Wildfire vulnerability for insurers and lenders
WildfireVuln estimates the chance a home is destroyed once fire reaches its neighbourhood, and how much each retrofit changes that. It is calibrated on homes and commercial and infrastructure buildings that CAL FIRE inspected after fires between and . It gives a mitigation credit only where the inspection evidence holds up.
Why most wildfire scores are wrong in the same way
After a fire, CAL FIRE inspectors walk every street inside the perimeter and record each structure: damage category, roof, eaves, vents, siding, decks and fences. It is the best open record of wildfire losses anywhere. It is also recorded after the fact.
A wooden fence attached to a house that burned is gone, so the form says no fence. Stucco walls stand in the ash, so the form says stucco. Eaves on a burnt house cannot be seen, so the form says unknown. A model trained on these records learns how the survey was filled in.
A model fitted naively to the survey concludes that an attached wooden fence lowers the chance of losing the house by .
We checked every feature three ways and only give credit where the answers agree. We also stopped the model from reading the word "unknown". A gradient-boosted model that is allowed to use it scores an AUC of on withheld fires. Once "unknown" cannot be told apart from a real value, the same model scores . Restricted to features that are recorded the same way before and after a fire, it scores ; ours scores .
Share of inspected homes with each recorded value, –. The fence and deck rows are the evidence-survival effect: a combustible attachment that burned is recorded as absent. "Unknown" eaves are three times as common on destroyed homes.
Area under the ROC curve within each withheld fire, mean over fires. Hatched bars use survey fields whose recorded value depends on the outcome; they score well against the survey and cannot be used to price a house before a fire.
Mitigation credits, graded by evidence
Each estimate compares homes in the same fire that faced the same local fire, measured as the share of neighbours destroyed within 100 m, and differ in one feature. Intervals are 90% bootstrap ranges, resampling whole fires. A credit is supported when the full survey and the complete records agree, the standing homes do not contradict it, and the direction matches fire-lab testing.
From vulnerability to premium
WildfireVuln prices the house. The chance that fire reaches it each year comes from your hazard model. Annual expected loss is that chance times the expected damage ratio times the dwelling limit. The planner uses the home on the inspection card above.
The same home before and after retrofits, from a mild fire (10th percentile of the fires) to a severe one (90th). Fire-to-fire variation is larger than any retrofit, which is why the credit is a relative change and the hazard model carries the severity.
During the event
Before a fire, nobody knows how severe it will be, so the first estimate is wide. As inspection teams report, WildfireVuln updates its estimate of the fire's severity and narrows the range. Each replay below uses a model trained without that fire, and feeds it inspections in the order they were entered.
Destroyed share of inspected structures, homes and commercial, in 250 m cells. Cells with fewer than 5 structures are left out, so no single building can be located. Basemap © Esri and OpenStreetMap contributors.
Commercial real estate and fixed assets
Inspectors record every structure in the footprint, not only homes. The record holds commercial, institutional and infrastructure buildings from the same fires as the homes. WildfireVuln fits them in one model: commercial buildings get their own building-feature effects, and homes and commercial buildings share each fire's severity. That shared severity is what lets the first home inspections update the estimate for a lender's commercial collateral.
The inspection form splits commercial buildings by storey count, plus schools, churches, hospitals, mixed use and infrastructure. It has no warehouse, industrial, office or retail code, so those types are scored as the nearest class and marked as proxies. Contents, stock, machinery and business interruption are not in the record. Ports, refineries and large plants are not represented.
Same fire, same recorded features and spacing. Lines are 90% ranges from resampling whole fires. One-storey commercial buildings are close to houses; multi-storey, institutional and infrastructure buildings are destroyed far less often.
Brier score on the commercial buildings of each withheld fire after the first k home inspections, mean over fires. A model fitted to commercial records alone cannot use home inspections, so its score does not move.
Validation
Each of fires was removed in turn, the models were trained on the other fires, and the removed fire was scored. Fires qualify with at least 200 homes and at least 20 destroyed and 20 not destroyed. Lower is better for Brier score, log loss and fire-share error; higher is better for AUC.
Bars are WildfireVuln's 80% range for the destroyed share of each fire, made without any information about that fire. Dots mark what happened.
Brier score on the homes not yet inspected, mean over withheld fires. Boosting gets the same early inspections and has its average re-centred on them. WildfireVuln's update is Bayesian: with few inspections it leans on what it knows about fires in general, so it does not over-react.
Where we win. Against every model that can be used for pricing, WildfireVuln has the lowest fire-share error and log loss, the best update once inspections arrive, and the only calibrated range for a fire's total. It ranks homes about as well as gradient boosting on the same features (AUC against ).
Where we do not. Gradient boosting on the same features has a slightly lower Brier score before the fire and a slightly higher AUC. Models that read the outcome-dependent survey fields rank homes better against the survey; we explain above why that is not usable.
Score a loan book or a schedule of locations
Paste or upload a CSV of collateral, residential and commercial together, and every row is scored in your browser; nothing is sent anywhere. Buildings in one fire share its severity, so the range for the book is wider than scoring each loan on its own suggests. Losses are to the building, before any property insurance recovery. The rows loaded below are examples.
asset_type: sfr_1, sfr_2, mobile, motorhome, multifamily, or a commercial type: warehouse, industrial, logistics, retail, office, hotel, commercial_1, commercial_2, school, church, hospital, mixed_use, utility. Roof asphalt, tile, metal, wood · eaves open, enclosed, none · vents coarse, fine, open, none · anything else or blank is unknown. n30 and n100 are structures within 30 m and 100 m; nn_m is metres to the nearest structure; blank spacing uses the suburban median for the asset class. A severe fire is the 90th percentile of the fires in the record.
Method and limits
CAL FIRE Damage Inspection (DINS) records, structures pulled from the public feature service. Kept: residential, commercial and infrastructure structures from fires starting or later with an assessed damage category: homes and commercial and infrastructure buildings. Sheds, barns and garages are left out. Earlier fires use a coarser form and mostly record damaged homes only. Street addresses are dropped at download.
Logistic regression for P(destroyed) with structure type, year built, roof, eaves and vents in their graded form, spacing to neighbouring structures, and a partially pooled effect for each fire (spread τ = on the logit scale). For a new fire the prediction is averaged over that effect. Unknown fields take the average contribution of the known values, so "unknown" never carries information.
Destroyed counts as a total loss. Homes that survive are assigned the observed mix of DINS categories for their structure type, at 5%, 17.5% and 37.5% for affected, minor and major. Smoke, ash and contents losses are not in the inspection record and are not modelled.
Every inspected home was inside or near a fire perimeter. The model answers "if fire reaches this neighbourhood", and needs a hazard model for the chance that it does.
Comparing neighbours in the same fire removes most differences in exposure, but not all. Homes built after 2008 also differ in ways the form does not record. Credits for fences, decks and patio covers cannot be estimated from post-fire surveys at all; they need inspections made before the fire.
DINS records commercial buildings by storey count, plus schools, churches, hospitals, mixed use and infrastructure. It has no warehouse, industrial, office or retail code, so those collateral types are mapped to the nearest class and marked as proxies. Contents, stock, machinery and business interruption are not in the record. Ports and large industrial plants are not represented.
California only. Other states and countries have different building stock, codes and fire behaviour, and no comparable open inspection record.
Commercial catastrophe models are licensed and were not available to test. Every comparison here is against a model built and scored on the same data, and the code is public.
Pilot
Most carriers already hold underwriting inspections or application answers for homes that later burned. Joined to DINS outcomes, they record the house as it stood, which is exactly what the post-fire survey cannot, and they make the fence, deck and siding credits estimable.
Location, construction and mitigation fields for policies in the footprint of any fire since 2018. Coordinates can be rounded to 50 m.
Each policy is matched to its inspected structure and neighbourhood. The same withheld-fire test runs on your fields.
Credits for your own features with the same three-way check, ready for a Safer from Wildfires rate filing.