The interesting relationships live off the diagonal.
Every cell is measured on the cases two models both answered, in our own ledger. Cells rest on different shared-case sets, so this is a map of relationships — not a league table.
public snapshot public-20260913140717-080dd4d9 · as of 2026-09-13 14:07 UTC
sea · temperature_2m · 10 models with measured shared cases · 1,474 cases over 30 days
Model matrix
each square is one pair of models on the cases they both answered — tap one to open it
Pairwise comparison of models on the cases they both answered, temperature_2m at +24h
row vs column
ECMWF AIFS 0.25° single
ECMWF IFS 0.25°
NOAA GFS
DWD ICON
JMA seamless
Météo-France seamless
Met Office seamless
CMA GRAPES
ECCC GEM
ECMWF IFS 0.25° ensemble
ECMWF AIFS 0.25° single
ECMWF IFS 0.25°
NOAA GFS
DWD ICON
JMA seamless
Météo-France seamless
Met Office seamless
CMA GRAPES
ECCC GEM
ECMWF IFS 0.25° ensemble
paired absolute-error difference — mean over shared cases of ( |f_row − y| − |f_col − y| ). Negative means the row model was closer, in °C.
row closercolumn closer– means no shared measured cases
ECMWF AIFS 0.25° single vs ECMWF IFS 0.25°
ECMWF AIFS 0.25° single
ECMWF · forecast · ecmwf_aifs025_single
ECMWF IFS 0.25°
ECMWF · forecast · ecmwf_ifs025
On the 1,414 cases both answered, ECMWF AIFS 0.25° single was closer by 0.08°C on average (± 0.01 over 29 day blocks).
ECMWF AIFS 0.25° single closer on 48% of shared cases · error correlation 0.71 · +24h · sea temperature_2m
Does the advantage hold over time?
Mean of |error of ECMWF AIFS 0.25° single| − |error of ECMWF IFS 0.25°| on each day. Below the centre line, ECMWF AIFS 0.25° single was closer.
1,414 shared cases over 29 days · values in °C · shaded band = ±1 block standard error (0.01)
The dashed trend rises across the window — the advantage is changing, not fixed.
Do their errors move together?
Horizontal: signed error of ECMWF AIFS 0.25° single. Vertical: signed error of ECMWF IFS 0.25°. Points along the rising diagonal mean the two miss in the same direction at the same time.
1,414 shared cases (472 plotted, evenly sampled) · r = 0.71 · errors in °C
Correlation is association, not joint failure. Two strongly correlated models can both be accurate; two uncorrelated models can both be wrong.
Does it depend on how far ahead you ask?
The same paired difference at every horizon in the cohort. Left of centre, ECMWF AIFS 0.25° single was closer; right of centre, ECMWF IFS 0.25°.
+6h
−0.03 ±0.02 · n750
+24h
−0.08 ±0.01 · n1,414
+48h
−0.10 ±0.01 · n1,040
+72h
−0.07 ±0.02 · n1,180
Mean absolute error at each horizon: +6h 0.36/0.39 · +24h 0.33/0.42 · +48h 0.34/0.44 · +72h 0.41/0.48 (ECMWF AIFS 0.25° single / ECMWF IFS 0.25°).
each bar carries its own shared-case count · values in °C
One real run of cases
Site south-china-sea: what actually happened, and what each model said would happen.
observed ECMWF AIFS 0.25° single ECMWF IFS 0.25°
72 consecutive shared cases · 2026-09-08 14:00 → 2026-09-11 23:00 UTC · values in °C
This is one site's run. It shows what the aggregate looks like; it does not substantiate it.
These illustrate the aggregate; they do not substantiate it.
matrix-v3 · content e7e47751cba27646 · matched on task definition (domain + channel) · native unit · entity / site · target time · horizon bucket (±1h) · outcome vintage used · issue and revision policy: the receipt as stored, never re-issued
This selection is in the address bar — the link you copy opens exactly this comparison.
How to read this
Each cell rests on its own set of shared cases, so the matrix cannot be read as a global ranking.
Only cases both models answered are scored. Abstentions and non-answers are counted beside every model, because scoring answered cases alone favours selective models.
Uncertainty is a block standard error over shared days, not an assumption of independent cases.
A single illustrated case shows what the aggregate looks like; it does not substantiate it.
Coverage behind each model
ECMWF AIFS 0.25° singleecmwf_aifs025_singleecmwfanswered 1,463 · abstained 0 · no answer 3 · mae 0.34°C
ECMWF IFS 0.25°ecmwf_ifs025ecmwfanswered 1,463 · abstained 0 · no answer 3 · mae 0.41°C
NOAA GFSgfs_seamlessgfsanswered 1,463 · abstained 0 · no answer 3 · mae 0.55°C
DWD ICONicon_seamlessiconanswered 1,463 · abstained 0 · no answer 3 · mae 0.44°C
JMA seamlessjma_seamlessjmaanswered 1,463 · abstained 0 · no answer 3 · mae 0.52°C
Météo-France seamlessmeteofrance_seamlessmeteofranceanswered 1,463 · abstained 0 · no answer 3 · mae 0.54°C
Met Office seamlessukmo_seamlessukmoanswered 1,463 · abstained 0 · no answer 3 · mae 0.54°C
CMA GRAPEScma_grapes_globalcmaanswered 773 · abstained 0 · no answer 0 · mae 0.54°C
ECCC GEMgem_seamlessgemanswered 773 · abstained 0 · no answer 0 · mae 0.49°C
ECMWF IFS 0.25° ensembleecmwf_ifs025_ensecmwfanswered 737 · abstained 0 · no answer 5 · mae 0.38°C
A second body of forecasting evidence, read from a cached notflat.world capture. It is shown here for reference and coverage: it is not mixed into the matrix above, because a comparison is only meaningful when task, unit, target, horizon and outcome vintage genuinely match.