Monthly reviews · 2026-09
September 2026
30 calendar days. Over this month's days alone, the lowest mean absolute error at lead day 1 was 2.24 °F — ECMWF AIFS Single, on 30 of the 30 days. That is the lowest among the models scored in this month, each over its own days; it is not a claim that one model is better than another. 11 forecast systems were scored, over 23 stations. This page is generated from the tables below; it contains no assessment of why any number is what it is.
Ranked on this month's dayslead day 1
MAE and bias are recomputed over 2026-09 alone, with the same two collapses as every published number — the mean over the four synoptic instants of a station-day, then the mean over the stations present that day (methodology §4). A calendar month is not one of the published rolling windows, so no confidence interval is quoted here: a month is 28–31 days, close to the minimum a bootstrap needs, and an interval printed at that length would suggest more precision than the sample carries. Every row links to the permanent page where the intervals over the published windows do live.
| # | Model | MAE °F | Bias °F | Days | Coverage | Stations |
|---|---|---|---|---|---|---|
| 1 | ECMWF AIFS Singleaifs_single | 2.24 | −1.11 | 30 | 100% | 23.0 |
| 2 | ECMWF IFS HRESifs_hres | 2.38 | −0.75 | 30 | 100% | 23.0 |
| 3 | Aurora (IFS init)aurora_ifs | 2.65 | −1.78 | 26 | 87% | 23.0 |
| 4 | GraphCast (IFS init)graphcast_ifs | 2.65 | −1.84 | 29 | 97% | 23.0 |
| 5 | NCEP GFSgfs | 2.68 | +0.02 | 30 | 100% | 23.0 |
| 6 | Pangu-Weather (GFS init)pangu_gfs | 2.71 | −1.20 | 20 | 67% | 23.0 |
| 7 | Pangu-Weather (IFS init)pangu_ifs | 2.80 | −1.35 | 27 | 90% | 23.0 |
| 8 | FourCastNet v2 (IFS init)fourcastnet_ifs | 2.86 | −1.43 | 28 | 93% | 23.0 |
| 9 | FourCastNet v2 (GFS init)fourcastnet_gfs | 2.91 | −1.60 | 20 | 67% | 23.0 |
| 10 | GraphCast (GFS init)graphcast_gfs | 2.99 | −2.15 | 20 | 67% | 23.0 |
| 11 | Aurora (GFS init)aurora_gfs | 3.00 | −2.26 | 20 | 67% | 23.0 |
How complete the month was
Counted against what a fully observed month would hold: 23 stations × 30 days. A rate below 100 % is missing observation, not missing forecast; which forecast runs arrived is on the status page, and how many days each model was scored on is the Days column above.
| Observation | Present | Expected | Rate | What one is |
|---|---|---|---|---|
| Instantaneous observations | 2760 | 2760 | 100.0% | one routine METAR per station at each of 00/06/12/18 UTC |
| Daily-extreme reports | 690 | 690 | 100.0% | one NWS Daily Climate Report (or CF6 fallback) per station-day |
Quality-control events741 in the month
Every flag raised on an observation whose timestamp falls in this month. A flagged observation is kept and scored — the flag travels with it into the published tables rather than removing the day (methodology §6). One observation can carry more than one flag, and each is counted separately.
| Flag | Count |
|---|---|
| obs_fallback (daily extreme) | 690 |
| obs_diff_gt2f (daily extreme) | 43 |
| suspect (instant) | 3 |
| cf6_missing_value (daily extreme) | 1 |
| cli_implausible (daily extreme) | 1 |
| cli_implausible_dropped (daily extreme) | 1 |
| cli_missing_value (daily extreme) | 1 |
| gap_gt35min (instant) | 1 |
Largest single-day error
The biggest absolute error any one model made at any one station on any one day of the month, at lead day 1. It is a single station-day, not evidence about a model: a station-day is the smallest unit the site scores and the tails of a month of them are wide.
Open the permanent page for KPHX · graphcast_gfs · lead 1 — the daily-error chart there plots this day among the rest, and the address does not change.
Upstream changes0 in the month
A model version that appears for the first time in this month. Days either side of a version change are not pooled: scores are computed on the latest version segment only (methodology §9), so a change here resets the record that the leaderboard shows for that model.
No model version appeared for the first time in this month.