A storm object is not a radar pixel and it is not a warning. It is one convective cell, isolated as a polygon, given an identity, and scored.
The cadence
A new picture of every convective cell in the lower 48 lands every two minutes: roughly 200 storms in an active period, each with a polygon, an identity and its readings.
Two minutes is the cadence, not two minutes of new information. The radar network underneath only revisits a given point when it sweeps it, which at the lowest tilt takes between 42 and 182 seconds. Between sweeps, the mosaic carries the same observation forward.
A scan also describes a moment about two minutes before it is published, before any processing of ours, so that is the floor on how early anything here can possibly be.
What each storm carries
Four hazard readings:
| Reading | What it answers |
|---|---|
| Severe | Will this storm produce any severe hazard? |
| Hail | Severe hail specifically |
| Wind | Damaging wind specifically |
| Tornado | Is it rotating, and is that rotation going anywhere? |
Plus around fifty diagnostic fields: hail size and vertically integrated liquid, echo tops, azimuthal shear at two levels, satellite lightning flash extent and density, the lightning-jump statistic, and environmental fields such as low-level storm-relative helicity sampled at the storm.
The tornado reading is not the severe reading
This is the single most misread thing in the product, so it gets its own section.
They are not the same scale and they do not mean the same thing at the same number. The severe reading is a union model - it answers “is this storm dangerous at all”, and on an active day plenty of storms sit high on it. The tornado reading is far more selective. On an ordinary convective day the overwhelming majority of storms sit at the bottom of it and nothing anywhere in the country reaches the top of the scale. It stays dark unless something is genuinely rotating, and when it lights up, that is the signal.
Which is why the two have different vocabularies. The tornado axis has the five bands - Watching through Extreme - and the severe axis has none: it is carried on the map by the white ring and in the ladder by the two thunderstorm tiers.
What each band is worth
A tornado report counts as a hit when it falls within 40 km of the storm and within the next 30 minutes. Each tracked storm is counted once if it ever reached the band, across a full year rather than a single outbreak:
| Band | Produces a tornado | False alarm |
|---|---|---|
| Watching | 13% | 87% |
| Elevated | 20% | 80% |
| Significant | 27% | 73% |
| High | 37% | 63% |
| Extreme | 52% | 48% |
Those are year-round rates with quiet days included, which is the conservative view. On a single outbreak day the same bands run far higher: Significant is a 42% tornado call on a day like that.
The false-alarm column is the honest one. Even at Extreme, about half the time no tornado is reported. That is what a nowcast on a rotating storm can do, and a product claiming better than this on the tornado hazard alone is not measuring itself against ground truth.
The same storms as a severe-hazard call
Run the identical triggers against any severe report - wind, hail or tornado - rather than tornadoes alone. Both columns are the same storms on the same day:
| Band | Produces a tornado | Produces some severe hazard |
|---|---|---|
| Elevated | 28% | 81% |
| Significant | 42% | 85% |
The same storms at the same threshold go from a 42% tornado call to an 85% severe-hazard call: a false-alarm rate of 15% instead of 58%. For context, 12% of all tracked storms produce a severe hazard, and 19% of those lasting 20 minutes or more, so Significant is roughly a 4.5× lift rather than an artifact of a busy day.
The two thunderstorm tiers
Below the tornado bands sit the two severe tiers, set on the severe reading rather than the tornado one:
- Thunderstorm Risk - a storm capable of severe weather is heading this way. Between 35 and 67% of the storms it fires on produce a severe report, covering 31 to 82% of the reports on a day.
- Thunderstorm Detected - a severe storm is on you now. A higher bar on the same reading, set where a squall line and a discrete supercell are scored about equally well; a bar tuned for supercells alone misses 20 to 32 points of a squall line’s damage reports, because a line is many short-lived cells that each do damage without any one of them looking impressive.
Hail size and integrated liquid ride in the content of those alerts - “2.1 inch hail, 60 kt wind” - and never in the decision to send one.