The Copernicus Multi-System

A map showing ten models in agreement is the most persuasive thing this site could publish, and one of the easiest ways it could mislead you. These systems are not independent witnesses. They share ancestry, they share resolution limits, several share components outright, and they share the same well-known biases in how the tropical ocean drives the atmosphere. When they agree, they may simply be agreeing on an error — and a consensus of ten wrong answers looks exactly like a consensus of ten right ones.

That is why the agreement layer is stippled on skill rather than left clean, why the scoreboard below is on the same page rather than a link away, and why the panel worth reading first is not the consensus but the gain — what the other nine centres actually add to ECMWF alone, measured, and drawn whichever way it comes out.

Temperature

Rainfall

"Systems agreeing" counts how many put their highest probability on the same tercile as the published consensus, averaged over land. The floor is not zero but roughly 3 — with three categories and no signal at all, a third of the systems land on the same one by chance.

Every system scored inside one leave-one-out, over land, on identical footing: for each held-out year each system's tercile boundaries are rebuilt from its own other 23 reforecast years before the ten are averaged. Calibrating once on all 24 and holding out only at the averaging step would score each model partly against its own answer, and the consensus would inherit ten times that flattery rather than escape it. The best single system in each row is marked; the last two columns say by how much the consensus beat it and how many systems beat the consensus.

Temperature

Rainfall

The rainfall result is the strongest case this route makes. Taken one at a time, seasonal rainfall terciles are close to worthless: at every full lead between 7 and 10 of the 10 systems score negative RPSS over land — issuing their probabilities is worse than admitting you do not know — and in September 2026 that is 10 of 10, with ecmwf-51 itself at -0.0109. Averaged, the same ten systems are positive at 5 of 5 leads. Route 10 concluded that one model's rainfall terciles were mostly not worth issuing, and that stands. This is the amendment: as a consensus they cross into being worth issuing, barely — which is a smaller claim than it sounds and a larger one than nothing.

Twenty-four years is not many, and a per-system RPSS from 24 cases is noisy — the ranking between two close systems is not meaningful, though the gap between the consensus and the field usually is. The benchmark column uses ecmwf-51, the system route 10 publishes on its own.

Membership is not a fixed club. CDS is asked which systems carry each initialisation month and the answer changes through the year — the Met Office moved from system 604 to 605 to 610 across 2026, and JMA has no system at all for some months. What follows is who was actually in this run.

Member counts range from 20 to 155, which is why each system gets one vote rather than one vote per member: pooling members would let the largest ensembles outvote the rest and produce a map mostly of who runs the most members. Note also that a system's reforecast is usually much smaller than its live ensemble — the probabilities are as fine-grained as the live members allow, but the skill estimate is only as good as the reforecast behind it.