Storm Forecast by Address: Value and People at Risk

What it is: a family of per-peril machine-learning models. They identify the addresses inside the forecast footprint of a storm over the next 24 hours to 10 days, across 10 perils: TORNADO, HAIL, WIND, FLASH_FLOOD, HEAVY_RAIN, WINTER, ICE, COASTAL_SURGE, HURRICANE, WILDFIRE. Could happen, not will.

What you send: one or more US counties as County, ST, a forecast window, and optionally one peril.

What you get per row: one address with the structure replacement value and estimated occupants attached to it, the storm type, forecast time, a risk label, the count of prior storm hits at the address, the window, a reference_id and the forecast generation time.

Cost: pricing is not enabled yet. The planned rate is $0.05 per delivered address. A preview of the address count and estimated cost is logged per county before delivery, and maxAddresses and maxChargeUsd bound the run.

Models are refreshed 4 times per day.

What you get

One flat record per forecast address. These are exactly the fields delivered in the dataset.

Field	Meaning
structure_value	Estimated structure replacement value in USD (rebuild cost, not market value)
people	Estimated number of occupants
full_address, city, state, zip, county	Street address and its locality
lat, lon	Address coordinates
storm_type	The peril (one of the 10 above) the record refers to
forecast_time	Time the forecast event is expected to affect the address
risk_label	The model's risk label for this address and peril (for example Extreme)
storm_hits	Count of past storm hits at the address
facility_name, facility_type	Facility name and type when the address is a recognized facility, otherwise empty
window	The forecast window requested (24h to 10d)
reference_id	Identifier of the forecast pull, for citation and reproducibility
forecast_generated	Timestamp at which the forecast was generated

Methodology

Each peril has a dedicated machine-learning model that identifies, address by address, which addresses fall inside the forecast footprint of that peril within the window. Models are trained and validated chronologically: they are fit on earlier periods and evaluated on later, held-out reported events, so evaluation reflects forecasting forward in time. Models are versioned and the pipeline is reproducible, and every pull carries a reference_id and a forecast_generated timestamp so any record can be traced to a specific forecast run. Output is at the level of the individual address, not a county or polygon summary, and is refreshed four times per day.

Interpretation

A record means a storm could affect that address within the window, not that it will. A record is a forecast record, not a damage assessment, an insurance determination or a claim of loss. risk_label is a category expressing the model output; structure_value and people are exposure context, not a loss estimate. Use forecast_generated to judge freshness and reference_id to cite the run.

Input

{ "counties": ["Dallas, TX"], "window": "24h", "peril": "", "maxAddresses": 100 }

• counties (required): list of County, ST strings.
• window: 24h, 48h, 3d, 5d, 7d or 10d. Default 24h.
• peril: optional, one of the 10 perils. Empty means all 10.
• maxAddresses: cap on delivered records, 1 to 5000. Default 100.
• maxChargeUsd: optional spend cap.

Example record

{ "full_address": "405 Evergreen Trail, Cedar Hill, TX 75104", "county": "Dallas", "state": "TX",
  "structure_value": 300882, "people": 3.5, "storm_type": "HEAVY_RAIN", "risk_label": "Extreme",
  "storm_hits": 2, "window": "24h", "reference_id": "TFX-202610-3A48FA8C",
  "forecast_generated": "2026-10-04T03:32:44.962420Z" }

Compliance

Forecasts describe what could happen, not what will happen. This is data, not insurance, investment or safety advice, and no guarantee of accuracy or results is made. Not affiliated with NOAA or the National Weather Service.