Interstate 90 · Seattle to Boston · 3,020 milesThe season ahead,
The season ahead,
read three ways.
Seasonal forecasting is the long haul: one month out to seven, where the weather has stopped being weather and become odds. Forecast Season takes the systems that do it — ECMWF's SEAS5, the North American Multi-Model Ensemble, and the Copernicus multi-system — and reads them for the United States, for Europe, and for the world.
Every engine on the board is running. The SEAS5 Explorer puts ECMWF's seasonal system on an interactive map — sixteen fields worldwide, seven monthly leads, and one initialisation held up against another — with the run-to-run board underneath it that shows how much the model has changed its mind. The model garage, the drivers and the methods are open alongside it.
Three lanes
United States · Europe · GlobalUnited States
U.S. route
CPC's monthly and seasonal outlooks, the NMME behind them, and the drought outlook — read against the models rather than instead of them.
Take the exit→
Europe
European route
The C3S multi-system and SEAS5 over the continent, and the honest story about winter: a low signal-to-noise ratio, and what you can still say through it.
Take the exit→
Global
World route
Where seasonal forecasting actually works — the tropics, the monsoons, and the ENSO teleconnections that reach out of them into everyone else's winter.
Take the exit→
The route board
What is open, what is being pavedEvery product on this site is a numbered route. A route is only ever signed open once it is genuinely here — there are no placeholder maps and no invented numbers anywhere on this site.
Route 01Open
SEAS5 Explorer
ECMWF's seasonal system, seven months out, on a map you can drag, zoom, re-colour and read values out of — sixteen fields worldwide at 0.25°, every monthly lead, six initialisations to compare against each other, and the run-to-run board that shows how the outlook for one target month has changed run by run. There is a *[locked, region-by-region viewer](/seas5/legacy/)* beside it for reading the same data the way a chart desk does.
ECMWF · 0.25° global · 7 leadsOpen →
Route 02Open
The NMME
The North American Multi-Model Ensemble, pooled the way it is meant to be read — as a multi-model probability rather than any one centre's best guess — and counted honestly: of the eight models that fly under the banner, the ones with a reforecast behind them.
One vote per model · calibrated · scoredOpen →
Route 03Open
Copernicus Multi-System
The C3S seasonal multi-system: ten systems from nine centres on one grid and one calendar, each calibrated against its own reforecast. The place to see whether an outlook is a signal or one model's opinion — and whether agreeing means being right.
10 systems · 704 members · one vote eachOpen →
Route 04Open
The Model Garage
Every operational long-range system, what it actually is under the hood, how many members it carries, how far it runs and when it publishes. The reference you want open in the next tab while you read an outlook.
ReferenceOpen →
Route 05Open
Drivers
ENSO, the Indian Ocean Dipole, the NAO and the polar vortex, the QBO, the MJO, and the slow ocean modes. What each one is, how it is measured, and what it does — and does not do — to a season.
ReferenceOpen →
Route 06Open
Verification & Skill
The uncomfortable half, and the only route that scores the others. Every scorecard the site publishes in one table — RPSS by lead and by region for all three systems — plus the three questions no single route can ask about itself: which one actually verifies best, whether pooling models pays, and which multi-model to believe where they disagree.
RPSS · land-masked · three pipelinesOpen →
Route 07Open
The Season Ahead
City briefings: the full 51-member distribution for one place, month by month, with the tercile probabilities against the model's own climate — and, beside each one, whether a forecast for that gridcell has ever actually verified.
40 cities · 51 members · skill attachedOpen →
Route 11Open
The Second Opinion
Two of this site's routes forecast the same seasons and disagree over a quarter of the world. This one scores them against each other on identical footing and says which to believe where — including where the answer is not the flattering one.
Matched verification · 24 yearsOpen →
Route 10Open
The Probability Plate
All 51 SEAS5 members rather than their mean: tercile and decile probabilities in raw and trend-adjusted form, the ensemble median, and the reforecast skill that says where issuing a probability beats saying nothing at all.
51 members · 24-year hindcast · RPSSOpen →
Route 09Open
The Analog Bench
The site's own forecast, and the only route that does not render somebody else's model: the years whose ocean state most resembles this one, what actually followed them, and the cross-validated skill that says where to believe it and where not.
86 years · constructed analog · skill-maskedOpen →
Route 08Open
The Wire
Every outlook the site publishes, in two files meant for something other than a person: one plain-text dump for a broadcast desk or a grep, one JSON feed for a model or an agent, and an index to every route's own data. The same format the rest of the Forecast Network carries.
JSON · text · ASCII · indexedOpen →
Route 12Open
Weather Regimes
The seven recurring shapes of the Euro-Atlantic circulation, and how SEAS5's fifty-one members distribute the season across them — Greenland blocking, zonal flow, the Scandinavian ridge — with the reforecast skill printed beside every number, which is honest enough to be near zero in the shoulder season.
Grams 7 · 51 members · drift-corrected · scoredOpen →
Route 13Open
The Polar Vortex
Twenty kilometres above route 12, the ring of wind that decides how much Arctic cold the mid-latitudes actually get. The probability of a *major sudden stratospheric warming* from all 51 members, the wave driving that would cause one, the QBO phase it would happen under — and the reforecast skill, which is the honest reason to read the climatological rate printed beside every probability.
Charlton-Polvani · 51 members · 24 winters scoredOpen →
Route 15Open
The SFS Beta
NOAA's next seasonal system, published as a beta while it is built. It runs **twelve months** where the rest of this site runs six or seven, and carries a 1991–2025 reforecast — the most recent baseline here by nine years. That is what lets this route put a measured skill number against every one of the twelve leads, and so show where a seasonal forecast stops being a forecast rather than settling it by truncating the axis.
12 leads · beta1 + beta2 · 35-year reforecastOpen →
Route 16Open
The Daily Seasonal Forecast
Route 14's axis with route 10's evidence: SEAS5 **day by day for seven months, all 51 members**, eight fields from the Copernicus archive rather than a published mean. That is what lets it draw the thing a mean and a spread cannot — the **chance** of frost, of a 30-degree day, of ten millimetres of rain, counted member by member against an absolute threshold. Click anywhere for the plume. Still no skill score: there is no reforecast behind it, and the page says so on its face.
215 days · 51 members · 35 layers · no skill scoreOpen →
Route 17Open
The Storm Track
The only route here that publishes **objects** rather than a field. Every one of SEAS5's 51 members is run through *bombcyclone.earth*'s cyclone tracker, and what is counted is storms: where they pass, where they form, and where they **explosively deepen**, month by month with the inter-member spread beside every number. The axis is months on purpose — at seasonal lead one member's low on a given day is a draw from a distribution, not a forecast. Absolute counts only: no reforecast, so no anomaly and no skill score.
51 members tracked · 3 densities · no skill scoreOpen →
Route 18Open
The Subseasonal Explorer
The gap between the weather and the season: ECMWF's **extended-range ensemble**, run every day to 46 days, as **weekly and fortnightly means** on a map you can drag and read values out of. Temperature, rainfall, pressure and SST anomalies, the **actual rainfall total**, and the **spread** of the members beside every field; every day's run in the picker, so the forecast for one calendar week can be watched as it firms up. Anomalies are against ERA5's 1991–2020 WMO normal, and the page says what that carries.
EC46 · 6 weeks · daily runs · mean + spreadOpen →
Route 19Open
Degree Days
**Heating and cooling demand** from every one of SEAS5's 51 members: each member's daily high and low becomes a daily degree day, summed by calendar month, then **population- or area-weighted** over the US, its nine census divisions, Canada and Europe's largest markets, plus maps and thirty cities. US °F-days against 65 °F, Eurostat °C-days for Europe. **Corrected** against AgERA5 through the monthly reforecast, and **last winter's runs are scored** raw and corrected.
51 members · pop-weighted · corrected · last winter checkedOpen →
Route 21Open
Winter Folklore, Tested
Siberian snow, the QBO, the Blob, a cold October: **ten October indicators** people quote for the coming winter, scored against **76 winters of US gas-weighted heating degree days**, as correlations and as real forecasts made only from earlier winters. Ten fail. **El Niño strength** is the one signal that survives.
10 indicators · 76 winters · as-of backtestOpen →
Next departures
Published issue schedulesThe major systems run to a timetable. These are their routine publication times — schedules, not forecasts.
T-20DAYS
ECMWF SEAS5
Thu 5 Nov 12:00 UTC
5th of the month, 12 UTC
T-20DAYS
Copernicus C3S multi-system
Thu 5 Nov 12:00 UTC
5th of the month, 12 UTC
T-23DAYS
NMME
Sun 8 Nov 12:00 UTC
around the 8th
T-35DAYS
NOAA CPC outlooks
Thu 19 Nov 20:00 UTC
third Thursday
How to read a seasonal forecast
Four things worth knowing01
Ensembles, not forecasts
A seasonal model is run many times from slightly different starting points. No single member is the forecast; the spread between them is the forecast. Read the distribution — where it sits, and how wide it is.
02
Anomalies against the model's own past
A coupled model drifts. The fix is to re-run it over decades of past years — the hindcast — and express the live run as a departure from its own climatology. That is why a SEAS5 anomaly is not an ERA5 anomaly, and why the two should never be quietly swapped.
03
Terciles and the odds
Seasonal outlooks are framed as probabilities of the coming season landing in the coldest, middle or warmest third of the record. A 40 % chance of the warm tercile is a real, useful shift from the 33 % you would say knowing nothing — and it is still more likely than not that something else happens.
04
Skill has a geography
Seasonal forecasts are good in the tropics, useful in ENSO-teleconnected regions in winter, and close to worthless for mid-latitude summer rainfall. A site that shows you a map without telling you where the map is any good is selling you something.
The one rule this site is built on. A seasonal forecast is a statement about the odds over three months, not about a day. If a page here ever shows you a number that looks like a weather forecast, it is doing something wrong — and if it shows you a map without telling you what the skill is under it, so is that.