A scan of storm objects is not yet a product. Three things happen to it first: the storm gets a motion, it gets an identity that survives the feed losing it, and it gets an alerting rule that will not wake you twice for the same storm.
Storm motion
The upstream feed carries motion fields of its own, and they are not usable: at their best the residual error runs 38 to 74% of a typical storm’s speed, and even the direction is right only 70 to 93% of the time.
So motion is fitted here instead, by least squares over the last five centroids - ten minutes of history - with any physically impossible jump thrown out first. That predicts where a storm will be ten minutes from now to a median error of 5 to 6 km, and everything about the cone rests on it.
A storm with too few scans, or one moving too slowly for the grid to resolve, reports no motion at all rather than a guess. That is why a cell can appear on the map with no cone and a popup that says not yet fitted.
One storm stays one storm
Within a single upstream ID the feed is excellent: about 99.97% of IDs are a contiguous run of scans, with no reuse and no jumping between storms. The failure is at the seams.
51% of significant tracks end while the storm is still obviously going. The storm does not decay; it merges with or splits from a neighbor and comes back under a new number, often rated much weaker than when it vanished. Read literally, that sequence is severe, then nothing, then severe - and keying notifications off the raw ID produces 22% duplicates for the same physical storm.
So when a track ends, it is linked to the nearest new object within 25 km over the next five scans, and the two become one storm. On a single outbreak day that merges 5,694 upstream IDs into 2,456 physical storms, and roughly doubles the usable life of a significant storm: a median of 112 minutes where the raw track gives 60.
When a cell’s popup says “re-identified by the feed 3×, merged back into one storm here”, this is what it is telling you.
Why one storm alerts once
A storm sitting near a threshold will cross it repeatedly, and firing on every crossing is how one storm wakes you six times.
A storm alerts when it reaches the Significant band, and clears only once it has fallen well below that band for three consecutive scans, rather than the moment it dips under. That takes the chance of a storm re-alerting from 56 to 74% down to 6 to 8%.
The clearing rule is deliberately slower than the firing rule. A storm pulsing around the threshold is one storm having one bad hour, and it should cost you one notification.
The cone
The cone is not a forecast track. It is the region the storm might occupy out to a lead time you choose, built from the storm’s own measured size, speed and bearing.
It is lopsided on purpose. Storms under-run a straight line: the distance a storm actually travels is a median 0.85 to 0.91 of what its own motion predicts, so a cone centered on the prediction wastes most of its area out ahead of a storm that will never get there. The cone is pulled back accordingly, and it widens with lead time at about 22° of half-angle for a storm moving 67 km/h.
What a longer lead costs you
Capture is the chance the storm is really inside the cone when it arrives:
| Lead | Capture | Area |
|---|---|---|
| 10 min | 99% | 1,787 km² |
| 20 min | 98% | 2,913 km² |
| 30 min | 97% | 4,280 km² |
| 45 min | 96% | 6,705 km² |
| 60 min | 95% | 9,509 km² |
Accuracy barely moves. Area grows 5.3× between 10 and 60 minutes, and 9,509 km² is roughly a 100 km swath with everyone in it notified. That is why the lead selector is a real choice rather than a display preference, and why the notification lead is set separately from whatever the map happens to be drawing.
The cone follows the storm rather than predicting its future: it is rebuilt from scratch every scan, so a weakening storm’s cone shrinks on its own and places fall out of it without anything having to forecast the decay.