what is my 1-in-100 year loss?" and "what is my expected annual loss?" and "what is my maximum foreseeable loss at the 1-in-250 year return period?" These are qualitatively different questions, and the shift to asking them represents a fundamental change in how the market understands and manages catastrophe risk.
The first commercial hurricane and earthquake models are built and validated against Andrew and Northridge (1994). Lloyd's and the major London market insurers begin requiring cat model output for risk assessment and reinsurance pricing. The models are computationally limited, running a large portfolio takes hours or days on dedicated workstations. The vendor-client relationship is one of tool licensing, not data analytics. Most practitioners receive model output as a black box with limited ability to interrogate assumptions.
Cat models expand to cover European windstorm, Japan earthquake and typhoon, and the first flood models. The Lloyd's market establishes formal catastrophe reporting requirements, the Realistic Disaster Scenarios (RDS) that become mandatory capital assessment tools. Hurricane Katrina (2005) delivers the first major test of the models at scale, and reveals material weaknesses in storm surge modelling, levee reliability assumptions, and demand surge estimation. The gap between model output and actual loss for Katrina drives the first major wave of model revision and the recognition that model uncertainty is not merely a fringe concern.
Tōhoku (2011) produces the largest insured earthquake loss in history and drives complete revision of Japan subduction zone models, the existing models had not included an event of this magnitude as credible. Superstorm Sandy (2012) exposes compound risk modelling gaps. The 2010s see major advances in model resolution, street-level geocoding, high-resolution flood hazard mapping, and building-level vulnerability functions become standard. The decade also sees the first serious attention to "secondary perils", severe convective storms, wildfire, flash flood, as the relative importance of these perils in global loss totals begins to grow relative to the primary perils the models were originally built for.
The 2020s confront the cat modelling industry with two simultaneous challenges. The first is non-stationarity, the recognition that climate change is altering the frequency and severity distributions of multiple perils in ways that make historical calibration data increasingly unreliable. California wildfire, Atlantic hurricane intensification, and extreme heat are all demonstrating loss patterns that historical models did not predict. The second is the emergence of cyber as a cat peril, the NotPetya attack (2017) and CrowdStrike outage (2024) demonstrated that technology concentration risk can produce correlated catastrophe-scale losses with no geographic footprint, requiring entirely new modelling approaches. The industry is simultaneously revising existing models for a changing climate and building new model categories for new risk types.
The vendor cat model market has consolidated significantly since the early 1990s. The three original firms have been absorbed into larger data and analytics businesses, and the market is now dominated by two primary vendors, Moody's RMS and Verisk AIR, with a growing ecosystem of specialist and regional model providers alongside them.
| Vendor | Platform | Parent | Primary Strengths |
|---|---|---|---|
| Moody's RMS | RiskLink | Moody's Corporation | Earthquake, European windstorm, North Atlantic hurricane. Historically strong in seismic engineering. Expanding into climate risk analytics. Broad global peril coverage. |
| Verisk AIR | Touchstone (AIR) | Verisk Analytics | North Atlantic hurricane, severe convective storm, Japan typhoon and earthquake. Actuarial methodology leadership. |
| Sequel / Verisk | Sequel Impact | Verisk Analytics | Aggregation management and accumulations. Widely used in Lloyd's market for managing aggregate exposure across multiple perils and regions rather than individual risk assessment. |
| Corelogic | Hazard HQ | Independent | High-resolution U.S. hazard data, property data, and claims analytics. Increasingly important as a data layer underlying other vendors' models. |
| JBA Risk Management | JBA Flood | Independent | Global and regional flood modelling specialist. Often used where the primary vendors' flood models are considered less strong, particularly for European and Asian flood risk. |
| Cytora / CyberCube / Kovrr | Various | Various | Emerging cyber cat model providers, building probabilistic frameworks for a peril that the primary vendors have been slower to address at full depth. |
Different vendors produce materially different loss estimates for the same portfolio, sometimes by 50% or more for the same peril and geography. This is not because one model is right and one is wrong. It reflects genuine scientific uncertainty about hazard frequency, vulnerability functions, and financial assumptions. Multi-vendor analysis, running a portfolio through both primary vendors and comparing the outputs, is standard practice at Lloyd's and among sophisticated reinsurers precisely because the divergence between vendors is informative about the range of credible outcomes. This is covered in depth in the Intermediate course Module 3.
The history of cat modelling is simultaneously a history of the discipline's achievements and a catalogue of its failures. Every major catastrophe since Andrew has exposed something the models did not adequately capture, and the subsequent wave of model revision has addressed that specific gap, only for the next event to reveal a different one. This pattern is not evidence that cat models are unreliable; it is evidence that they are calibrated to the events they have been through, and struggle systematically with events outside that experience.
The models assumed Florida's building stock met the codes on the books. Many buildings, especially those built before the 1992 reforms that Andrew itself prompted, were constructed to lower standards than the codes specified. The models learned: physical construction quality matters, not just the code year.
The primary Katrina models treated the levee system as providing reliable protection up to its design level. The catastrophic levee failures that flooded 80% of New Orleans were not in any model's scenario set. Storm surge was undermodelled for the specific geographic configuration of the Gulf Coast. The models learned: infrastructure reliability must be probabilistic, not binary.
The existing Japan subduction zone models had not included a magnitude 9.0 event as a credible scenario for the Tohoku region. The models learned: geological science must explicitly account for events at the tail of the magnitude distribution, not just those observed in the modern instrumental record.
Harvey's catastrophic losses came primarily from rainfall-driven flooding across the Houston metropolitan area, a loss type that the existing hurricane models' wind and surge modules did not capture. The models learned: compound perils require compound models, not peril-by-peril single-hazard approaches.
The most fundamental current challenge: models calibrated to historical event data may systematically underestimate future risk in perils where climate change is altering the frequency and severity distribution. This is not a gap the industry has yet fully solved, it is an active area of model development and the source of significant uncertainty in current outputs.
The senior underwriter in the opening scenario was right that the models have been wrong. But the alternative, pricing catastrophe risk without any probabilistic framework, produces systematic underpricing of tail risk, as Andrew demonstrated definitively. The models are imperfect tools for a genuinely difficult problem. The practitioner's skill is knowing where each model's limitations lie, which is exactly what this course is building toward.
A specific defined event used to estimate portfolio loss, "what happens if a Category 5 hurricane hits Miami?" Produces a single loss estimate for one scenario without assigning a probability to that scenario.
A large synthetic set of simulated events covering the full distribution of possible hazards, each with an assigned annual rate. Running a portfolio through the full catalogue produces the EP curve, the core output of a probabilistic cat model.
A model that estimates the full probability distribution of possible annual losses rather than the loss from a single scenario. The output is an EP curve, not a single number. The standard approach in commercial cat modelling since 1992.
Exceedance Probability Curve, a mapping from every possible loss level to the annual probability of exceeding it. The primary output of a probabilistic cat model. Covered in detail in Foundation Lesson 3.1.
A suite of specific catastrophe scenarios defined by Lloyd's, used for mandatory capital assessment and regulatory reporting. Every Lloyd's syndicate must estimate its exposure to each RDS event. Covered in Foundation Lesson 3.3.
The condition in which the statistical distribution of a hazard is changing over time, meaning historical data is not a reliable guide to future risk. Climate change is making non-stationarity a first-order challenge for cat model calibration across multiple perils.
The range of credible loss estimates that arise from legitimate differences in scientific assumptions between vendors, not from data errors, but from genuine uncertainty about hazard, vulnerability, and financial parameters. The divergence between vendor models is a quantification of this uncertainty.
The increase in repair costs in the aftermath of a major catastrophe due to simultaneous excess demand for labour and materials across an entire region. First systematically quantified after Andrew and now a standard component of financial module loss estimation.
Five questions · 4 of 5 correct (80%) to pass