The peril model, explained.

Atlas United grades are a purpose-built deterministic peril model. Eight independently calibrated layers, fifty-four perils, validated against decades of observed severe events. No black-box blending, no proprietary aggregator math — every score traces to a documented model component with version, calibration vintage, reason code (patent claim 30).

ATMOS
Atmospheric & tropical layer
Hurricane intensity, track climatology, wind-field reconstruction, and high-resolution forecast wind grids. Calibrated against the full Atlantic basin tropical-cyclone catalog from 1851 forward.
  • Tropical-cyclone catalog · 1851–today
  • 4x-daily forecast wind grid
  • Surge basin envelope
  • Extreme precipitation atlas
Hazards servedHurricane · Storm surge · Pluvial flood · Wind climatology · SLR
v2024.3Live
SURFACE
Riverine & coastal flood layer
Special Flood Hazard Areas, multi-hazard risk index, and surge inundation by basin. Cross-validated against post-event flood-extent observations and disaster-declaration history.
  • SFHA polygons · live
  • Multi-hazard risk index
  • Disaster-declaration history
  • Mitigation project pipeline
Hazards servedRiverine flood · Multi-hazard · Disaster history
v2024.2Live
SEISMIC
Earthquake, landslide & volcanic layer
National seismic hazard model with peak-ground-acceleration grids, plus landslide susceptibility and volcanic exposure. 50-state coverage for seismic, regional coverage where geology requires it.
  • 50-state seismic hazard · 2023
  • Landslide susceptibility
  • Volcanic exposure zones
  • 3D elevation model
Hazards servedEarthquake · Landslide · Volcanic · Tsunami
2023 catalogQuarterly
WILDFIRE
Burn perimeter & WUI exposure layer
Burn-probability raster, fuel-model classification, and Wildland-Urban Interface exposure scoring. Calibrated against the full national fire-occurrence record and post-fire perimeter data.
  • Burn-probability grid
  • Fuel-model classification
  • Fire-occurrence catalog
  • Ember-zone overlay
Hazards servedWildfire · Burn probability · WUI exposure · Ember
v2024.1Live
CONVECTIVE
Hail, tornado & severe-wind layer
Convective-event archive spanning seven decades of tornado tracks, severe-hail reports, and damaging-wind events. The calibration set behind every convective peril Atlas grades.
  • Severe-event database
  • Tornado-track archive
  • Severe-hail catalog
  • Day-1 convective outlook
Hazards servedTornado · Hail · Severe wind · Lightning
1950–todayDaily
CHRONIC
Chronic-climate & long-horizon layer
Climate normals, billion-dollar disaster trends, and forward projections at 2030 / 2050 / 2100 horizons. Powers chronic peril scores: heat stress, drought, sea-level rise.
  • 1991–2020 climate normals
  • Monthly climate grids
  • Billion-dollar disaster catalog
  • 2030 / 2050 / 2100 projections
Hazards servedHeat · Cold · Drought · Sea-level · Trend
1991–2020Annual
INFRA
Infrastructure & environmental layer
Levee and dam exposure, air-quality monitoring, water-system records, and grid-reliability indicators. Captures the man-made systems that modulate peril outcomes around a parcel.
  • Levee & dam inventory
  • Air-quality monitoring
  • Water-system records
  • Grid-reliability metrics
Hazards servedLevee failure · Air quality · Outage · Habitability
v2024Quarterly
FORECAST
Forward probability layer
Sixteen-day forward-looking peril probability, refreshed four times daily. This is the AtlasCast layer: trained storm models rather than fixed formulas, documented in full in the storm data section below.
  • 16-day forward grid
  • 4x-daily convective refresh
  • Tropical-track ensemble
  • Daily severe outlook
Hazards servedAll convective · Tropical · Forward windows
Live4x-daily
Storm data methodology

AtlasCast and TrueFixR: how the storm data is made.

The A–F peril grades above are deterministic. The storm data is different: AtlasCast forecasts come from trained machine-learning models, and TrueFixR damage data comes from observed and reported events. Both are described here, including where they fall short.

ATLASCAST
Storm forecast models
One trained model per storm peril (gradient-boosted trees), scored against the address universe every cycle. Inputs are NOAA ensemble forecasts (GEFS), high-resolution HRRR fields, and radar-derived features (MRMS hail size and rotation tracks). Models are validated on a chronological, purged hold-out so future weather never leaks into training.
  • NOAA GEFS, HRRR and MRMS inputs
  • Calibrated probability per address
  • Reason and severity on every lead
  • County grades and exposure roll-ups
Hazards servedTornado · Hail · Wind · Flash flood · Heavy rain · Winter · Ice · Wildfire · Hurricane · Coastal surge
Live4×/day
EXPOSURE
Property value and people in the path
Each address inherits the value and population of its building from the USACE National Structure Inventory (133M structures). Totals are rolled up per county so a forecast reads as dollars and people, not just a count of addresses.
  • USACE National Structure Inventory
  • Replacement cost, not market price
  • Shared across units in one building
  • County and state roll-ups
Fields servedStructure value · People · County exposure
NSIStatic
TRUEFIXR
Reported storm-damage layer
Address-level leads built from storm reports and radar-reported hail, matched to real property addresses. Served as county and state histories (up to 365 days), peril breakdowns and flood-risk context. Radar-reported hail (MRMS MESH) is archived back to November 2019 and available as a bulk dataset.
  • Damage leads by county or state
  • MRMS MESH radar hail, 2019 onward
  • Peril breakdown and flood context
  • 13-endpoint JSON API
Hazards servedHail · Wind · Heavy rain · Flood · Tornado
DailyLive
Known limits

What the storm data does not claim.

A likelihood, not a guarantee

AtlasCast values are modeled likelihoods. They rank where damaging weather is more probable, and are not a promise that any single roof will be hit.

Area risk, not rooftop risk

Weather-model cells are tens of kilometers wide, so every address in a cell shares its signal. Treat ratings as neighborhood-to-county risk.

Address points, not unique buildings

The 282.9M figure counts address points. Roughly half sit within a few meters of another point for the same street address, so distinct deliverable addresses are fewer.

Replacement cost, not market value

Exposure dollars are what it costs to rebuild a structure, the number insurance cares about. They exclude land and location premium.

Perils vary in skill

Winter, ice, wildfire and wind models are the strongest. Tornado, hail, heavy rain and flash flood are useful for ranking areas but weaker, and hurricane and coastal surge have few training events. Ask us for per-peril validation figures.

Not every peril is backfilled

Hindcast and historical runs do not cover hurricane, coastal surge, winter and ice, or stream-gauge flood data. Forecast lead time in a hindcast is at most about 15 hours.

Validation

Calibrated against decades of observed events.

Every Atlas peril layer is independently calibrated against a documented event catalog spanning at least 22 years of validated observations — and continuously back-tested against new severe events as they occur. We publish hit-rate statistics. Most peril vendors don't.

Citable in filings
Every score response includes model version, calibration vintage, and reason code — language ready to paste into a rate-filing exhibit.
Versioned and replayable
Pin a score to a specific model version and replay it years later. Useful for litigation, audits, and back-testing.
Published hit rates
We publish address-level verification statistics against post-event observations, layer by layer. Transparency is a feature, not a marketing asterisk.