About this product
The ensemble is the same model run fifty-one times. ECMWF perturbs the analysis and the model physics and integrates fifty-one forecasts at once — one unperturbed control and fifty perturbations — because the atmosphere's sensitivity to its own initial state is not a nuisance to be engineered away but the thing a forecast beyond about day four is mostly about. What comes out is not one answer. It is a distribution, and the two numbers that summarise it are on this page.
The mean is the forecast. Averaging fifty-one members cancels the detail they disagree about and keeps the structure they share, and past roughly day four that is measurably a better forecast than any single run — including the high-resolution one. It is also, deliberately, a smoother field than HRES: a low that fifty-one members put in fifty-one slightly different places comes out broader and shallower than any of them drew it. That is not the mean losing the storm. It is the mean saying it does not know where the storm will be, and a chart that sharpened it back up would be inventing confidence nobody has.
The members, used properly. Fifty rasters of the same field is not a product; what the members are for is arithmetic across them, and that is what the field picker now offers, quantity by quantity: the mean, the spread, the 10th, 50th and 90th percentiles, and the probability of exceeding a threshold that matters — ten millimetres of rain, thirty degrees, a gust of twenty-five metres a second. The temperatures add a daily high and low, folded from the members' three-hourly extremes. All of it is computed on this site's own hardware from every member of every cycle, and written out as Zarr so it can be shared as data, not only as pictures.
Click a city and the point view draws the plume: all fifty members as lines, the 10th–90th band, the median, and the statistics at the hour on screen. The calendar lays the members out by day — each row a member, each column a day, each cell that member's daily high, or its rain — so a heat wave the ensemble agrees on and a weekend it does not read at a glance. Spaghetti in the map options draws each member's 552 and 576 dam contour over the mean: a tight bundle is agreement, a fan is doubt.
Uncertainty, on the mean. The map options can hatch the mean where the members disagree — sparse where the spread is in the lower half of its own table, dense in the upper quarter — and the pointer then reads the pair together: the mean, and ± the spread at the same point.
The spread is how much to believe it — and it is drawn here as a field in its own right, on its own scale, beside every mean. It is the standard deviation across the same fifty-one members: small where they agree and the mean is a real forecast, large where the mean is an average of genuinely different weather and should be read as a statement of ignorance rather than a prediction. Every field button has a spread twin next to it for exactly this reason. The pair is the product; either half alone is half the story.
Every field here is computed from the members themselves. ECMWF's open data carries the fifty perturbed members at 0.25°; they land on this site's own hardware four times a day, and the mean, the spread, the percentiles, the probabilities and the daily extremes are all arithmetic over those fifty, done in one pass and written out as GRIB2 and Zarr before a single frame is drawn. The fields are the surface set a forecaster reads — 2-metre temperature, its daily high and low, precipitation, snowfall, 10-metre wind and gust, mean sea-level pressure, CAPE, cloud and column water vapour — and the synoptic skeleton: geopotential height at 1000 and 500 hPa, temperature at 850 hPa, wind at 250 hPa.
The winds are speeds rather than components, which is why this page offers no streamlines where the other two do. Averaging a direction across fifty-one members is not a well-defined operation, so ECMWF ships the scalar and this page draws the scalar rather than reconstructing a direction that would be a fiction.
The grid is 0.25°, the run goes to fifteen days three-hourly to day six and six-hourly after it, and there are four cycles a day, of which 06Z and 18Z run to six days rather than fifteen. All of that is a property of the ensemble rather than a choice made here: fifty members at HRES's resolution and cadence would cost more than the forecast is worth, and the trade is the right way round. There is also no analysis frame, so the timeline starts at +3 h: the perturbations are zero at initialisation, so an ensemble mean at hour zero is just the analysis and a spread at hour zero is a field of zeros.
Compare & verify in the map options draws HRES or AIFS over the ensemble mean at the same valid time, or subtracts one from the other. Mean sea-level pressure is the one field all three publish, so it is the one that pairs — and an HRES-minus-EPS difference map is worth reading carefully: it shows one forecast against the average of fifty-one, so the largest differences are where the ensemble is least certain rather than where either model is wrong.
As on the other explorers, one raster covers the planet and every one of the 226 countries, twelve continental frames and eight ocean basins is a viewport on a field that is already drawn, so moving from Iceland to New Zealand is instant. The pointer reads a value back out of the picture and reports the band it falls in rather than a precision a banded chart never claimed. The polar caps are on the polar explorer, where the ensemble has the most to say: the vortex at day ten is a large-scale medium-range structure, which is exactly what a mean is good at and exactly where the spread is worth knowing.
Data © ECMWF. Forecast Euro is not affiliated with ECMWF.