MeteoSignal

30 DWD stations in Germany.

AIFS 2 against IFS and GFS: lower error in 17 of 28 tested comparisonsSample, sources and status
Scored against
Hourly station observations (DWD open data)
Sample
AIFS 2: 174 days, 23 May 2026 to 12 November 2026. Other AI models: 68 days
Result so far, AIFS 2
Against IFS and GFS: AIFS's error lower in 17 of 28 tested comparisons, higher in 0, not separated in 11
Status
Provisional: no full year yet. The main result, with tested verdicts.

Last reviewed 16 November 2026Measurements to 12 November 2026How verdicts are reachedData access

Sankt Peter-OrdingDörnickUeckermündeNorderneyHamburg-FuhlsbüttelBremenDiepholzBraunschweigLindenbergLügde-PaenbruchLeinefeldeGörlitzDresden-KlotzscheKöln/BonnAachen-OrsbachGießen/WettenbergNürburg-BarweilerLautertal-OberlauterWürzburgMannheimNürnbergWaldmünchenSaarbrücken-EnsheimStuttgart (Schnarrenberg)FürstenzellAugsburgWeihenstephan-DürnastMühldorfChiemingKonstanz
The 30 DWD stations scored on this page. Hover over or tap a point to see the site.

Results at the German stations

Show
Version
Reference
Forecasts

AIFS 2: in service from 12 May 2026; forecasts issued from 13 May 2026, scored for valid days up to 12 November 2026.

2 m temperature: skill against climatology
-200204060▲ violet: lower error than climatologyDay 1Day 3Day 5Day 7Day 10Skill, %AIFSIFSGFS

Raw forecasts, 174 days, 23 May 2026 to 12 November 2026. Hover for values; AIFS shows its 95% range, others on hover.

AIFS's RMSE against each model, in %, over 174 days
Compared withDay 1Day 3Day 5Day 7Day 10
IFS−10%, AIFS lower error, tested−13%, AIFS lower error, tested−13%, AIFS lower error, tested−9%, not separated−11%, not separated
GFS−29%, AIFS lower error, tested−29%, AIFS lower error, tested−24%, AIFS lower error, tested−20%, not separated−15%, not separated
10 m wind speed: skill against climatology
-40-2002040▲ violet: lower error than climatologyDay 1Day 3Day 5Day 7Day 10Skill, %AIFSIFSGFS

Raw forecasts, 174 days, 23 May 2026 to 12 November 2026. Hover for values; AIFS shows its 95% range, others on hover.

AIFS's RMSE against each model, in %, over 174 days
Compared withDay 1Day 3Day 5Day 7Day 10
IFS−1%, not separated−3%, not separated−7%, AIFS lower error, tested−5%, not separated0%, not separated
GFS−4%, not separated−8%, AIFS lower error, tested−10%, AIFS lower error, tested−10%, AIFS lower error, tested−3%, not separated
  • AIFS lower error, tested (better)
  • AIFS higher error, tested (worse)
  • Not separated: no clear difference, too few days, or the verdict depends on how the station is read or on the height correction (hover says which)

Each number is AIFS's RMSE minus the other model's, as a percentage of the other model's RMSE: minus means AIFS's was lower (better). Hover or tap a cell for the difference in units, its 95% interval and the p-value.

Skill against climatology at day 7, 2 m temperature: AIFS 29%, IFS 22%, GFS 11%. Over these days, AIFS had the highest skill against climatology at this lead. The climatology is not adjusted for warming and ran 0.9 K too cold over these days, which raises every model's temperature skill against it.

Skill against climatology at day 7, 10 m wind: AIFS −5%, IFS −11%, GFS −16%. Over these days, no model had a lower RMSE than climatology at this lead, and AIFS came closest to it.

Bias, activity and eventsWhether a model runs high or low, how much it varies, and how it does in strong wind and frost

What an average error hides: whether a model runs high or low, whether it varies as much as the weather, and how it does when it matters. Bold and underlined marks the value nearest the ideal at each lead, as printed. It is not a verdict: these values are not tested against each other. Hover an event score for its 95% interval. The thresholds were fixed before scoring.

2 m temperature: bias, in °CForecast minus measurement, on average. Plus: too high. Minus: too low. Zero is best.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS+0.08+0.08+0.21+0.43−0.15
IFS+0.11+0.14+0.36+0.53+0.47
GFS−0.14−0.08+0.02−0.02−0.21
2 m temperature: activityVariability of the forecast divided by that of the measurements, both as departures from climatology, with each station centred on its own mean at each valid hour first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.980.970.990.980.95
IFS0.950.940.970.960.99
GFS0.940.940.940.960.98

2 m temperature of 0 °C or belowObserved 20 times on 7 days in 10401 station-times (0.2%). Too few events to score: at least 30, on at least 10 days, are needed.

10 m wind: bias, in m/sForecast minus measurement, on average. Plus: too high. Minus: too low. Zero is best.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS−0.26−0.34−0.38−0.45−0.42
IFS−0.21−0.26−0.27−0.34−0.38
GFS+0.02−0.06−0.07−0.06−0.08
10 m wind: activityVariability of the forecast divided by that of the measurements, both as departures from climatology, with each station centred on its own mean at each valid hour first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.910.890.850.830.87
IFS0.950.940.920.890.86
GFS1.000.980.930.950.95
10 m wind of 8 m/s or more: frequency biasObserved 285 times on 78 days in 10351 station-times (3%). Forecast events per observed event. 1.00 is best. Above 1 the model calls the event too often, below 1 too rarely.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.470.380.340.260.36
IFS0.620.640.570.490.44
GFS1.091.010.880.920.85
10 m wind of 8 m/s or more: equitable threat scoreHow well forecast and observed events coincide, beyond chance. 1 is perfect, 0 is chance, higher is better.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.310.230.190.120.05
IFS0.350.310.200.140.07
GFS0.360.270.210.140.10
10 m wind of 10.8 m/s or more: frequency biasObserved 68 times on 34 days in 10351 station-times (0.7%). Forecast events per observed event. 1.00 is best. Above 1 the model calls the event too often, below 1 too rarely.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.150.160.070.100.18
IFS0.680.620.720.480.63
GFS0.680.600.530.540.52
10 m wind of 10.8 m/s or more: equitable threat scoreHow well forecast and observed events coincide, beyond chance. 1 is perfect, 0 is chance, higher is better.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.150.140.070.06-0.00
IFS0.310.250.170.060.07
GFS0.340.270.130.060.10
Against ICON, up to 7 November 2026A closed set that no longer grows

ICON's only archive of past runs was Open-Meteo's. Its last valid day read here is 7 November 2026, and the archive is no longer read. It is compared with AIFS here on the days it has, as a set of its own that no longer grows, while the other sets move forward each week. It returns as a new series from DWD's own files once enough runs have been collected.

2 m temperature: skill against climatology
-200204060▲ violet: lower error than climatologyDay 1Day 3Day 5Day 7Day 10Skill, %ICONAIFSIFSGFS

Raw forecasts, 170 days, 23 May 2026 to 8 November 2026. This sample no longer grows.

2 m temperature, AIFS 2: how AIFS's error compared, over 170 days
Compared withDay 1Day 3Day 5Day 7Day 10
ICON−12%, AIFS lower error, tested−16%, AIFS lower error, tested−16%, AIFS lower error, tested––
10 m wind speed: skill against climatology
-40-2002040▲ violet: lower error than climatologyDay 1Day 3Day 5Day 7Day 10Skill, %ICONAIFSIFSGFS

Raw forecasts, 170 days, 23 May 2026 to 8 November 2026. This sample no longer grows.

10 m wind speed, AIFS 2: how AIFS's error compared, over 170 days
Compared withDay 1Day 3Day 5Day 7Day 10
ICON−11%, AIFS lower error, tested−12%, AIFS lower error, tested−17%, AIFS lower error, tested––
  • AIFS lower error, tested (better)
  • AIFS higher error, tested (worse)
  • Not separated: no clear difference, too few days, or the verdict depends on how the station is read or on the height correction (hover says which)

ICON's 10 m wind ran low at every lead at these stations, on average by 0.59 to 0.62 m/s. One day ahead, AIFS's error was 11% below ICON's in raw forecasts and 4% below after bias correction, so the raw comparison mostly reflects that offset. The cause of the offset is not established.

2 m temperature: bias, in °CForecast minus measurement, on average. Plus: too high. Minus: too low. Zero is best.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS+0.08+0.09+0.22+0.45−0.14
ICON+0.11+0.06+0.19––
2 m temperature: activityVariability of the forecast divided by that of the measurements, both as departures from climatology, with each station centred on its own mean at each valid hour first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.980.970.990.980.95
ICON0.970.950.97––

2 m temperature of 0 °C or belowObserved 15 times on 5 days in 10161 station-times (0.1%). Too few events to score: at least 30, on at least 10 days, are needed.

10 m wind: bias, in m/sForecast minus measurement, on average. Plus: too high. Minus: too low. Zero is best.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS−0.26−0.34−0.39−0.46−0.42
ICON−0.62−0.61−0.59––
10 m wind: activityVariability of the forecast divided by that of the measurements, both as departures from climatology, with each station centred on its own mean at each valid hour first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.920.890.850.830.87
ICON0.870.880.85––
10 m wind of 8 m/s or more: frequency biasObserved 283 times on 77 days in 10118 station-times (3%). Forecast events per observed event. 1.00 is best. Above 1 the model calls the event too often, below 1 too rarely.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.460.380.330.260.37
ICON0.870.930.85––
10 m wind of 8 m/s or more: equitable threat scoreHow well forecast and observed events coincide, beyond chance. 1 is perfect, 0 is chance, higher is better.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.300.230.180.120.05
ICON0.380.310.22––
10 m wind of 10.8 m/s or more: frequency biasObserved 66 times on 33 days in 10118 station-times (0.7%). Forecast events per observed event. 1.00 is best. Above 1 the model calls the event too often, below 1 too rarely.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.150.170.080.110.18
ICON1.231.331.27––
10 m wind of 10.8 m/s or more: equitable threat scoreHow well forecast and observed events coincide, beyond chance. 1 is perfect, 0 is chance, higher is better.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.150.150.070.06-0.00
ICON0.360.300.18––
Month by month, three days aheadA pattern to read, without intervals

Monthly values carry no intervals, so read them as a pattern and not as results. November 2026 holds 12 days, fewer than 20.

2 m temperature: raw skill against persistence
0204060▲ violet: lower error than persistenceJunJulAugSepOctNovSkill, %AIFSIFSGFS
10 m wind speed: raw skill against persistence
01020304050▲ violet: lower error than persistenceJunJulAugSepOctNovSkill, %AIFSIFSGFS
All scores: 2 m temperature, rawRMSE, MAE, bias and skill for every model and lead
°C, with 95% intervals in brackets
SeriesLeadRMSEMAEBiasSkill against persistenceSkill against climatology
AIFSDay 11.38 (1.28 to 1.46)1.04 (0.97 to 1.10)0.08 (-0.07 to 0.23)61% (57% to 64%)66% (62% to 70%)
IFSDay 11.52 (1.41 to 1.63)1.12 (1.04 to 1.21)0.11 (0.02 to 0.21)56% (52% to 60%)63% (58% to 67%)
GFSDay 11.93 (1.79 to 2.07)1.47 (1.35 to 1.59)-0.14 (-0.44 to 0.15)45% (39% to 50%)53% (46% to 58%)
PersistenceDay 13.50 (3.22 to 3.76)2.73 (2.49 to 2.96)0.06 (-0.16 to 0.26)14% (−1% to 27%)
ClimatologyDay 14.09 (3.46 to 4.74)3.19 (2.70 to 3.73)-0.94 (-1.90 to -0.02)−17% (−36% to 1%)
AIFSDay 31.60 (1.50 to 1.69)1.20 (1.13 to 1.27)0.08 (-0.06 to 0.24)66% (62% to 68%)61% (55% to 65%)
IFSDay 31.84 (1.72 to 1.96)1.37 (1.28 to 1.45)0.14 (-0.01 to 0.31)61% (56% to 64%)55% (48% to 61%)
GFSDay 32.24 (2.09 to 2.38)1.72 (1.60 to 1.84)-0.08 (-0.38 to 0.19)52% (46% to 56%)45% (37% to 52%)
PersistenceDay 34.67 (4.10 to 5.20)3.63 (3.20 to 4.08)0.17 (-0.41 to 0.71)−14% (−30% to −1%)
ClimatologyDay 34.09 (3.45 to 4.74)3.19 (2.70 to 3.73)-0.94 (-1.89 to 0.00)13% (1% to 23%)
AIFSDay 52.13 (1.97 to 2.29)1.60 (1.48 to 1.73)0.21 (-0.01 to 0.45)57% (53% to 61%)48% (40% to 54%)
IFSDay 52.45 (2.26 to 2.64)1.83 (1.71 to 1.96)0.36 (0.10 to 0.63)51% (45% to 56%)40% (31% to 48%)
GFSDay 52.80 (2.59 to 3.00)2.14 (1.99 to 2.30)0.02 (-0.45 to 0.42)44% (38% to 49%)31% (21% to 40%)
PersistenceDay 55.02 (4.51 to 5.56)3.96 (3.56 to 4.39)0.25 (-0.65 to 1.06)−23% (−38% to −10%)
ClimatologyDay 54.09 (3.46 to 4.74)3.19 (2.71 to 3.73)-0.94 (-1.89 to -0.01)19% (9% to 28%)
AIFSDay 72.91 (2.68 to 3.12)2.20 (2.01 to 2.38)0.43 (0.09 to 0.82)48% (41% to 54%)29% (17% to 38%)
IFSDay 73.20 (2.86 to 3.52)2.46 (2.20 to 2.71)0.53 (0.17 to 0.91)43% (37% to 48%)22% (12% to 30%)
GFSDay 73.64 (3.20 to 4.06)2.76 (2.44 to 3.09)-0.02 (-0.53 to 0.45)35% (28% to 42%)11% (−1% to 21%)
PersistenceDay 75.60 (4.90 to 6.27)4.44 (3.88 to 5.01)0.31 (-0.87 to 1.38)−37% (−52% to −23%)
ClimatologyDay 74.09 (3.46 to 4.73)3.19 (2.70 to 3.73)-0.94 (-1.89 to -0.01)27% (19% to 34%)
AIFSDay 103.93 (3.51 to 4.38)3.03 (2.71 to 3.38)-0.15 (-0.66 to 0.37)36% (25% to 44%)4% (−10% to 14%)
IFSDay 104.44 (3.97 to 4.89)3.43 (3.09 to 3.77)0.47 (-0.06 to 1.04)27% (17% to 36%)−9% (−22% to 2%)
GFSDay 104.62 (4.06 to 5.14)3.61 (3.19 to 4.03)-0.21 (-0.90 to 0.47)24% (14% to 33%)−13% (−26% to −3%)
PersistenceDay 106.11 (5.17 to 7.06)4.78 (4.06 to 5.59)0.34 (-1.21 to 1.72)−50% (−66% to −35%)
ClimatologyDay 104.09 (3.46 to 4.73)3.19 (2.70 to 3.72)-0.94 (-1.90 to -0.00)33% (26% to 40%)

Persistence's own row and the skill against it rest on 174 days, 23 May 2026 to 12 November 2026, on the 10,241 of 10,401 station-times where persistence can be scored as well. Every other score needs no earlier measurement.

All scores: 10 m wind speed, rawRMSE, MAE, bias and skill for every model and lead
m/s, with 95% intervals in brackets
SeriesLeadRMSEMAEBiasSkill against persistenceSkill against climatology
AIFSDay 11.20 (1.15 to 1.24)0.92 (0.88 to 0.95)-0.26 (-0.32 to -0.19)39% (34% to 44%)30% (22% to 38%)
IFSDay 11.21 (1.16 to 1.26)0.91 (0.88 to 0.95)-0.21 (-0.27 to -0.16)39% (34% to 43%)29% (22% to 36%)
GFSDay 11.25 (1.19 to 1.31)0.94 (0.89 to 0.98)0.02 (-0.04 to 0.08)37% (32% to 40%)27% (19% to 34%)
PersistenceDay 11.98 (1.85 to 2.12)1.47 (1.36 to 1.58)-0.00 (-0.07 to 0.07)−17% (−25% to −9%)
ClimatologyDay 11.71 (1.55 to 1.90)1.31 (1.20 to 1.44)-0.02 (-0.28 to 0.25)14% (9% to 20%)
AIFSDay 31.31 (1.24 to 1.37)0.99 (0.94 to 1.03)-0.34 (-0.42 to -0.27)42% (36% to 48%)24% (14% to 32%)
IFSDay 31.35 (1.28 to 1.42)1.01 (0.96 to 1.06)-0.26 (-0.32 to -0.20)40% (35% to 46%)21% (13% to 29%)
GFSDay 31.41 (1.35 to 1.48)1.06 (1.01 to 1.11)-0.06 (-0.11 to 0.00)37% (32% to 42%)17% (10% to 24%)
PersistenceDay 32.26 (2.07 to 2.48)1.67 (1.52 to 1.85)0.01 (-0.15 to 0.17)−33% (−41% to −26%)
ClimatologyDay 31.71 (1.55 to 1.90)1.31 (1.20 to 1.44)-0.02 (-0.29 to 0.25)25% (20% to 29%)
AIFSDay 51.50 (1.43 to 1.57)1.12 (1.07 to 1.17)-0.38 (-0.47 to -0.30)36% (28% to 44%)12% (3% to 21%)
IFSDay 51.60 (1.52 to 1.69)1.21 (1.15 to 1.27)-0.27 (-0.37 to -0.17)32% (24% to 40%)6% (−3% to 15%)
GFSDay 51.67 (1.57 to 1.76)1.25 (1.18 to 1.32)-0.07 (-0.16 to 0.02)29% (21% to 37%)3% (−7% to 11%)
PersistenceDay 52.36 (2.11 to 2.65)1.74 (1.57 to 1.96)0.01 (-0.21 to 0.24)−39% (−49% to −29%)
ClimatologyDay 51.71 (1.55 to 1.89)1.31 (1.20 to 1.44)-0.02 (-0.28 to 0.26)28% (23% to 33%)
AIFSDay 71.80 (1.68 to 1.92)1.33 (1.26 to 1.42)-0.45 (-0.58 to -0.31)27% (19% to 34%)−5% (−13% to 2%)
IFSDay 71.90 (1.79 to 2.01)1.42 (1.34 to 1.51)-0.34 (-0.48 to -0.18)23% (15% to 30%)−11% (−20% to −3%)
GFSDay 71.99 (1.86 to 2.14)1.49 (1.39 to 1.61)-0.06 (-0.21 to 0.11)19% (12% to 25%)−16% (−24% to −10%)
PersistenceDay 72.45 (2.21 to 2.72)1.83 (1.66 to 2.04)0.03 (-0.25 to 0.31)−44% (−55% to −34%)
ClimatologyDay 71.71 (1.55 to 1.90)1.31 (1.20 to 1.44)-0.02 (-0.28 to 0.25)31% (26% to 35%)
AIFSDay 102.19 (2.00 to 2.39)1.63 (1.51 to 1.78)-0.42 (-0.64 to -0.19)9% (1% to 16%)−28% (−36% to −21%)
IFSDay 102.19 (2.02 to 2.38)1.64 (1.51 to 1.77)-0.38 (-0.60 to -0.15)9% (2% to 16%)−28% (−36% to −21%)
GFSDay 102.26 (2.09 to 2.46)1.67 (1.56 to 1.81)-0.08 (-0.28 to 0.12)6% (−1% to 13%)−32% (−39% to −25%)
PersistenceDay 102.39 (2.19 to 2.60)1.76 (1.62 to 1.92)0.02 (-0.29 to 0.33)−41% (−51% to −32%)
ClimatologyDay 101.71 (1.55 to 1.90)1.31 (1.20 to 1.44)-0.02 (-0.28 to 0.24)29% (24% to 34%)

Persistence's own row and the skill against it rest on 174 days, 23 May 2026 to 12 November 2026, on the 10,081 of 10,351 station-times where persistence can be scored as well. Every other score needs no earlier measurement.

Summer and autumn 2026: 174 days from 23 May to 12 November 2026, 9 of them in spring, at 30 German stations, raw forecasts. AIFS 2 had the lower RMSE against IFS at 3 of 5 leads for temperature, 1 of 5 leads for wind and 3 of 3 leads for solar; against GFS at 3 of 5 leads for temperature, 3 of 5 leads for wind and 4 of 5 leads for solar. Of the rest, 11 could not be separated. The 28 comparisons share days and stations, so they are not 28 independent results. IFS and GFS are scored at 0.25°, as openly distributed.

Skill compares each model with a simple reference: climatology (the average for that date and hour) or persistence (the latest measurement carried forward). In the violet area a model beats that reference; skill is not tested, so the area is not the green of a verdict. IFS and GFS are scored at 0.25° and ICON at 0.125°, as openly distributed, not at their native resolution. Each AIFS version has its own days and seasons, so do not compare the two.

How much data this rests onDays scored, seasons still missing, and the smallest gap the sample can detect

No AIFS version has been running for a full year yet, so every result on this page is provisional. A result would stop being provisional once a version covers at least 365 days and every season.

Days scored, period, seasons missing and effective days, by AIFS version and variable
VersionVariableDays scoredPeriodSeasons still missingEffective days
AIFS 22 m temperature17423 May 2026 to 12 November 2026winter, springabout 37
AIFS 210 m wind speed17423 May 2026 to 12 November 2026winter, springabout 64
AIFS 2Solar energy17423 May 2026 to 12 November 2026winter, springabout 51
AIFS 1.12 m temperature2487 September 2025 to 12 May 2026summerabout 62
AIFS 1.110 m wind speed2487 September 2025 to 12 May 2026summerabout 93
AIFS 1.1Solar energy2477 September 2025 to 12 May 2026summerabout 87

The direction of the differences was steadier than their size. These counts cover raw forecasts at the 30 German stations, both AIFS versions, every lead and model compared. The sign was the same in both halves of the period in 52 of 56 comparisons, and AIFS had the lower error in 378 of 420 monthly comparisons. The smallest RMSE difference a sample would detect 80% of the time, after the Holm correction, depends on the version and the lead. For AIFS 2, over 174 days and leads of 1 to 10 days, it is 0.10 to 1.04 °C for 2 m temperature, 0.06 to 0.30 m/s for 10 m wind and 0.10 to 0.30 kWh/m²/day for daily solar energy. For AIFS 1.1, over 248 days and leads of 1 to 10 days, it is 0.19 to 0.93 °C for 2 m temperature, 0.06 to 0.49 m/s for 10 m wind and 0.05 to 0.28 kWh/m²/day for daily solar energy. A smaller real gap may not show as a verdict.

AIGFS at the same stationsNOAA's operational AI model, in a set of its own

AIGFS is NOAA's operational AI model. It starts from the GFS analysis, not from ECMWF's, and it publishes no surface height, so its temperature is moved to station height from the GFS surface height. It is scored here in a set of its own with AIFS, IFS and GFS, at the 30 German stations in the AIFS 2 period: 156 days, 23 May 2026 to 4 November 2026, with no winter or spring. Its archive has no solar, so this block shows 2 m temperature and 10 m wind at 00:00 and 12:00 UTC. The Reference and Forecasts switches apply; the Version switch does not.

2 m temperature: skill against climatology
-200204060▲ violet: lower error than climatologyDay 1Day 3Day 5Day 7Day 10Skill, %AIFSIFSAIGFSGFS

Raw forecasts, AIFS 2, 156 days, 23 May 2026 to 4 November 2026. Hover for values; AIFS shows its 95% range, others on hover.

2 m temperature, AIFS 2: how AIFS's error compared, over 156 days
Compared withDay 1Day 3Day 5Day 7Day 10
IFS−10%, AIFS lower error, tested−12%, AIFS lower error, tested−12%, AIFS lower error, tested−9%, not separated−11%, not separated
GFS−28%, AIFS lower error, tested−28%, AIFS lower error, tested−24%, AIFS lower error, tested−20%, not separated−16%, not separated
AIGFS−16%, AIFS lower error, tested−16%, AIFS lower error, tested−13%, AIFS lower error, tested−11%, not separated−14%, not separated

Each number is AIFS's RMSE minus the other model's, as a percentage of the other model's RMSE: negative means AIFS's was lower. Only comparisons with AIFS are tested, so the card does not say how AIGFS compares with GFS.

10 m wind speed: skill against climatology
-40-2002040▲ violet: lower error than climatologyDay 1Day 3Day 5Day 7Day 10Skill, %AIFSIFSAIGFSGFS

Raw forecasts, AIFS 2, 156 days, 23 May 2026 to 4 November 2026. Hover for values; AIFS shows its 95% range, others on hover.

10 m wind speed, AIFS 2: how AIFS's error compared, over 156 days
Compared withDay 1Day 3Day 5Day 7Day 10
IFS−1%, not separated−3%, not separated−7%, AIFS lower error, tested−5%, not separated0%, not separated
GFS−4%, not separated−7%, AIFS lower error, tested−11%, AIFS lower error, tested−9%, AIFS lower error, tested−3%, not separated
AIGFS−2%, not separated−5%, not separated−8%, not separated−6%, not separated0%, not separated

Each number is AIFS's RMSE minus the other model's, as a percentage of the other model's RMSE: negative means AIFS's was lower. Only comparisons with AIFS are tested, so the card does not say how AIGFS compares with GFS.

  • AIFS lower error, tested (better)
  • AIFS higher error, tested (worse)
  • Not separated: no clear difference, too few days, or the verdict depends on how the station is read or on the height correction (hover says which)

Bias, activity and events, with AIGFS

AIFS 2, the same 156 days. Bold and underlined marks the value nearest the ideal at each lead, as printed. It is not a verdict: these values are not tested against each other.

2 m temperature: bias, in °CForecast minus measurement, on average. Plus: too high. Minus: too low. Zero is best.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS+0.10+0.12+0.27+0.50−0.10
IFS+0.11+0.14+0.39+0.58+0.49
GFS−0.21−0.16−0.07+0.00−0.20
AIGFS−0.19−0.30−0.32−0.23−0.58
2 m temperature: activityVariability of the forecast divided by that of the measurements, both as departures from climatology, with each station centred on its own mean at each valid hour first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.980.981.000.990.94
IFS0.950.950.970.960.99
GFS0.950.960.960.970.98
AIGFS0.940.930.950.950.95

2 m temperature of 0 °C or belowObserved 1 times on 1 days in 9325 station-times (0.0%). Too few events to score: at least 30, on at least 10 days, are needed.

10 m wind: bias, in m/sForecast minus measurement, on average. Plus: too high. Minus: too low. Zero is best.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS−0.26−0.35−0.38−0.47−0.41
IFS−0.21−0.27−0.27−0.35−0.42
GFS+0.04−0.04−0.05−0.09−0.11
AIGFS−0.01−0.12−0.23−0.34−0.24
10 m wind: activityVariability of the forecast divided by that of the measurements, both as departures from climatology, with each station centred on its own mean at each valid hour first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.920.890.850.820.86
IFS0.950.940.920.870.85
GFS1.010.970.920.940.94
AIGFS1.000.950.880.870.87
10 m wind of 8 m/s or more: frequency biasObserved 274 times on 71 days in 9283 station-times (3%). Forecast events per observed event. 1.00 is best. Above 1 the model calls the event too often, below 1 too rarely.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.470.380.340.260.37
IFS0.620.640.570.480.43
GFS1.111.020.880.910.80
AIGFS0.850.650.460.340.50
10 m wind of 8 m/s or more: equitable threat scoreHow well forecast and observed events coincide, beyond chance. 1 is perfect, 0 is chance, higher is better.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.300.230.180.120.05
IFS0.340.300.200.130.07
GFS0.360.270.210.140.10
AIGFS0.330.230.130.090.04
10 m wind of 10.8 m/s or more: frequency biasObserved 64 times on 31 days in 9283 station-times (0.7%). Forecast events per observed event. 1.00 is best. Above 1 the model calls the event too often, below 1 too rarely.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.160.170.080.110.19
IFS0.670.610.720.500.64
GFS0.700.630.550.560.52
AIGFS0.440.340.110.130.08
10 m wind of 10.8 m/s or more: equitable threat scoreHow well forecast and observed events coincide, beyond chance. 1 is perfect, 0 is chance, higher is better.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.150.150.080.06-0.00
IFS0.300.240.170.060.08
GFS0.340.270.120.060.10
AIGFS0.310.190.060.010.01
Other AI models at the same stationsGraphCast, FourCastNet, Pangu-Weather and Aurora, as re-run by NOAA

GraphCast, FourCastNet v2-small (not the newer FourCastNet 3), Pangu-Weather and Aurora are NOAA's re-runs from its open AIWP archive, started from IFS analyses. They are not the developers' own forecasts. AIGFS, AIFS and IFS are shown on the same days; AIGFS has more days in its own block above. The archive has no solar and covers only part of the AIFS 2 period, so this panel shows 2 m temperature and 10 m wind at 00:00 and 12:00 UTC. It rests on 68 days, with no winter or spring, fewer than the 15 blocks a verdict needs, so it shows differences without verdicts. The Reference and Forecasts switches apply; the Version switch does not.

2 m temperature
-50-2502550▲ violet: lower error than climatologyDay 1Day 3Day 5Day 7Day 10Skill, %

Raw forecasts against climatology, AIFS 2. 68 days, 21 July 2026 to 4 November 2026. Bars are 95% intervals.

2 m temperature, AIFS 2: how AIFS's error compared, over 68 days
Compared withDay 1Day 3Day 5Day 7Day 10
IFS−11%, not separated−14%, not separated−14%, not separated−2%, not separated−13%, not separated
AIGFS−16%, not separated−15%, not separated−10%, not separated−10%, not separated−17%, not separated
GraphCast−4%, not separated−5%, not separated−1%, not separated−2%, not separated−5%, not separated
FourCastNet v2−22%, not separated−21%, not separated−14%, not separated−2%, not separated−7%, not separated
Pangu−9%, not separated−11%, not separated−10%, not separated−11%, not separated−9%, not separated
Aurora−4%, not separated−4%, not separated0%, not separated+9%, not separated+11%, not separated
  • AIFS lower error, tested (better)
  • AIFS higher error, tested (worse)
  • Not separated: no clear difference, too few days, or the verdict depends on how the station is read or on the height correction (hover says which)
10 m wind speed
-60-40-2002040▲ violet: lower error than climatologyDay 1Day 3Day 5Day 7Day 10Skill, %

Raw forecasts against climatology, AIFS 2. 68 days, 21 July 2026 to 4 November 2026. Bars are 95% intervals.

10 m wind speed, AIFS 2: how AIFS's error compared, over 68 days
Compared withDay 1Day 3Day 5Day 7Day 10
IFS−3%, not separated−4%, not separated−6%, not separated−5%, not separated+5%, not separated
AIGFS−2%, not separated−2%, not separated−6%, not separated−7%, not separated+4%, not separated
GraphCast−2%, not separated0%, not separated−4%, not separated0%, not separated+13%, not separated
FourCastNet v2−12%, not separated−11%, not separated−9%, not separated−7%, not separated−4%, not separated
Pangu−4%, not separated−4%, not separated−6%, not separated−7%, not separated+6%, not separated
Aurora+2%, not separated+2%, not separated−4%, not separated−11%, not separated−5%, not separated
  • AIFS lower error, tested (better)
  • AIFS higher error, tested (worse)
  • Not separated: no clear difference, too few days, or the verdict depends on how the station is read or on the height correction (hover says which)

Bias, activity and events, other AI models

AIFS 2, the same 68 days. Bold and underlined marks the value nearest the ideal at each lead, as printed. It is not a verdict: these values are not tested against each other.

2 m temperature: bias, in °CForecast minus measurement, on average. Plus: too high. Minus: too low. Zero is best.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS−0.01−0.02+0.01+0.24−0.02
IFS+0.10+0.17+0.46+0.76+0.20
AIGFS−0.16−0.32−0.38−0.23−0.69
GraphCast+0.20+0.040.00+0.45−0.35
FourCastNet v2+0.24−0.17−0.44−0.37−0.24
Pangu+0.24+0.16+0.17+0.51+0.07
Aurora+0.24+0.15+0.18+0.24+0.43
2 m temperature: activityVariability of the forecast divided by that of the measurements, both as departures from climatology, with each station centred on its own mean at each valid hour first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.950.910.920.970.97
IFS0.930.910.920.991.03
AIGFS0.910.870.871.010.99
GraphCast0.920.890.880.990.95
FourCastNet v20.850.760.750.780.81
Pangu0.930.920.931.000.97
Aurora0.930.870.810.790.71

2 m temperature of 0 °C or belowObserved 1 times on 1 days in 4065 station-times (0.0%). Too few events to score: at least 30, on at least 10 days, are needed.

10 m wind: bias, in m/sForecast minus measurement, on average. Plus: too high. Minus: too low. Zero is best.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS−0.24−0.36−0.41−0.50−0.40
IFS−0.21−0.28−0.33−0.35−0.50
AIGFS−0.03−0.20−0.27−0.31−0.43
GraphCast−0.21−0.27−0.28−0.36−0.31
FourCastNet v2−0.13−0.30−0.38−0.55−0.61
Pangu−0.20−0.26−0.22−0.28−0.46
Aurora−0.19−0.34−0.68−1.14−1.56
10 m wind: activityVariability of the forecast divided by that of the measurements, both as departures from climatology, with each station centred on its own mean at each valid hour first. 1.00 is best. Below 1 the forecast is too smooth, which flatters RMSE.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.920.890.850.860.93
IFS0.950.940.940.910.82
AIGFS0.990.940.910.940.84
GraphCast0.910.900.890.870.78
FourCastNet v20.960.910.850.860.87
Pangu0.900.870.890.880.91
Aurora0.890.850.760.650.51
10 m wind of 8 m/s or more: frequency biasObserved 103 times on 34 days in 4059 station-times (3%). Forecast events per observed event. 1.00 is best. Above 1 the model calls the event too often, below 1 too rarely.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.410.310.270.230.52
IFS0.590.610.640.590.38
AIGFS0.870.710.560.390.32
GraphCast0.440.350.380.320.28
FourCastNet v20.670.450.370.300.38
Pangu0.450.400.480.520.59
Aurora0.420.310.150.020.00
10 m wind of 8 m/s or more: equitable threat scoreHow well forecast and observed events coincide, beyond chance. 1 is perfect, 0 is chance, higher is better.
ModelDay 1Day 3Day 5Day 7Day 10
AIFS0.250.200.110.100.04
IFS0.270.230.210.140.05
AIGFS0.300.240.160.120.06
GraphCast0.260.230.140.120.10
FourCastNet v20.270.180.130.070.01
Pangu0.240.190.130.130.08
Aurora0.300.230.100.000.00

10 m wind of 10.8 m/s or moreObserved 17 times on 11 days in 4059 station-times (0.4%). Too few events to score: at least 30, on at least 10 days, are needed.

The figures, written outThe headline numbers for AIFS 2, raw forecasts, as sentences
  • After bias correction, none of the 28 comparisons can be judged yet. That sample holds 14.4 blocks of 10 days; a verdict needs 15.
  • Three days ahead, AIFS 2's RMSE was 66% below persistence for temperature, 42% for wind and 60% for solar; against climatology, 61%, 24% and 47%. The 2015–2024 station climatology is not adjusted for warming and ran 0.9 K too cold over these days, which raises every model's temperature skill against it. By day 7, its temperature RMSE was still 29% below climatology's; its wind RMSE could not be separated from climatology's, with IFS and GFS scoring lower against climatology.
  • RMSE favours smooth forecasts. At day 10, AIFS 2 carried 65% of the measured variability in daily solar energy, against 86% for GFS. Strong wind and frost are scored as events in the fold "Bias, activity and events".
  • AIFS 2, 2 m temperature, day 3, over 174 days: skill against climatology 61% (55% to 65%) and against persistence 66% (62% to 68%). Month by month, skill against climatology ranged from 52% to 70%. On the 144 days that can be bias-corrected, the correction moves RMSE from 1.60 to 1.49 °C.
  • AIFS 2, 10 m wind speed, day 3, over 174 days: skill against climatology 24% (14% to 32%) and against persistence 42% (36% to 48%). Month by month, skill against climatology ranged from 8% to 41%. On the 144 days that can be bias-corrected, the correction moves RMSE from 1.31 to 1.14 m/s.
  • AIFS 2, solar energy, day 3, over 174 days: skill against climatology 47% (41% to 53%) and against persistence 60% (56% to 63%). Month by month, skill against climatology ranged from 26% to 51%. On the 143 days that can be bias-corrected, the correction moves RMSE from 0.77 to 0.78 kWh/m²/day.
  • AIFS 1.1, 2 m temperature, day 3, over 248 days: skill against climatology 61% (56% to 65%) and against persistence 67% (64% to 69%). Month by month, skill against climatology ranged from 46% to 69%. On the 218 days that can be bias-corrected, the correction moves RMSE from 1.53 to 1.48 °C.
  • AIFS 1.1, 10 m wind speed, day 3, over 248 days: skill against climatology 29% (25% to 33%) and against persistence 47% (43% to 50%). Month by month, skill against climatology ranged from 8% to 37%. On the 218 days that can be bias-corrected, the correction moves RMSE from 1.41 to 1.27 m/s.
  • AIFS 1.1, solar energy, day 3, over 247 days: skill against climatology 45% (38% to 50%) and against persistence 59% (54% to 62%). Month by month, skill against climatology ranged from 22% to 52%. On the 216 days that can be bias-corrected, the correction moves RMSE from 0.58 to 0.59 kWh/m²/day.