Building a Modern Catastrophe Model: Open Data, Open Methodology

Every catastrophe model ever built has exactly three modules. AIR Worldwide (now Verisk) has them. RMS (now Moody's) has them. CoreLogic has them. The academic literature defines them. They are:

Module 1: Hazard. What physical intensity hits this location? For an earthquake, this is Modified Mercalli Intensity or peak ground acceleration. For a hurricane, wind speed in knots and storm surge height in feet. For a wildfire, fire perimeter proximity and ember exposure. The output is a number: intensity at a point.

Module 2: Vulnerability. Given intensity X, what percentage of the structure's value is damaged? A wood-frame single-family home (Hazus code W1) at MMI VIII experiences approximately 31% mean damage ratio. A reinforced concrete building (C1) at the same intensity experiences approximately 8%. These are damage functions — lookup tables that map intensity to damage by construction type.

Module 3: Financial. Damage ratio multiplied by replacement value equals dollar loss. A $400,000 home with a 31% damage ratio produces a $124,000 loss. Aggregate thousands of properties, account for deductibles, limits, and reinsurance terms, and you have portfolio-level loss estimates. Run them through a probability distribution and you have exceedance probability curves, AAL, and PML.

The Black Box Problem

When AIR and RMS built their cat models in the early 1990s, seismic data was hard to obtain, computing was expensive, and the modeling expertise resided in a handful of specialized firms. Proprietary models made sense. Insurers paid millions per year for model licenses because no alternative existed.

In 2026, the inputs are open. USGS publishes the National Seismic Hazard Model with full source code and data files. NOAA publishes hurricane track archives, storm surge models (SLOSH), and flood gauge data in real time. NIFC publishes wildfire perimeters as GeoJSON. FEMA publishes the complete Hazus vulnerability methodology, including every damage function table, as a free download.

The modeling expertise is published. Gutenberg-Richter frequency distributions, Atkinson & Wald intensity prediction equations, Knaff & Zehr wind-pressure relationships — all peer-reviewed, all cited thousands of times, all available to anyone who reads the journals.

What remains proprietary at incumbent firms is scale (decades of engineering investment), calibration (proprietary claim data for validation), and market position (regulatory acceptance). The methodology itself is not secret. It is published science.

The Calchis Approach

Calchis implements the three-module architecture using exclusively open data and published methodologies. Model parameters are stored in a structured database with source_description fields citing the originating publication, table number, and page.

Hazard module: USGS NSHM 2023 for seismic hazard. IBTrACS for hurricane frequency. NOAA SLOSH for storm surge. NIFC perimeters for wildfire. FEMA flood zones for flood. Each data source is documented in docs/data-lineage.md with provenance, update frequency, and license.

Vulnerability module: FEMA Hazus damage functions (Earthquake Model v6.1; Hurricane and Flood Models v5.1). They are published by FEMA in the public domain with a documented technical basis. When an actuarial team asks "what vulnerability functions do you use?", the answer is "FEMA Hazus, the published standard" — with one qualification we state on the methodology page: the hurricane model publishes continuous per-building-type curves rather than a ratio table, so our hurricane ratios are a screening-level tabulation shaped to those curves.

Financial module: Microsoft Building Footprints for structure counts and areas (free, 130M+ US structures). ATTOM property data for insured values and construction types (Phase 5 integration). OASIS Loss Modeling Framework output format for compatibility with Moody's RMS and Verisk integration pipelines.

Parameter Governance

Model parameters should evolve as science evolves, and every change should be traceable. Each parameter Calchis scores with is a database row that carries its source description, a verification status (verified against the cited source, needs verification, or an estimate) and the URL it was checked against. A parameter that is not verified marks every score it touches as estimated, and the score says so. When a value changes it is superseded, not overwritten, and a database trigger records the old value, the new value, who changed it and why.

There is no automated literature pipeline. An earlier version of this post described a weekly Semantic Scholar ingest; that was removed, because a parameter should change when a person has read the source, not when a script has found a paper. Changes arrive by pull request with a citation. A reviewer can trace every score component to the row and the row to the publication.

Why This Matters for Acquisition

Verisk, Moody's, and Palantir evaluate cat modeling platforms on methodology credibility, data quality, and engineering rigor. A platform that shows its math — with traceable parameters, published sources, and documented validation — is more credible than one that hides behind proprietary labels. Transparency is not a weakness. In 2026, it is the competitive advantage.