The same target month, as called by every initialisation in the store. Where the lines sit on top of one another the model has been saying the same thing for months; where the newest one jumps, it has changed its mind, and that is worth more than any single run.
ECMWF SEAS5 is the reference seasonal forecast system: 51 members, run once a month, out to seven months. What this page draws is the ensemble mean of those members on a 0.25° grid — the middle of the distribution, not its spread. For the spread, and for the probability that a month lands in the upper or lower third of its climate, see the Probability Plate, which uses all 51.
The timeline under the map carries the whole run — OCT 2026 to APR 2027 — one frame a month. Lead 1 is the initialisation’s own month, which is already part-elapsed by the time anyone reads it, so the page opens on the first whole month and marks any frame that has already happened with a hatched rail down the left of the map. That is not a technicality: a “forecast” for a month half of which is in the past is a different object from a forecast for next March, and presenting the two identically would be the most misleading thing this page could do.
Every anomaly here is measured against the SEAS5 system's own 1993–2016 hindcast climatology — the model re-run over past years, not an observed normal. That is deliberate and it matters: a model-relative departure cancels the system’s own systematic drift, so it is a like-for-like statement about this forecast rather than a mixture of forecast error and model bias. It also means a SEAS5 anomaly is not an ERA5 anomaly and the two should never be quoted against each other. See the systems page.
Months or seasons. The toggle beside the transport switches the axis from seven one-month leads to the five rolling three-month means a seasonal outlook is actually quoted in — ASO, SON, OND, NDJ, DJF. Three months is roughly where a monthly signal stops being noise, and the windows overlap on purpose, because watching one develop is most of what the axis is for. Temperature, pressure and the rest are averaged over the window; rainfall and snowfall are totalled, and the legend changes unit to say so — a three-month total on a scale built for monthly ones would read as an extreme that is really just a longer window.
The distribution, not just the middle. For 9 of these fields the page can draw the 51 members themselves rather than their mean: the 10th and 90th percentiles — the cold or dry and the warm or wet edges of the run — the median, and the spread between them, which is the map that says how much of the rest to believe. Click any labelled place and the panel under the map draws a box and whisker for every month or season of the run, with the model’s own 1993–2016 hindcast median beside it: a 2 °C box is wide in the tropics and narrow in Siberia, and a spread means nothing until it is set against something on the same footing.
Two things about the members are worth stating plainly. They come from the Copernicus archive at 1.0°, not the 0.25° of the mean fields, and the maps are published at that resolution rather than upsampled to match — a smoother picture that implied detail it did not have would be worse than an honest coarse one. And they reach 6 leads where the mean reaches 7: C3S stops one month short for this system. The page reports that where it happens rather than drawing a blank frame. 4 initialisations carry members, so the compare control will hold one run’s percentiles against another’s at the same target month. The rest of the archive is mean-only, because each initialisation needs its own re-forecast.
Colour tables are pinned to absolute values, never re-fitted per frame, so stepping through the months shows the forecast changing rather than the palette changing. The value under the pointer is read back out of the drawn pixel through the palette the frame was written with, so the readout cannot disagree with the picture; it is precise to about half a colour step and no further.