Tracking forecast error by month, updated every week
RMSE of each model relative to IFS, at stations and against ERA5, from open data.
- Scores grouped by
- Calendar month
- Refreshed
- Every week, on Mondays
- Last update
- 16 November 2026
- Station observations to
- 12 November 2026
- ERA5 scores to
- 9 November 2026
- Models scored
- 8: AIFS and seven others, 6 of them AI
Each value is the error over one calendar month. The weekly update adds the new days, so the current month fills up week by week and appears once it holds 10 days. A finished month can still move a little when a source revises its measurements. Monthly values are point scores without confidence intervals. They show how the error moves with the season and with model versions. Whether a difference holds is decided only on the whole sample, on the scores pages.
RMSE by model and lead time, latest month scored.
Raw model output. The percentage beside each RMSE compares it with IFS over the same days. A month is shown once it holds at least 10 days.
−8% RMSE 8% lower than IFS's: better. +8% RMSE 8% higher than IFS's: worse. These monthly differences are not tested, so violet and yellow-brown show the sign only, not the green or red of a tested result.
Against observations at 30 DWD stations, Germany: operational models
Every day with a forecast from AIFS, IFS and GFS.
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 1.34−14% (lower than IFS, better) | 1.45−27% (lower than IFS, better) | 1.80−16% (lower than IFS, better) | 2.15−23% (lower than IFS, better) | 2.67−18% (lower than IFS, better) |
| IFS | 1.56 | 2.00 | 2.16 | 2.81 | 3.28 |
| GFS | 2.17+39% (higher than IFS, worse) | 2.46+23% (higher than IFS, worse) | 2.61+21% (higher than IFS, worse) | 3.03+8% (higher than IFS, worse) | 3.59+9% (higher than IFS, worse) |
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 1.05−5% (lower than IFS, better) | 1.09−11% (lower than IFS, better) | 1.35−8% (lower than IFS, better) | 1.52−16% (lower than IFS, better) | 1.91+2% (higher than IFS, worse) |
| IFS | 1.11 | 1.22 | 1.47 | 1.80 | 1.87 |
| GFS | 1.09−1% (lower than IFS, better) | 1.27+4% (higher than IFS, worse) | 1.55+6% (higher than IFS, worse) | 2.07+15% (higher than IFS, worse) | 2.13+14% (higher than IFS, worse) |
Against observations at 107 stations in six countries, Europe: operational models
Every national-service station, Germany included, pooled.
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 1.37−10% (lower than IFS, better) | 1.51−19% (lower than IFS, better) | 1.83−16% (lower than IFS, better) | 2.23−15% (lower than IFS, better) | 2.73−15% (lower than IFS, better) |
| IFS | 1.52 | 1.86 | 2.17 | 2.62 | 3.21 |
| GFS | 2.08+37% (higher than IFS, worse) | 2.43+31% (higher than IFS, worse) | 2.57+19% (higher than IFS, worse) | 2.99+14% (higher than IFS, worse) | 3.66+14% (higher than IFS, worse) |
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 1.14−4% (lower than IFS, better) | 1.18−8% (lower than IFS, better) | 1.44−9% (lower than IFS, better) | 1.67−14% (lower than IFS, better) | 1.96−4% (lower than IFS, better) |
| IFS | 1.19 | 1.27 | 1.58 | 1.93 | 2.04 |
| GFS | 1.29+9% (higher than IFS, worse) | 1.42+12% (higher than IFS, worse) | 1.68+6% (higher than IFS, worse) | 2.32+20% (higher than IFS, worse) | 2.43+19% (higher than IFS, worse) |
Against observations at 30 DWD stations, Germany: AI models
Only the days on which AIGFS, GraphCast, FourCastNet v2, Pangu and Aurora also have a forecast, so the months hold fewer days than above. AIGFS is NOAA's operational model and starts from the GFS analysis. The other four are NOAA's re-runs, started from IFS analyses.
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 1.27+1% (higher than IFS, worse) | 1.47−7% (lower than IFS, better) | 2.13−8% (lower than IFS, better) | 3.33+45% (higher than IFS, worse) | 3.61−6% (lower than IFS, better) |
| IFS | 1.26 | 1.58 | 2.32 | 2.29 | 3.86 |
| AIGFSAI | 1.58+26% (higher than IFS, worse) | 1.78+12% (higher than IFS, worse) | 2.25−3% (lower than IFS, better) | 2.76+21% (higher than IFS, worse) | 2.97−23% (lower than IFS, better) |
| GraphCastAI | 1.30+3% (higher than IFS, worse) | 1.67+6% (higher than IFS, worse) | 1.96−16% (lower than IFS, better) | 3.01+32% (higher than IFS, worse) | 3.27−15% (lower than IFS, better) |
| FourCastNet v2AI | 1.58+25% (higher than IFS, worse) | 1.81+14% (higher than IFS, worse) | 2.33+1% (higher than IFS, worse) | 3.00+31% (higher than IFS, worse) | 4.11+6% (higher than IFS, worse) |
| PanguAI | 1.41+12% (higher than IFS, worse) | 1.96+24% (higher than IFS, worse) | 2.57+11% (higher than IFS, worse) | 3.63+59% (higher than IFS, worse) | 3.42−11% (lower than IFS, better) |
| AuroraAI | 1.34+6% (higher than IFS, worse) | 1.71+8% (higher than IFS, worse) | 2.07−11% (lower than IFS, better) | 2.54+11% (higher than IFS, worse) | 3.22−17% (lower than IFS, better) |
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 1.09−2% (lower than IFS, better) | 1.14−8% (lower than IFS, better) | 1.36−6% (lower than IFS, better) | 1.64−6% (lower than IFS, better) | 2.59+34% (higher than IFS, worse) |
| IFS | 1.11 | 1.23 | 1.45 | 1.74 | 1.93 |
| AIGFSAI | 1.12+1% (higher than IFS, worse) | 1.20−3% (lower than IFS, better) | 1.47+1% (higher than IFS, worse) | 1.87+7% (higher than IFS, worse) | 2.20+14% (higher than IFS, worse) |
| GraphCastAI | 1.14+2% (higher than IFS, worse) | 1.16−6% (lower than IFS, better) | 1.37−5% (lower than IFS, better) | 1.75+1% (higher than IFS, worse) | 1.91−1% (lower than IFS, better) |
| FourCastNet v2AI | 1.24+12% (higher than IFS, worse) | 1.41+15% (higher than IFS, worse) | 1.69+16% (higher than IFS, worse) | 1.86+7% (higher than IFS, worse) | 2.50+30% (higher than IFS, worse) |
| PanguAI | 1.15+3% (higher than IFS, worse) | 1.17−5% (lower than IFS, better) | 1.48+1% (higher than IFS, worse) | 1.90+9% (higher than IFS, worse) | 1.85−4% (lower than IFS, better) |
| AuroraAI | 1.08−3% (lower than IFS, better) | 1.15−7% (lower than IFS, better) | 1.41−3% (lower than IFS, better) | 1.86+7% (higher than IFS, worse) | 2.07+7% (higher than IFS, worse) |
Against ERA5 over Europe, 35 to 72°N, 12.5°W to 40°E
Latitude-weighted, on the days every model has a forecast. ERA5 is a model product and AIFS was trained mostly on it, so this is a second view, not an observation.
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 4.7+9% (higher than IFS, worse) | 15.8−1% (lower than IFS, better) | 33.1−17% (lower than IFS, better) | 75.6−1% (lower than IFS, better) | 130.9−2% (lower than IFS, better) |
| IFS | 4.3 | 16.0 | 40.0 | 76.2 | 133.6 |
| GFS | 7.3+68% (higher than IFS, worse) | 23.1+44% (higher than IFS, worse) | 51.1+28% (higher than IFS, worse) | 86.9+14% (higher than IFS, worse) | 140.4+5% (higher than IFS, worse) |
| AIGFSAI | 7.5+72% (higher than IFS, worse) | 23.3+45% (higher than IFS, worse) | 46.0+15% (higher than IFS, worse) | 75.0−2% (lower than IFS, better) | 128.8−4% (lower than IFS, better) |
| GraphCastAI | 6.2+43% (higher than IFS, worse) | 18.5+15% (higher than IFS, worse) | 40.10% (equal to IFS) | 78.8+3% (higher than IFS, worse) | 131.2−2% (lower than IFS, better) |
| FourCastNet v2AI | 7.5+72% (higher than IFS, worse) | 25.1+56% (higher than IFS, worse) | 51.1+28% (higher than IFS, worse) | 97.1+27% (higher than IFS, worse) | 142.1+6% (higher than IFS, worse) |
| PanguAI | 6.0+38% (higher than IFS, worse) | 20.8+30% (higher than IFS, worse) | 43.7+9% (higher than IFS, worse) | 87.7+15% (higher than IFS, worse) | 133.00% (equal to IFS) |
| AuroraAI | 4.4+2% (higher than IFS, worse) | 16.9+6% (higher than IFS, worse) | 38.1−5% (lower than IFS, better) | 69.7−9% (lower than IFS, better) | 110.5−17% (lower than IFS, better) |
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 0.76−34% (lower than IFS, better) | 0.99−27% (lower than IFS, better) | 1.43−20% (lower than IFS, better) | 2.08−11% (lower than IFS, better) | 2.93−5% (lower than IFS, better) |
| IFS | 1.14 | 1.35 | 1.79 | 2.32 | 3.09 |
| GFS | 1.32+16% (higher than IFS, worse) | 1.58+17% (higher than IFS, worse) | 2.02+13% (higher than IFS, worse) | 2.43+5% (higher than IFS, worse) | 3.04−2% (lower than IFS, better) |
| AIGFSAI | 1.140% (equal to IFS) | 1.37+2% (higher than IFS, worse) | 1.780% (equal to IFS) | 2.17−6% (lower than IFS, better) | 2.88−7% (lower than IFS, better) |
| GraphCastAI | 1.05−8% (lower than IFS, better) | 1.30−3% (lower than IFS, better) | 1.77−1% (lower than IFS, better) | 2.31−1% (lower than IFS, better) | 2.91−6% (lower than IFS, better) |
| FourCastNet v2AI | 1.27+12% (higher than IFS, worse) | 1.59+18% (higher than IFS, worse) | 1.92+7% (higher than IFS, worse) | 2.46+6% (higher than IFS, worse) | 3.14+2% (higher than IFS, worse) |
| PanguAI | 0.92−19% (lower than IFS, better) | 1.21−10% (lower than IFS, better) | 1.74−2% (lower than IFS, better) | 2.35+1% (higher than IFS, worse) | 3.15+2% (higher than IFS, worse) |
| AuroraAI | 0.97−15% (lower than IFS, better) | 1.19−12% (lower than IFS, better) | 1.60−10% (lower than IFS, better) | 2.00−14% (lower than IFS, better) | 2.56−17% (lower than IFS, better) |
| Model | Day 1 | Day 3 | Day 5 | Day 7 | Day 10 |
|---|---|---|---|---|---|
| AIFSAI | 0.76−28% (lower than IFS, better) | 1.23−16% (lower than IFS, better) | 1.80−18% (lower than IFS, better) | 2.78−3% (lower than IFS, better) | 3.26−3% (lower than IFS, better) |
| IFS | 1.06 | 1.47 | 2.20 | 2.88 | 3.37 |
| GFS | 1.41+33% (higher than IFS, worse) | 1.91+30% (higher than IFS, worse) | 2.51+14% (higher than IFS, worse) | 3.16+10% (higher than IFS, worse) | 3.78+12% (higher than IFS, worse) |
| AIGFSAI | 1.16+10% (higher than IFS, worse) | 1.59+8% (higher than IFS, worse) | 2.26+3% (higher than IFS, worse) | 2.880% (equal to IFS) | 3.40+1% (higher than IFS, worse) |
| GraphCastAI | 0.87−18% (lower than IFS, better) | 1.35−8% (lower than IFS, better) | 2.04−7% (lower than IFS, better) | 2.76−4% (lower than IFS, better) | 3.35−1% (lower than IFS, better) |
| FourCastNet v2AI | 1.17+10% (higher than IFS, worse) | 1.70+16% (higher than IFS, worse) | 2.31+5% (higher than IFS, worse) | 2.93+2% (higher than IFS, worse) | 3.53+5% (higher than IFS, worse) |
| PanguAI | 0.92−13% (lower than IFS, better) | 1.44−2% (lower than IFS, better) | 2.10−5% (lower than IFS, better) | 2.86−1% (lower than IFS, better) | 3.39+1% (higher than IFS, worse) |
| AuroraAI | 0.80−24% (lower than IFS, better) | 1.26−14% (lower than IFS, better) | 2.13−3% (lower than IFS, better) | 3.17+10% (higher than IFS, worse) | 3.88+15% (higher than IFS, worse) |
RMSE relative to IFS at day 3.
The zero line is IFS. Below the line, on the violet side, a model had a lower RMSE than IFS in that month over the same days. Above the line its RMSE was higher. These monthly differences are not tested. AIFS 2 replaced the earlier version on 12 May 2026. A month is scored for a version only when that version has at least 10 days in it, so May 2026 shows the earlier version alone. A month with no scores leaves a gap on the axis.
Against observations at 30 DWD stations, Germany: operational models
Every day with a forecast from AIFS, IFS and GFS.
Sep 2025 to Nov 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Sep 2025 to Nov 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Against observations at 107 stations in six countries, Europe: operational models
Every national-service station, Germany included, pooled.
Sep 2025 to Nov 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Sep 2025 to Nov 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Against observations at 30 DWD stations, Germany: with ICON, up to 7 November 2026
ICON's archive of past runs ended on 7 November 2026, so this group no longer grows.
Sep 2025 to Oct 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Sep 2025 to Oct 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Against observations at 30 DWD stations, Germany: AI models
Only the days on which AIGFS, GraphCast, FourCastNet v2, Pangu and Aurora also have a forecast, so the months hold fewer days than above. AIGFS is NOAA's operational model and starts from the GFS analysis. The other four are NOAA's re-runs, started from IFS analyses.
Jul 2026 to Oct 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Jul 2026 to Oct 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Against ERA5 over Europe, 35 to 72°N, 12.5°W to 40°E
Latitude-weighted, on the days every model has a forecast. ERA5 is a model product and AIFS was trained mostly on it, so this is a second view, not an observation.
Sep 2025 to Oct 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Sep 2025 to Oct 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Sep 2025 to Oct 2026. Hover for each month's values: minus is a lower RMSE than IFS's (better), plus a higher one (worse).
Day 1The same charts for forecasts made 1 day ahead
Against observations at 30 DWD stations, Germany: operational models
Against observations at 107 stations in six countries, Europe: operational models
Against observations at 30 DWD stations, Germany: with ICON, up to 7 November 2026
Against observations at 30 DWD stations, Germany: AI models
Against ERA5 over Europe, 35 to 72°N, 12.5°W to 40°E
Day 5The same charts for forecasts made 5 days ahead
Against observations at 30 DWD stations, Germany: operational models
Against observations at 107 stations in six countries, Europe: operational models
Against observations at 30 DWD stations, Germany: with ICON, up to 7 November 2026
Against observations at 30 DWD stations, Germany: AI models
Against ERA5 over Europe, 35 to 72°N, 12.5°W to 40°E
Day 7The same charts for forecasts made 7 days ahead
Against observations at 30 DWD stations, Germany: operational models
Against observations at 107 stations in six countries, Europe: operational models
Against observations at 30 DWD stations, Germany: with ICON, up to 7 November 2026
Against observations at 30 DWD stations, Germany: AI models
Against ERA5 over Europe, 35 to 72°N, 12.5°W to 40°E
Day 10The same charts for forecasts made 10 days ahead
Against observations at 30 DWD stations, Germany: operational models
Against observations at 107 stations in six countries, Europe: operational models
Against observations at 30 DWD stations, Germany: with ICON, up to 7 November 2026
Against observations at 30 DWD stations, Germany: AI models
Against ERA5 over Europe, 35 to 72°N, 12.5°W to 40°E
Eight scored, two being collected, three open and waiting.
A model is scored here only if its forecast files are published routinely, free of charge and without an account. File listings were last checked on 10 November 2026.
| Model | Type | Open feed | Starts from | Status |
|---|---|---|---|---|
| AIFS SingleECMWF | AI | ECMWF open data, operational, four runs a day | ECMWF's own analysis | ScoredThe model every comparison on this site is made from. Versions 1.1 and 2 are scored apart and never pooled. |
| IFSECMWF | Physics-based | ECMWF open data, operational | ECMWF's own analysis | ScoredThe reference line on the tracking charts. |
| GFSNOAA/NCEP | Physics-based | NOAA open data on AWS, operational | GDAS/GFS analysis | Scored |
| AIGFSNOAA/NCEP | AI | NOAA open data on AWS, operational, four runs a day | GDAS/GFS analysis | ScoredArchive on AWS from 16 April 2026, so in the AIFS 2 samples only. No surface height is published: station temperature uses the GFS height. |
| GraphCastGoogle DeepMind, re-run by NOAA | AI | NOAA AIWP archive on AWS (GRAP_v100) | IFS analysis | ScoredA re-run, not an operational service: scored on fewer days. The version is the one NOAA's archive labels v100. |
| FourCastNet v2-smallNVIDIA, re-run by NOAA | AI | NOAA AIWP archive on AWS (FOUR_v200) | IFS analysis | ScoredA re-run, not an operational service: scored on fewer days. This is version 2 (small), not the newer FourCastNet 3. |
| Pangu-WeatherHuawei, re-run by NOAA | AI | NOAA AIWP archive on AWS (PANG_v100) | IFS analysis | ScoredA re-run, not an operational service: scored on fewer days. The version is the one NOAA's archive labels v100. |
| AuroraMicrosoft, re-run by NOAA | AI | NOAA AIWP archive on AWS (AURO_v100) | IFS analysis | ScoredA re-run, not an operational service: scored on fewer days. The version is the one NOAA's archive labels v100. |
| ICONDWD | Physics-based | DWD open data, kept for about a day | DWD's own analysis | Collecting runs32 runs kept, from 8 November 2026, 18 UTC.Scored up to 7 November 2026 from Open-Meteo's archive, as a set that no longer grows. A new series from DWD's own files is being collected. |
| AICONDWD | AI | DWD open data, kept for about a day | DWD's own analysis | Collecting runs32 runs kept, from 8 November 2026, 18 UTC.No public archive, so runs are captured as they appear. Scores follow once the sample is long enough. |
| AIFS ENSECMWF | AI ensemble | ECMWF open data, operational | ECMWF's own analysis, perturbed | Open, not yet scoredArchive from 2 July 2025. Needs ensemble scores (CRPS, spread and error), which this site does not compute yet. |
| AIGEFSNOAA/NCEP | AI ensemble | NOAA open data on AWS, operational, 31 members | GEFS member analyses | Open, not yet scoredArchive from 1 June 2025. Needs ensemble scores. |
| GDPS-GEMLECCC | AI | MSC Datamart, labelled experimental, kept for 30 days | ECCC's own analysis | Open, not yet scoredNo long archive, so it would have to be captured as it appears. |
| FourCastNet 3NVIDIA | AI ensemble | Model and weights only | – | No open forecast feedNo routine forecast files were found, and NOAA's AIWP archive holds versions 1 and 2 only (checked 11 November 2026). It could be scored only by running the model here. |
| WeatherNext 2Google | AI ensemble | Google Cloud, with an account and a request form | – | No open forecast feedNot available without an account, so outside the open-data rule of this site. |
| FastNetMet Office | AI | Model weights only | – | No open forecast feedNo routine forecast files were found. |
| FuXi, FengWuFudan University, Shanghai AI Laboratory | AI | Code, weights and sample files only | – | No open forecast feedNo routine forecast files were found. |
What each update changed.
The scores are recomputed every Monday, and the same run publishes the site when its checks pass. If a source is late, the affected scores keep their earlier end date; if a run fails, the earlier results stay in place. The date under "Last update" says which run you are reading. Each update fetches the new observations, forecasts and ERA5 fields and rescores everything from the raw files. ERA5 arrives about five days late, so the area scores end earlier than the station scores. Verdicts rest on blocks of 10 days or more and change slowly; the monthly values above change with every update.
| Update | Stations to | ERA5 to | 10 m wind, stations | 2 m temperature, stations | 500 hPa height, ERA5 | 2 m temperature, ERA5 |
|---|---|---|---|---|---|---|
| 16 November 2026 | 12 November 2026174 days | 9 November 2026171 days | −3.0%cannot be separated | −13.2%AIFS lower error | −7.9%AIFS lower error | −28.6%no verdict |
| 12 November 2026 | 8 November 2026170 days | 5 November 2026167 days | −2.9%cannot be separated | −12.3%AIFS lower error | −8.2%AIFS lower error | −28.7%no verdict |
| 11 November 2026 | 7 November 2026169 days | 3 November 2026165 days | −2.9%cannot be separated | −12.2%AIFS lower error | −8.2%AIFS lower error | −28.7%no verdict |
| 10 November 2026 | 7 November 2026172 days | 3 November 2026168 days | −6.3%depends on how the station is read | −13.4%AIFS lower error | −8.1%AIFS lower error | −28.7%no verdict |
- 16 November 2026. The verdict level, the sample rule for persistence, activity and solar persistence changed on this day, so verdicts and figures differ from the update before without new data. The changes are listed on the method page.
- 11 November 2026. The scored reading became interpolation, day 10 was added as a lead time and the intervals were widened on this day, so the sample and the figures differ from the update before without new data. The changes are listed on the method page.
The verdicts are tested again at every update, so a verdict needs a Holm-adjusted p-value below 0.01, and every verdict is interim until a version covers a full year.
The log starts on 10 November 2026. Earlier results are in the scores themselves, which are always rebuilt from the first day of each AIFS version.