The models

Four, in order of how close they get to the ground. COSMO is the reason this site exists; the other three are what you have after it runs out at two and a half days, and carrying three of them rather than one is what makes disagreement visible.

Key	Model	Grid km	Range	Runs	Latest cycle	Source and licence
Loading…

Why these four and not others

IMGW COSMO 2.8 km

Poland's national mesoscale model, run four times a day to T+60. It is the only model here that resolves the Kłodzko valley, the Tatra ridge line or the sea breeze on the Hel peninsula, and it is published as GRIB under the EU's high-value-dataset rules — free, keyless and reusable. The site takes hourly to T+48 and three-hourly to T+60, which is about 9.7 GB a cycle, because IMGW's endpoint supports no byte ranges and a step must be streamed whole.

DWD ICON-EU 7 km

The natural continuation: the same family of physics on a coarser grid, out to five days, from a service that has published openly for a decade. It picks up exactly where COSMO stops, and it is the only model here besides COSMO that carries a snow-fall limit and a cloud ceiling.

Met Office Global 10 km

Carried because it is a genuinely independent forecasting system — its own data assimilation, its own dynamical core, its own physics — rather than another derivative of the same European chain. When it and IFS agree about day six, that agreement means something.

ECMWF IFS 0.25°

The highest-scoring global model in the world, and coarse over Poland: a quarter of a degree is about 17 by 28 km here, so the Carpathians are three grid cells. It is still the one worth reading on day eight, which is the only reason it is here.

NCEP GFS is deliberately absent. At 0.25° it would be a fourth global model at the same resolution as IFS with worse scores, over a domain small enough that another coarse global is noise rather than signal.

Against the thermometers

Loading…

Model	Field	n	MAE	Bias	RMSE	Units	Worst two stations
Loading…

What is being compared. The values come out of the published label sidecars — the same files the map labels read — not out of the model in memory. So this checks the whole site: if the grid geometry, the city sampler or the label writer were wrong, this table would show it. It is a test of Forecast Poland at least as much as of COSMO.

Fair, not flattering. Every model is asked about the same valid time, so a model whose cycle is older is answering a harder question and the lead time in the first column says so. The observations are never in the forecast: COSMO's 00Z run assimilated data up to 00Z, and this compares its T+17 against the 17Z reports.

A model with no row had no frame valid at that hour — ECMWF's open data is three-hourly, so five hours in eight it simply has nothing to compare. Wind is compared as a speed only: the synoptic reports give direction in tens of degrees and the anemometer exposures across this network range from an oil platform to a 1991 m summit, so a direction statistic here would measure the network rather than the models.