The NMME
The multi-model behind CPC's outlooks, and the other half of what route 03 does for Europe. Two independent panels — different centres, different reforecast windows, different grids — asked the same question about the same months. Where they agree, that is worth more than either alone; where they do not, that is worth knowing too. And as on every page here, each model is calibrated against its own past runs and scored before anything is drawn.
scripts/route02.shIt needs ~4 GB of NMME raw, which
scripts/nmme_ingest.py pull will restore from the archive if it has been pruned.How to read this
Four things it is notCoverage
Not the whole NMME
This is the part of the ensemble that can be calibrated and scored: models with a reforecast long enough to build tercile boundaries and measure skill against. CPC's own realtime feed carries more models than this page does, and no reforecast at all — which would mean probabilities with nothing behind them and no way to say whether they have ever verified.
Independence
Fewer minds than models
Two of the models here are one institution's two configurations, sharing components outright. Treating the count of agreeing models as a sample size — as though three of four were three independent confirmations — overstates the evidence considerably at this ensemble size.
Grid
Not route 03's
The NMME arrives on whole degrees; the C3S archive on half-degree centres. This page therefore is not pixel-comparable with the multi-system without an explicit regrid. The observations are interpolated onto the models here rather than the models onto a shared grid, because you should never resample the forecast when you can resample the thing it is scored against.
Calibration
Each against itself
Every model's probabilities come from its own reforecast over the common 1991–2016 window — not a shared climatology, and not each model's own longest record. A coupled model's drift is its own; but scoring them over different periods and different sample sizes would not be a comparison.