How wildfire risk is changing insurance underwriting

In 2023, U.S. natural-disaster losses exceeded $100 billion, according to the Congressional Budget Office, and insurance covered about 70% of that total. That leaves a protection gap large enough to matter even to companies that do not sell insurance, because the same forces that strain carriers also change where capital is willing to stay.

Climate and wildfire risk are not only raising expected losses, they are also increasing uncertainty about future losses, and that is what reprices insurance. Broad territory scores can no longer do the job alone, because two homes on the same street can have very different exposure once structure separation, roof condition, nearby vegetation, access, and mitigation are inspected at the parcel level.

The practical response is not a single model. It is a workflow that combines geospatial data, catastrophe models, and underwriter review, so the carrier can separate a genuinely high-risk property from one that was flattened by a coarse regional assumption.

Wildfire pricing breaks when the map is too coarse

The recent wildfire losses show why pricing by broad geography is breaking down. According to Swiss Re, the Palisades and Eaton fires generated an estimated $40 billion in combined insured losses by July 2025, the largest insured wildfire losses it had recorded. Earlier, Moody’s RMS estimated eventual insured losses at approximately $20 to $30 billion, which reflects different event assumptions, settlement timing, and definitions of loss.

The Congressional Budget Office has tied rising disaster severity and rising uncertainty to reduced insurance availability and affordability, which is exactly what underwriters are seeing when a territory that once looked manageable starts to produce surprises that models did not price well enough. The problem is mostly operational. Traditional workflows often start with an address-level peril score or a regional classification, then layer rules and judgment on top. That is efficient until the coarse score hides the one thing that matters, such as whether a home is separated from neighboring structures, surrounded by vegetation, or exposed because the roof and exterior are not built for the hazard.

The decision improves when the property becomes the unit of analysis

Property-level geospatial analysis changes the unit of work. The Society of Actuaries and Milliman case study looked at Scripps Ranch, Grizzly Flats, and Fountaingrove and used building footprints, vegetation, and structure separation to improve wildfire underwriting at the property level. The point is not that imagery becomes truth by itself. The point is that it can distinguish homes that regional models may classify together but that do not actually face the same loss profile.

That distinction matters because wildfire risk is spatial, not abstract. Slope, aspect, elevation, vegetation continuity, road access, hydrant proximity, historical fire perimeters, and distance to wildland fuel all affect exposure, and they do not show up cleanly in a postal code. If those inputs live in separate systems, the underwriter sees a score without the evidence needed to defend the decision.

At that point, the carrier is not really lacking data. It is lacking a usable identity layer that joins parcel records, building footprints, imagery, hazard history, and claims into one property dossier. That dossier is what lets a review team explain why one property is acceptable, another needs inspection, and another should be referred or declined under filed rules.

The workflow is where carriers make the economics real

The operating model matters as much as the model itself. A score that cannot be reviewed, justified, or challenged will either be ignored or forced through manual exception handling, and both outcomes are expensive. Ambiguous addresses, stale mitigation records, unusual construction, and incomplete property attributes are the places where underwriters already spend time, which is why the sensible design is to route only those cases that need attention.

In practice, the useful architecture is straightforward. Source data lands in a property identity layer that joins parcel, footprint, imagery, terrain, vegetation, weather, fire-history, infrastructure, and claims records. A feature service then calculates slope, separation, fuel proximity, and access indicators, while a catastrophe model or risk model estimates ignition, spread, intensity, and loss. The underwriter then receives a dossier with the evidence behind the score, not just the score itself.

That is also where retrieval helps. A review queue can surface the imagery date, parcel boundary, model version, mitigation records, inspection photos, and the underwriting rule that applies, so the reviewer can see what drove the outcome and what would change it. This does not replace professional judgment. It gives judgment a record to stand on.

Stage 1 is source ingestion, where parcel, footprint, imagery, weather, vegetation, fire-history, infrastructure, and claims data are collected into an event stream and document store. Stage 2 is property identity resolution, where addresses, parcels, and building records are matched into one geospatial record with versioned lineage. Stage 3 is feature engineering, where terrain, separation, fuel proximity, access, and mitigation signals are derived for risk scoring. Stage 4 is risk modeling, where catastrophe models and rule engines combine hazard and vulnerability assumptions into a property-level assessment. Stage 5 is review and decision support, where a dossier is presented to the underwriter with source evidence, conflict flags, and a case queue for exceptions. Stage 6 is outcome feedback, where inspections, claims, and post-event observations are written back for validation and rule updates.

Where the approach fails is also where it becomes useful

This is not a magic fix, and it fails in predictable ways. Nonstationarity means historical loss patterns are less reliable because temperature, drought, wind, fuel conditions, and fire behavior are changing. Model disagreement also matters, since different catastrophe models use different hazard, vulnerability, and financial assumptions. A carrier that treats one score as final will still be surprised.

There are also costs. High-resolution imagery, property matching, model governance, and human review all add operational overhead, and regulatory constraints can slow how quickly new data enters production. Consumer protection rules, explanation requirements, and rate filing obligations do not disappear because the model is better. If anything, they make it more important to know which data source drove the decision and whether the decision can be explained in plain language.

That is why the best use of automation here is selective. Low-risk, well-supported properties can be auto-cleared or renewed under standard terms, while ambiguous or high-risk cases go to a specialist underwriter with the evidence already assembled. The point is not to remove people from the loop. It is to keep them focused on the cases where their judgment actually changes the outcome.

Insurance is becoming a data quality problem with financial consequences

The larger lesson reaches beyond property insurance. Whenever a company faces growing uncertainty, coarse segmentation stops being a back-office convenience and becomes a capital problem. If the data layer cannot distinguish risk at the level where decisions are made, then pricing, reserves, and coverage rules all drift away from reality at the same time.

That is why this topic belongs on an operations and technology agenda, not only on an actuarial one. The work is not glamorous: property identity, geospatial joins, evidence retrieval, review queues, audit logs, and model comparison. But that is where the economics of wildfire risk are now being decided, one parcel at a time.

If you are deciding whether to keep this kind of work in spreadsheets and manual review, or move it into a governed workflow with evidence attached to every exception, QueryNow can build the workflow in your environment in two weeks. Tell us the workflow you want gone, then start at /build.