ENPT
Open Marble

Note 01 · Forecast accuracy

How close is the forecast?

Against about 1,860 airport stations, the temperature on Marble’s map missed by a typical 1.6–1.7 °C at the start of a forecast and 2.2–2.5 °C five days out.

November 20264 min readSuperlumine

A summary of internal testing. Methods are described only in outline. Not an official forecast or warning.

  • 108,741comparisons per modelIFS and AIFS runs, 17–22 Nov 2026
  • 1.64 °Ctypical temperature error at the start (AIFS; IFS 1.69 °C)17,324 comparisons
  • 2.19 °Cfive days ahead (AIFS; IFS 2.52 °C)1,597 comparisons
  • 98%of stations kept their average error within ±3 °C over the first day1,858 stations

What we tested

Marble shows two global forecast models from the European Centre for Medium-Range Weather Forecasts (ECMWF): the physics-based IFS and the machine-learning AIFS. We took eleven runs of each, issued between 17 and 22 November 2026, and compared the value exactly as Marble displays it with the official weather report (METAR) from about 1,860 airport stations, at the same time to within 15 minutes.

We did this every 12 hours of lead time, from the start of the forecast to five days ahead: 108,741 comparisons per model, from 17,324 at the start to 1,597 at five days. The score is the typical error (root-mean-square error), the usual size of a miss with large misses counting more. For readers in Fahrenheit: an error of 1 °C is 1.8 °F.

Temperature error by lead time, IFS and AIFS

Temperature error by lead time, IFS and AIFSRoot-mean-square error of 2 m temperature as displayed against airport reports, 17–22 November 2026. IFS rises from 1.69 °C at the start to 2.52 °C at five days; AIFS from 1.64 °C to 2.19 °C. Shaded: 95 % interval.1.52.02.5012345Typical temperature error, °CForecast lead time, daysIFS 2.52AIFS 2.19
Typical temperature error by lead time. Both models stay near 2 °C for most of the week; the AIFS was the more accurate of the two at every lead in this sample.
Chart in words

Root-mean-square error of 2 m temperature as displayed against airport reports, 17–22 November 2026. IFS rises from 1.69 °C at the start to 2.52 °C at five days; AIFS from 1.64 °C to 2.19 °C. Shaded: 95 % interval.

Temperature

The typical temperature error grew from 1.69 °C (IFS) and 1.64 °C (AIFS) at the start to 2.52 °C and 2.19 °C at five days. The AIFS was clearly ahead from one day on (1.78 against 1.93 °C at 24 hours). The IFS ran slightly cold (by 0.2 to 0.6 °C on average), the AIFS slightly warm (0.1 to 0.5 °C). The largest station errors were in complex terrain, such as cold valleys of the western United States and high Andean sites.

Marble showing the IFS temperature forecast on the globe over Europe, North Africa and Asia
The temperature layer as tested: ECMWF IFS, 22 November 2026, light theme. Real Marble screenshot.

Wind

Wind at 10 metres was about 0.5 m/s too weak at every lead, with a typical error of 1.9 to 2.4 m/s. That is expected: the model gives an average over a square about 28 km wide, while airports measure on an exposed mast. The error in wind direction grew from about 21° at the start to 36° at five days.

Rain

Both models showed rain about twice as often as airports reported it: one day ahead, in 25 % (IFS) and 32 % (AIFS) of cases, against 12.5 % reported. Part of this is a difference in what is measured: the map shows an average over a wide square and a few hours, the airport report says whether it is raining at one spot at that moment. The IFS placed rain slightly better; the AIFS caught more rain events, with more false alarms.

How often rain is shown versus reported

How often rain is shown versus reportedAt 24 hours ahead the IFS showed rain in 25 % and the AIFS in 32 % of cases, while airports reported precipitation in 12.5 %.0%10%20%30%40%Share of cases flagged as wet, 24 h aheadReported12.5%IFS25%AIFS32%
Share of cases flagged as wet one day ahead, against the share reported at airports.
Chart in words

At 24 hours ahead the IFS showed rain in 25 % and the AIFS in 32 % of cases, while airports reported precipitation in 12.5 %.

Data and attribution

Forecast data: ECMWF open data (IFS, AIFS), CC BY 4.0. Observations: airport reports (METAR) from the Iowa Environmental Mesonet archive.

A summary of internal testing. Methods are described only in outline. Not an official forecast or warning.