Tornado forecast

Our own model's chance of a tornado within 25 miles of each point in the lower 48, made twice a day from NOAA's HRRR, tested on years it had not seen, and scored every day against what happened and against the Storm Prediction Center.

How the forecast is made

Twice a day, after the 00 and 12 UTC runs of NOAA's High-Resolution Rapid Refresh model (the HRRR), the model reads the afternoon and evening that run forecasts for each 39 km cell: instability, wind shear, storm-relative helicity, the height of the cloud base, updraft helicity, simulated radar and the indices built from them. Gradient-boosted decision trees, trained on tornado reports from 2018 to 2023, turn those numbers into a probability. The probability is calibrated: over many days, cells marked 5% should have a tornado within 25 miles about 5% of the time. The 00 UTC forecast is made about two hours after the run, before the day begins at 12 UTC; the 12 UTC forecast about an hour and a half after the day has begun.

How it is tested

Every choice was made on 2018 to 2023. The model was then frozen and tested once on 2024 and 2025, and on 2026 as it happened, against the Storm Prediction Center's day 1 outlook of the matching hour: the 06 UTC outlook for the 00 UTC run, the 16:30 UTC outlook for the 12 UTC run. The score is the Brier skill score against climatology, the average chance of a tornado in that place at that time of year from 1995 to 2017: zero means no better than knowing the season, and higher is better. Intervals come from resampling whole days, since the cells on one day are not independent.

From the day it was switched on, every forecast is saved before its day begins and scored two days after it ends, from SPC's preliminary tornado reports. The log is never edited, and a forecast the system missed is left out, not filled in afterward.

What it is not

It is not an official forecast. It does not see radar, satellite or observations: it knows only what one run of one model forecast for the afternoon and evening. It is statistically indistinguishable from the Storm Prediction Center's outlook on the days tested, not better, and where both mark the same share of the map the Center's areas catch more of the tornadoes. A companion model for EF2 or stronger tornadoes was built and tested the same way; it under-forecast strong tornadoes in 2024 and 2025 and did not pass its test, so it is not in use and nothing here is a strong-tornado forecast. Reports of tornadoes depend on where people are, so the scores measure reported tornadoes, and a tornado in an empty place may never be counted. The National Weather Service and the Storm Prediction Center issue the forecasts, watches and warnings to act on.

Where the data comes from

The HRRR is a NOAA model; this site keeps its severe-storm fields from every run in its own archive. Tornado reports are from the Storm Prediction Center: its tornado database for 2018 to 2025 and its preliminary reports since. The outlooks it is compared with come from the Iowa Environmental Mesonet's archive of the Center's products.