Industries / Finance and business services / Insurance
Modeling catastrophes
Using computer simulations of thousands of possible storms, floods and earthquakes to estimate what disasters could cost before they happen.
5
Milestones achieved, from 1987 to 2018.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1987
The first catastrophe model
Karen Clark founded Applied Insurance Research (AIR) in 1987 and built the model that Insurance Journal says started the catastrophe modeling industry. AIR became AIR Worldwide and is now part of Verisk.
Credit
Karen Clark
1992
Andrew breaks the record
Before Hurricane Andrew, the largest US hurricane loss was about $4 billion, from Hugo in 1989. On the morning Andrew struck south of Miami on 24 August 1992, AIR's model put the loss at $13 billion; a reinsurer client bet Clark £5 it would not pass $6 billion. Insured losses reached $15.5 billion by the Insurance Information Institute's count, and some insurers became insolvent.
Lesson"Hurricane Andrew proved that past data is a poor gauge for future catastrophe exposure," in the words of the Insurance Information Institute.
Credit
Karen Clark and AIR, for the estimate
1995
Models put under review
The Florida Legislature created the Florida Commission on Hurricane Loss Projection Methodology to review hurricane models. Insurers filing home insurance rates in Florida must use methods the Commission finds accurate or reliable, and may not adjust them. Models it has found acceptable include those of Verisk, Moody's, Karen Clark & Company, Cotality, Applied Research Associates and the Florida Public Hurricane Loss Model.
Credit
Florida Legislature; Florida Commission on Hurricane Loss Projection Methodology
2018
Models checked against storms
The reinsurance broker JLT Re compared modeled losses with final losses for hurricanes since 2004 and found "an overriding trend towards significant loss underestimation". The misses were largest where losses spread into areas the models do not cover or understand well, and where complex storms struck dense cities. Models did relatively well for moderate wind losses.
LessonA model misses most where losses come from things it does not model.
Credit
JLT Re (David Flandro)
2018
An open model engine
In July 2018 the not-for-profit Oasis Loss Modelling Framework made all of its catastrophe modeling software open source and free to download from GitHub. It accepts models in a standard format from vendors, universities and research groups.
Credit
Oasis Loss Modelling Framework (chief executive Dickie Whitaker, chief technology officer Mark Pinkerton)
Watch for
Storms drawn by AI
Fathom, a Swiss Re subsidiary, says it uses diffusion models, a kind of generative AI, to produce tens of thousands of years of made-up weather events for a projected 2030 climate, trained on about 1,000 years of climate simulations. Verisk says it uses generative AI to model extreme wind and rain together. Fathom's scientific director, Oliver Wing, warned that such models can produce events that break basic physical laws.
Achieved whena model built on AI-generated events passes an independent review such as the Florida Commission's, or an outside party checks its losses against observed losses.
Claimed by
Fathom and Verisk, as reported in June 2026
Open
01
A scorecard for first estimates
A public table of every insured-loss range that modeling firms published in the first weeks after each major disaster since 2004, set beside the later loss figure. It reports how far each firm's first range missed, by firm and by type of event, and is updated after every major event.
The briefCredit · unclaimed02
Public tests for storm catalogs
Open code that runs one of the Florida Commission's hurricane checks on any public catalog of made-up storms: comparing the catalog's landfall counts in Florida, statewide and by region and by hurricane category, with the historical record. Publish results first for the STORM catalog and then for AI-generated catalogs as they appear.
The briefCredit · unclaimed
Open 01 · the brief
A scorecard for first estimates
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of every insured-loss range that modeling firms published in the first weeks after each major disaster since 2004, set beside the later loss figure. It reports how far each firm's first range missed, by firm and by type of event, and is updated after every major event.
Why it is still open
JLT Re's 2018 comparison was a one-off study of hurricanes. No continuing public scorecard of first estimates was found.
Built with what exists
modeling firms' press releases and trade-press archives such as Artemis and Reinsurance News; later loss figures as reported by PCS, PERILS, regulators or Swiss Re, with the source named for each; a language model to pull ranges and dates from press releases, with a person checking every row.
First step
Build the rows for the 2017 hurricanes Harvey, Irma and Maria from the firms' press releases.
Done looks like
At least 60 events, each first estimate with its date and link, and each later figure with its source. A median miss for each firm and event type, and a refresh within a month of each new major event.
Open 02 · the brief
Public tests for storm catalogs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Open code that runs one of the Florida Commission's hurricane checks on any public catalog of made-up storms: comparing the catalog's landfall counts in Florida, statewide and by region and by hurricane category, with the historical record. Publish results first for the STORM catalog and then for AI-generated catalogs as they appear.
Why it is still open
The Commission reviews modelers' trade secrets in closed meetings, so outsiders cannot see how a given catalog performed. AI-generated catalogs are new, and their makers warn they can produce events that are not physically possible.
Built with what exists
NOAA's HURDAT2 Atlantic hurricane database; the STORM dataset of 10,000 years of synthetic tropical cyclone tracks, released into the public domain; CLIMADA, an open-source risk platform from ETH Zurich; Python.
First step
Read the Commission's meteorological standards and code the annual landfall count for HURDAT2 alone.
Done looks like
Published code that reproduces the landfall comparison for HURDAT2 against STORM, statewide and for each region, with results that match a hand count for at least three regions, and results posted for at least two catalogs.