Industries / Finance and business services / Insurance / Reinsurance
Catastrophe models
Computer simulations of thousands of possible disasters, used to estimate how much a reinsurer could lose and how often.
7
Milestones achieved, from 2005 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2005
The water the models missed
When Katrina struck in August 2005, the modeling firms' first estimates of insured losses were $15 billion to $25 billion; on 10 September RMS raised its estimate to as much as $60 billion. Much of the damage came from flooding and storm surge, which some models did not include. AIR said its model covered surge but not inland flooding, since most insurers do not cover it.
LessonLosses from a cause the policies exclude, here flood water, can still reach insurers through disputes and courts.
Credit
RMS, AIR Worldwide and EQECAT, for the estimates; Insurance Journal, for the account
2011
One model update, big swings
RMS released version 11 of its US hurricane model in February 2011, with higher inland wind speeds, weaker buildings and more storm surge. One mid-Atlantic insurer saw its probable maximum loss, the loss it plans for in a severe year, rise by 150% to 200%, said John DeMartini of Towers Watson: "They were overly reliant on models, and many of them were overly reliant on one model." Guy Carpenter said this change and an AIR update helped make the June 2011 Florida renewals "volatile".
LessonA firm that relies on one model can see its view of its own risk change overnight.
Credit
RMS (Robert Muir-Wood); Business Insurance and Guy Carpenter, for the accounts
2011
A flood outside the models
Floods in Thailand in late 2011 swamped six industrial estates in Ayutthaya and Pathum Thani. A World Bank assessment put total damage and losses at THB 1.43 trillion (about $46.5 billion), with manufacturing bearing about 70%. Insured loss estimates then doubled to $15 billion to $20 billion, and Fitch, a rating agency, named "the limited accuracy of catastrophe model simulations for the Asia-Pacific region" as one possible reason.
LessonA flood that Fitch called a one-in-50-year event became one of the year's largest losses because so many factories stood together in a few flooded estates.
Credit
World Bank, for the assessment; Fitch
2013
An open earthquake engine
The Global Earthquake Model (GEM) Foundation described the OpenQuake engine, its open-source software for earthquake hazard and risk, in a paper in Natural Hazards published on 12 March 2013. The code is public under the AGPL-3.0 licence, and GEM publishes regional hazard models that run in it.
Credit
GEM Foundation
2021
A modeler joins a rating firm
Moody's agreed in August 2021 to buy RMS from Daily Mail and General Trust, which had owned it since 1998, for about $2.0 billion. RMS expected revenue of about $320 million for its year to September 2021.
Credit
Moody's; RMS
2025
AI beats official track forecasts
Google DeepMind's tropical cyclone model went public through Google Weather Lab in June 2025. In its verification report on the 2025 Atlantic season, the US National Hurricane Center found that the model's ensemble average "slightly outperformed" its own official track forecasts from 12 to 72 hours ahead. Gallagher Re, a reinsurance broker, noted that the AI models were weaker on storm strength.
Credit
Google DeepMind; US National Hurricane Center, for the verification
2025
Storms, fires and floods dominate
Swiss Re counted $107 billion of insured natural catastrophe losses in 2025, with no major US hurricane landfall. A record 92% came from "secondary perils" such as wildfires, floods and severe convective storms, the thunderstorms that bring hail and damaging winds; the Los Angeles wildfires alone cost insurers about $40 billion. Karen Clark & Company, a modeler, notes that the first catastrophe models were built mainly for rare, severe events such as hurricanes and earthquakes.
LessonFrequent, mid-sized events now cause most insured disaster losses, and models built for rare large ones need different methods.
Credit
Swiss Re Institute; Karen Clark & Company
Watch for
Models updated in days
Karen Clark & Company says catastrophe models that combine physics with AI could be updated "in days rather than months". It says its severe convective storm model processes more than 30 gigabytes of satellite, radar and weather data a day, and that since 2018 it has compared its event estimates with actual claims.
Achieved whenan outside party, such as a state model review body or a client publishing a back-test, confirms that a model updated in days matched later claims at least as well as the version it replaced.
Claimed by
Karen Clark & Company, August 2026
Open
01
An open flood model for Thailand
An open-source catastrophe model for river flooding of Thailand's industrial estates, run in the open Oasis Loss Modelling Framework, that anyone can inspect and rerun.
The briefCredit · unclaimed02
A log of model changes
A public log of every major release of the commercial catastrophe models since 2005, with what changed and how far modeled losses moved, by peril and region, as stated by the modeler, brokers or clients, each with a source.
The briefCredit · unclaimed
Open 01 · the brief
An open flood model for Thailand
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open-source catastrophe model for river flooding of Thailand's industrial estates, run in the open Oasis Loss Modelling Framework, that anyone can inspect and rerun.
Why it is still open
Fitch named the limited accuracy of Asia-Pacific models as one reason Thai loss estimates doubled in 2011. I found no open flood model for Thailand.
Built with what exists
the Oasis Loss Modelling Framework (BSD-3-Clause licence); the European Commission Joint Research Centre's global river flood hazard maps, which give water depth for floods from 1 in 10 to 1 in 500 years; OpenStreetMap outlines of the estates; the JRC's global depth-damage curves; Python.
First step
Trace the six estates from OpenStreetMap and lay them over the JRC 1-in-50-year flood depth map.
Done looks like
A published model that a second person can rerun to the same figures; a 1-in-50-year scenario that floods all six estates the World Bank names; and a modeled manufacturing loss for that scenario set beside the World Bank's 2011 figure, with the gap explained.
Open 02 · the brief
A log of model changes
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public log of every major release of the commercial catastrophe models since 2005, with what changed and how far modeled losses moved, by peril and region, as stated by the modeler, brokers or clients, each with a source.
Why it is still open
Karen Clark says updates to the traditional models "can shift loss estimates 20% to 30%", so brokers and reinsurers each rerun their portfolios to find what changed, "because the modeling companies don't make it easy". RMS's 2011 update moved one insurer's planning loss by 150% to 200%. No public log was found.
Built with what exists
modelers' release notes and press releases; brokers' renewal reports; trade-press archives; an AI model to pull figures from the text, with a person checking every row.
First step
Build the rows for the 2010 AIR and 2011 RMS US hurricane updates from Guy Carpenter's 2011 renewal reports and trade press.
Done looks like
At least 40 releases from at least four modelers, each with a date, the perils changed and at least one published figure for the change, every row linked.