Europe
The Copernicus multi-system, SEAS5 and the NMME over the same continent, each scored against its own reforecast — and the awkward truth about forecasting a European winter.
What the systems say
1 October 2026 · routes 10 · 03 · 02Temperature
| Target | lead | SEAS5 | C3S multi-system | NMME | |||
|---|---|---|---|---|---|---|---|
| P(above) | skill | P(above) | skill | P(above) | skill | ||
| October 2026 part elapsed | +0 | 35% | 77% | 36% | 81% | 38% | 82% |
| November 2026 | +1 | 36% | 38% | 38% | 70% | 40% | 68% |
| December 2026 | +2 | 49% | 46% | 44% | 74% | 32% | 42% |
| January 2027 | +3 | 47% | 29% | 45% | 44% | 46% | 19% |
| February 2027 | +4 | 46% | 34% | 43% | 82% | 40% | 28% |
| March 2027 | +5 | 39% | 47% | 37% | 68% | 41% | 62% |
They are furthest apart on December 2026: SEAS5 puts the odds of an above-normal month at 49% where NMME puts them at 32% — 17 points between two systems reading the same ground for the same month. The skill column is the one that governs how much of this to believe: 38–70% of this box beat climatology at November 2026, and 47–68% by March 2027.
Rainfall
| Target | lead | SEAS5 | C3S multi-system | NMME | |||
|---|---|---|---|---|---|---|---|
| P(above) | skill | P(above) | skill | P(above) | skill | ||
| October 2026 part elapsed | +0 | 37% | 64% | 39% | 86% | 36% | 67% |
| November 2026 | +1 | 37% | 29% | 38% | 47% | 34% | 35% |
| December 2026 | +2 | 36% | 26% | 39% | 40% | 37% | 29% |
| January 2027 | +3 | 40% | 32% | 39% | 64% | 38% | 43% |
| February 2027 | +4 | 46% | 27% | 42% | 59% | 38% | 38% |
| March 2027 | +5 | 37% | 38% | 41% | 46% | 40% | 42% |
They are furthest apart on February 2027: SEAS5 puts the odds of an above-normal month at 46% where NMME puts them at 38% — 8 points between two systems reading the same ground for the same month. The skill column is the one that governs how much of this to believe: 29–47% of this box beat climatology at November 2026, and 38–46% by March 2027.
And what this site’s own engine says
Route 09skilful over 45% of the box
skilful over 15% of the box
skilful over 4% of the box
skilful over 2% of the box
skilful over 6% of the box
The Analog Bench does not appear in the table above, and cannot: it forecasts rolling three-month seasons where the others forecast months, and it publishes an anomaly in degrees where they publish a tercile probability. Holding those side by side in one row would be arithmetic on unlike quantities.
It is here because it carries one number nothing else on this site does. Signal is the anomaly with the warming trend taken back out — what the analogue years actually contribute beyond what you would get by drawing a straight line through the last forty. Where the two are close, the forecast is mostly the trend and deserves to be read as such. The analogues are drawn from the ocean state through July–August–September 2026.
The systems behind those numbers
Routes 10 · 03 · 02Multi-system
Copernicus C3S
Eight centres — ECMWF, the Met Office, Météo-France, DWD, CMCC, NCEP, ECCC and JMA — regridded onto one grid with one calendar and one hindcast period. Europe has the best-organised seasonal archive in the world, and this lane is built on it.
Reference
SEAS5, in detail
All 51 members, seven monthly leads, and the evolution of the outlook for one target month across successive initialisations — which is usually more informative than any single run of it.
Winter
The NAO problem
The European winter lives and dies on the North Atlantic Oscillation, and the models predict it with a signal that is real but far too quiet relative to their own spread. Any honest European winter product has to deal with this explicitly. Ours will.
The signal-to-noise paradox
Why European winters are strangeTake a large seasonal ensemble, average it, and correlate that ensemble mean against the observed winter NAO. You get a respectable correlation — the models know something. Now correlate one member of the ensemble against the rest of the ensemble. You get a lower correlation.
That is backwards. A model should predict itself better than it predicts reality. The standard reading is that these models produce a predictable signal that is too weak relative to their own internally generated noise — so the ensemble mean, taken at face value, systematically under-states the forecast anomaly, and has to be scaled up before it can be believed.
The practical consequence for a reader: a European winter outlook showing a mild-looking +0.3 °C ensemble-mean anomaly may be a much stronger statement than it looks, and a site that shows you the raw ensemble mean without saying so is showing you a misleadingly timid map.