Hurricane Vulnerability, One Building at a Time

Every building.
Every storm.

HurricaneVuln tells insurers and lenders how a specific building responds to a specific storm. It is calibrated on buildings that Hurricane Michael passed over and FEMA-inspected homes from storms.

The vortex is drawn from the same wind-field model the product uses.

buildings matched to FEMA remote-sensing damage after Michael
owner-occupied homes with a FEMA inspection, 2016–2024
storms, each withheld in turn to test the loss model
withheld storms whose average loss fell inside our 80% range

October 9, 2018

Michael crosses the Gulf.

A tropical storm two days earlier, Michael deepens over warm water at a rate forecasters rarely see.

October 10, 17:30 UTC

Landfall at Mexico Beach.

Peak sustained winds of 140 knots, the first Category 5 landfall in the contiguous United States since Andrew in 1992.

The next six hours

Still a major hurricane, 100 miles inland.

The core races north-east into Georgia faster than most storms weaken, which is why inland Jackson and Seminole counties saw damage too.

What FEMA saw from the air

of buildings damaged.

Every orange cell is a 500-metre block of buildings that FEMA's imagery analysts marked as damaged. This is the record HurricaneVuln learns from, building by building.

–Max Wind
–Category
–UTC

Storm Replay

Nineteen storms. Your building.

Pick a storm, press play, and watch its wind field build from the official best track. Click anywhere on land to place a building and read the peak wind it would have felt.

–

Peak Intensity
–
Homes in the Loss Record
–
Median Verified Loss
–

Your Building

Location
Click the map
Peak Sustained Wind
–
Chance of Visible Damage
–

The building uses the configuration below. Wind is the peak 1-minute wind at 10 m in open terrain, the same reference as the best track.

Open in the Configurator

Building Unit

Configure the building.
See what the storm does to it.

Choose what the property record says. Options marked with a dot are recorded in the data but the evidence does not support pricing them separately, so they are scored as the reference option.

Result

–
–Chance of Visible Damage
–Chance Destroyed

Repair Cost If FEMA Inspects a Damaged Home

Expected Verified Loss, This Storm
–
Chance the Home Needs Repairs to Be Habitable
–

Repair costs are FEMA-verified real property losses in 2024 dollars under FEMA's current inspection practice. They measure what it takes to make a home safe and habitable, which is below a full insured loss.

How This Building Responds to Wind

Chance of visible damage from 60 to 160 knots, compared with the typical wood-frame house.

What the Evidence Supports

Effect of each recorded attribute on the odds of damage, with 90% ranges from resampling map tiles.

For Insurers and Lenders

Run a storm over your book.

Paste a schedule of buildings with coordinates and values, choose a historical storm, and every building is scored in your browser. Nothing leaves the page. The rows loaded below are examples.

Columns: id, lat, lon, construction (wood, masonry, concrete, steel, manufactured), use (single_family, mobile_home, condo, multi_family, commercial, industrial, institutional, other), storeys (1, 2, 3), year_built, sqft, value, loan, insured (1 or 0). Blank fields are treated as unknown.

Buildings–
Expected Damaged–
Expected Verified Loss–Residential units
Uninhabitable, Uninsured–Expected count

Evidence

Tested on what it has not seen.

Two models, two tests. The loss model is scored on each of 19 storms after training on the other 18. The fragility model is scored on map tiles and counties it was not fitted to.

Household Loss, Leave One Storm Out

Continuous ranked probability score of the full loss distribution in 2024 dollars, mean over withheld storms. Lower is better.

Average Loss per Withheld Storm

Bars are the 80% range predicted before the storm; dots are what FEMA inspectors verified.

Building Fragility, Hurricane Michael

Within assessed counties, new map tiles. Lower Brier and higher AUC are better.

Where HurricaneVuln Leads

Where It Does Not, Yet

Records

Everything on the record.

The model card, where every number came from, what we assumed, and how to rebuild it all. Model risk teams can audit this page end to end.

Read the Full Technical Documentation