Research · notes from testing
We check Marble against what actually happened.
Six short notes from our internal testing: what Marble shows, how we tested it and how well it did. Every headline number comes with its sample size and dates, and the results that are not flattering stay in.
The notes
-
Note 01 · Forecast accuracy
How close is the forecast?
1.64–2.52 °C typical temperature error, start to day 5
The forecast exactly as Marble shows it, checked against about 1,860 airport weather stations: 108,741 comparisons per model, out to five days.
4 min read
-
Note 02 · Terrain
Temperature where you actually stand
6.5 → 3.0 °C typical error in deep valleys, before and after
Global models smooth mountains into gentle hills. Marble adjusts temperature to the real ground height. We tested it at 8,932 stations.
4 min read
-
Note 03 · Radar nowcast
Rain in the next two hours
0.47 vs 0.31 rain placement score one hour ahead, Marble vs “no change”
Marble’s radar nowcast against the radar scans that followed: 30 real cases across the United States, scored minute by minute against “no change”.
4 min read
-
Note 04 · Rain chance · experimental
Will it rain in the next hour?
0.58 skill for 1 mm in the next hour, inside US radar (0 = no skill, 1 = perfect)
The experimental rain-chance layer against the hour that followed, measured by radar and rain gauges. What a 70 % should mean, and where it is too confident.
5 min read
-
Note 05 · Model comparison
Which forecast model is closest?
1.21 °C average temperature error one day ahead, AIFS as Marble shows it
The two ECMWF models Marble shows against NOAA GFS and the high-resolution HRRR, on the same 185 airports over one week of November weather.
4 min read
-
Note 06 · Limits
What Marble cannot do yet
28 km the grid of the global models Marble shows, and other honest limits
The honest limits: global resolution, experimental guidance, radar coverage, a short test period, no alerts and no native apps. And what we are testing next.
4 min read
How we test
- 01
As you see it
We score the values exactly as Marble displays them, not an idealised version.
- 02
Against what happened
Airport reports, national station networks, radar and satellite estimates, never the model checking itself.
- 03
Numbers with context
Sample size and dates next to every headline figure, with uncertainty where we have it.
- 04
Unflattering results stay in
Over-confident light rain, cold summits at night, short test periods: they are all here.
These notes summarise internal testing; methods are described only in outline. Marble is not an official warning service.
