How Wildfire Catastrophe Models Work

How Wildfire Catastrophe Models Work

A wildfire catastrophe model simulates many plausible events, estimates damage to a portfolio, applies insurance and reinsurance terms, and produces a distribution of financial loss.

It does not predict the exact date, ignition, path, or loss of the next fire.

That distinction explains why "the model missed this fire" can be either a valid criticism or a category error. The answer depends on what the model was designed to estimate, which version and data were used, and how the result supported the decision.

flowchart LR
    A["Hazard event set"] --> B["Vulnerability relationships"]
    B --> C["Exposure portfolio"]
    C --> D["Policy and reinsurance terms"]
    D --> E["Event losses"]
    E --> F["Loss distribution and decision outputs"]

The hazard module creates possible fires

The hazard component represents many potential events. For wildfire, a model may use assumptions and data about ignition, weather, wind, fuels, moisture, terrain, spread, intensity, suppression, and annual occurrence.

Each simulated event has a location, footprint, intensity, path, and probability or event rate. The catalog is not a list of scheduled future fires. It is a way to sample plausible catastrophe conditions.

The NAIC property catastrophe-model overview describes the hazard module as an event catalog with information about strength, size, location, path, and annual probability.

The vulnerability module estimates damage

Hazard intensity is not financial loss.

The vulnerability component connects local event conditions to expected damage for structures with particular characteristics. Construction, occupancy, roof, vents, siding, age, defensible space, maintenance, neighboring fuels, and other attributes may affect the relationship where the model and data support them.

A vulnerability function can be uncertain or incomplete. A property record can also be wrong. An insurer should not treat a modeled characteristic as a verified inspection fact.

Engineering evidence, claims experience, laboratory tests, post-fire analysis, and expert judgment can inform vulnerability. The source and calibration matter.

The exposure module describes the portfolio

Exposure data identifies what could be damaged and where. It can include geocoded locations, construction, occupancy, values, contents, business interruption, policy count, and other characteristics.

Address quality matters. So do valuations and building attributes. A misplaced coordinate, stale replacement value, generic construction code, or missing secondary characteristic can change estimated loss.

A portfolio model is not automatically a property score. The intended unit of analysis and decision should be stated.

The financial module applies contracts

Physical damage is translated into insured loss through policy terms. Deductibles, limits, sublimits, coverage, valuation, exclusions, occurrence rules, and reinsurance can change the financial result.

Two insurers can model the same fire and building damage but estimate different net loss because their portfolios and contracts differ.

The NAIC describes the financial module as applying policy conditions and reinsurance to event losses. It then aggregates simulated outcomes into a loss probability distribution.

Read the outputs as distributions

Average annual loss is the mean modeled loss across the event set and period definition. It is not the amount expected in every calendar year.

An exceedance probability curve connects a loss level to the modeled probability that loss exceeds it. Occurrence output focuses on the largest event in a year. Aggregate output can include all modeled events in the year.

A return-period loss is often misunderstood. A modeled 1-in-100-year loss does not mean one such event will happen on a 100-year schedule. It describes an annual probability or quantile under the model and version.

Probable maximum loss is also not one universal number. The probability basis, occurrence or aggregate view, gross or net treatment, time period, portfolio, and confidence should be stated.

Why models disagree

Models can use different event catalogs, physical science, vulnerability relationships, data enrichment, demand-surge assumptions, financial engines, and calibration.

A version update can change results because new events, research, resolution, claims data, or engineering evidence becomes available. The exposure portfolio may also change between runs.

Meaningful comparison requires the same portfolio, valuation date, terms, settings, output definition, and review design. A vendor ranking based on unmatched public numbers is not reliable.

How insurers use the output

Catastrophe-model output can inform underwriting, pricing, portfolio aggregation, concentration management, mitigation, reinsurance purchasing, capital, solvency, and scenario analysis.

The model does not make those decisions alone. Data quality, risk appetite, regulation, market conditions, human review, financial strength, strategy, and legal duties matter.

For mitigation, ask whether the model recognizes the exact feature and evidence. A reduced modeled loss is not a guarantee that a building survives or a premium falls.

California's model rules changed

California finalized wildfire catastrophe-model regulations in December 2024. The Department completed review of the first wildfire model in July 2025, as described in its model-review announcement.

The Department ties certain model or reinsurance-cost use to coverage commitments in wildfire-distressed areas. The exact regulation, reviewed version, filing, purpose, and approval control.

California's public wildfire model is separate. The Department's public-model page says SB 429 authorized the program in 2025. As of July 28, 2026, the Department had requested university expertise and anticipated a grant in October 2026. A completed operating public model did not yet exist.

Ask for a model claim record

Before accepting a wildfire-model number, ask for model name, vendor or owner, version, run date, peril, geography, event set, exposure data, valuation, vulnerability assumptions, financial terms, gross or net treatment, output, return period, sensitivity, validation, limitations, intended use, and reviewer.

A model can be useful without being certain. Its value comes from making a defined risk distribution more explicit, then keeping the assumptions and uncertainty attached to the decision.

This page was developed with AI assistance from the E053 transcript and linked regulator sources, then structured for catastrophe-modeler, actuarial, engineering, data, California regulatory, and editorial review. It does not provide property scoring, rate, capital, reinsurance, or actuarial advice.