One initialisation of ECMWF SEAS5, all fifty-one members, one frame a day, for seven months. Every other seasonal route on this site answers a question about a month — how warm December will run, how likely the upper tercile is, whether the polar vortex breaks. This one answers a smaller and more literal question: what does the model put on each individual day, and how many of its members agree?
That second half is the whole route. Route 14, which this replaces, read a published ensemble mean and spread — two numbers. Two numbers cannot be counted past a threshold: you cannot ask them how many members put a night below freezing, because the mean of fifty-one members is warmer than the cold ones and a spread carries no shape. Fifty-one members can be counted, and the chance layers on this page are that count and nothing more — no distribution fitted, no calibration applied, just the share of members on the far side of a line.
| Field | Units | Days | Layers | What it is |
|---|---|---|---|---|
| Daily mean temperature | °C | 215 | Mean · P10 · P90 · Spread | The average of the day across all four synoptic hours, member by member. The seasonal march dominates every one of these maps — that is what a seven-month run of daily frames mostly shows, and it is worth watching precisely because it is so much larger than anything the model is being clever about. |
| Daily maximum temperature | °C | 215 | Mean · P90 · Spread · P ≥30 °C · P ≥35 °C | The warmest point of each day, as the model itself reports it over the 24 hours ending at 00 UTC. The heat side of a run, and the field a summer is read in. |
| Daily minimum temperature | °C | 215 | Mean · P10 · Spread · P frost · P < −10 °C | The coldest point of each day. The field to watch through a winter: frost, hard freezes and cold outbreaks live here, and a monthly mean averages them away against the mild days either side. |
| Precipitation | mm | 215 | Mean · P90 · Spread · P ≥1 mm · P ≥10 mm | How much falls in the 24 hours ending at 00 UTC, member by member. The mean of 51 members is a smooth field and a real one; the individual members are far spottier, which is exactly what the chance layers beside it are for. |
| Snowfall | mm w.e. | 215 | Mean · P90 · Spread · P snow | Snowfall as water equivalent, which is what the model actually carries — roughly a tenth of it in depth, but the ratio is a function of temperature and this store does not compute it, so the honest field is the one the model publishes. |
| Mean sea level pressure | hPa | 215 | Mean · P10 · P90 · Spread | The daily mean of the four synoptic hours. Violet is low and stormy, amber is high and settled — the circulation that explains the other maps, so it is drawn in its own colours and never in theirs. The 10th percentile is the map of how deep the members are willing to go. |
| 500 hPa height | gpm | 215 | Mean · P10 · P90 · Spread | The mid-troposphere steering flow: ridges amber, troughs violet. Every surface field on this page is downstream of this one, and the spread layer here is the cleanest picture the site has of when the model stops knowing what the flow will be doing. |
| Soil temperature, 0–7 cm | °C | 215 | Mean · P10 · P90 · Spread | The top seven centimetres of ground, daily mean. It lags the air by days to weeks and keeps the memory of a cold spell long after the air has recovered — which is why the freeze line creeping across a continent looks so different here from the way it looks in the air-temperature map beside it. |
| Soil moisture, volumetric | m³/m³ | 215 | Mean · P10 · P90 · Spread | How much water the top soil layer is holding, as a volume fraction. The slowest field on this page and the one with the longest memory: it integrates months of rainfall and evaporation, and it is the land's own contribution to the next season's temperature. |
| 850 hPa temperature | °C | 215 | Mean · P10 · P90 · Spread | Temperature just above the boundary layer, where the surface's own day-to-night swing and its snow cover stop confusing the airmass. The field to read a cold outbreak's origin in rather than its arrival. |
| Layer | Threshold | What it counts |
|---|---|---|
| Chance of a 30 °C day | Max temp ≥ 30 °C | The share of the 51 members whose maximum reaches 30 °C on this day. An absolute threshold, not a percentile — so it carries the model's own warm or cool bias in the open rather than hiding it inside a climatology this store does not have. |
| Chance of a 35 °C day | Max temp ≥ 35 °C | The same count at 35 °C — the threshold at which heat stops being weather and starts being a health question. |
| Chance of frost | Min temp < 0 °C | The share of the 51 members that put this gridcell below freezing overnight. The single most useful thing 51 members can say that a mean and a spread cannot. |
| Chance below −10 °C | Min temp < -10 °C | A hard freeze rather than a touch of frost — the threshold that separates a cold night from a cold outbreak. |
| Chance of a wet day | Rain ≥ 1 mm | The share of members with at least a millimetre. This is the closest thing on the site to a daily rain probability, and it is an honest one: 51 members counted, nothing modelled on top. |
| Chance of 10 mm | Rain ≥ 10 mm | Ten millimetres in a day — a wet day rather than a damp one. |
| Chance of snowfall | Snow ≥ 1 mm w.e. | The share of members with at least a millimetre of water equivalent — about a centimetre of ordinary snow. |
Every threshold is absolute, and that is deliberate. Freezing, thirty degrees, a millimetre — physical lines, not percentiles of a climatology. This tree has no reforecast behind it, so there is no model climatology to take a percentile of; and even if there were, an absolute threshold is the more honest of the two here, because it carries the model’s own warm or cool bias in the open where “the coldest fifth of the reforecast” would hide it inside the baseline. Read a 60 % chance of frost as sixty per cent of SEAS5’s members, not as sixty per cent of reality.
The maximum and minimum are published by the model over the 24 hours ending at 00 UTC, and rainfall and snowfall accumulate over exactly that window. So every other field is made to match it: the mean temperature and the pressure are averaged from four synoptic hours inside that same window, and the two upper-air fields from the two the archive offers. One window, eight fields — a “day” that meant something different for temperature than it did for rain would be wrong in a way nothing on the page could show.
Why not simply take the 00 UTC value? Because on a global map a single synoptic hour is local midnight at Greenwich and local noon at the dateline, and the land diurnal range is ten to twenty degrees: the map would carry a permanent east–west band that looks like weather and is an artefact of when the snapshot was taken. Pressure has the same problem for a subtler reason — the atmospheric semidiurnal tide is over a hectopascal in the tropics, comparable to a real monsoon anomaly, and because that tide has a twelve hour period, averaging 00 and 12 UTC does not remove it. Only all four hours do.
The mean of 51 members is a smooth field and a real one, and it is also a map of no weather that will ever happen. Averaging 51 storm tracks gives a shallow smear where every member had a sharp low; averaging 51 rainfall fields gives drizzle everywhere and a downpour nowhere. That is what an ensemble mean is for and it is a poor picture of a day.
So the MEMBERS button leaves the ensemble behind and draws one run. Step through the days and you are watching one physically consistent sequence of weather — fronts arriving, a ridge building, a cold snap that lasts nine days. Step through the members on one day, with the up and down arrows, and you are watching the same Tuesday fifty-one times. That second motion is the honest picture of a seasonal forecast, and no summary statistic conveys it: the spread map says the members disagree, and flipping through them shows you what disagreeing looks like.
Read a member as a possible season, never as a forecast of a date. At three months no run is telling you about that Tuesday; the ensemble is telling you what the season can do, and each member is one draw from it. The members are not ranked and not ordered — member 07 is not more likely than member 44.
SEAS5 is integrated at one degree — roughly 100 km, a cell that averages an alpine valley together with the summit above it. Nothing here changes that. What the maps add is terrain: a static, observed correction from ERA5’s 0.25° 1991–2020 climatology, which knows where the mountains and the coastlines are even though the model’s grid does not.
The correction is the difference between the real climatology and a one-degree smearing of that same climatology, so it averages to zero over the model’s own cell — coarsen a downscaled map back to 1° and the model’s value comes back, to within 0.11 °C, half a per cent of the field’s own spatial variation. It carries no anomaly, no trend and no skill. If a member is three degrees above the model’s climate, the drawn map is three degrees above the observed climate, spread over the ground the way that month actually spreads it.
The size of what this fixes: at 1° the model puts Zermatt at 4.6 °C on a mid-September day, because its cell is half valley floor. Drawn at 0.25° the same member reads 1.0 °C. Kathmandu moves 2.8 °C cooler, Vancouver 1.0 °C milder. Over flat ground and open ocean it moves almost nothing — Kansas by 0.15 °C, the mid-Pacific by 0.03 °C. Over most of the world this is cosmetic; in the places people live in mountains and on coasts it is the difference between a usable number and a wrong one.
It will not sharpen a storm. The weather in a member is still 1° weather: fronts are as broad and lows as shallow as the model drew them. What moves is where the ground is high, low, wet or by the sea.
Pressure and 500 hPa height are deliberately not downscaled. Measured against their own 0.25° climatology they have nothing to restore — sea-level pressure has already been reduced to sea level, so its terrain signal has been removed by construction (0.002 hPa mean, 0.08 at worst), and 500 hPa is free atmosphere with no sub-degree structure at all (0.2 % of its colour range). Upsampling either would be a smoothing dressed as detail. The daily maximum and minimum borrow the mean temperature’s correction, because the archive holds no 0.25° climatology of the extremes and the correction is dominated by elevation, which acts on the whole day — an approximation, said here rather than hidden.
Clicking the map snaps to the nearest of 537 published places and fans the members across the whole run: the 10th to 90th percentile as the outer band, the 25th to 75th inside it, the median through the middle. It is the same distribution the map is drawing, read down a column instead of across a surface — and it is the clearest picture on the site of the thing a seasonal forecast actually does, which is to stop resolving individual days after about three weeks and keep only the seasonal signal underneath.
No skill score, and not yet a way to get one. The store behind this page is forecast only — the reforecast that would let it be scored is twenty-four years by twenty-five members, about twelve times the download, and it has not been taken. Every forecast on this site ships with its measured skill or says plainly that it has none; this one says it here rather than in a footnote. For SEAS5 with its skill attached, route 10 and route 06 are where it lives.
No anomalies and no terciles, for the same missing reason: a departure from normal needs a daily model climatology, and that climatology is the reforecast. An absolute map is honest about what it is; a departure from a borrowed baseline would not be.
No run-to-run comparison. Route 01 compares two initialisations holding one target month. On a daily axis the same date is a different lead in every run, so the comparison would be between two different questions. The run picker still moves between initialisations; it just does not overlay them.
One degree, not a quarter. The members are published at 1° and the rasters are drawn at two pixels per degree — headroom for the Mercator stretch, and not one pixel more than the data has. A finer canvas would imply detail that is not there.