Explainer
Ensemble forecasts, explained
Last updated: 16 November 2026
A single model run gives one answer and no hint of how far to trust it. An ensemble runs the forecast many times with small differences and looks at how much the answers vary. It is the main tool forecasters use to turn “the model says” into “the chance is”. This guide explains how ensembles work and how to read them.
The main global ensembles
| Ensemble | Run by | Members |
|---|---|---|
| ECMWF ENS | ECMWF | 51 |
| GEFS | NOAA (USA) | 31 |
| ICON-EPS | DWD (Germany) | 40 |
| GEPS | ECCC (Canada) | 21 |
| MOGREPS-G | Met Office (UK) | 18 per run |
Why one forecast is not enough
The atmosphere is a chaotic system. In the early 1960s the meteorologist Edward Lorenz found that two simulations started from almost identical conditions drift apart until they have nothing in common. The same thing happens in real forecasting, because the starting state is never known exactly: observations have gaps and errors, and the model itself is an approximation.
So a single forecast is one possible future out of many, and nothing in it tells you whether it is a robust one. Sometimes the atmosphere is in a stable pattern and small errors hardly matter. Sometimes it is balanced on a knife-edge. From one run, you cannot tell which.
How an ensemble works
An ensemble runs the same model many times. One run, the control, starts from the best estimate of current conditions. The others, the perturbed members, start from slightly altered versions of it, each change small enough to be consistent with the uncertainty in the observations. Many systems also vary the model’s physics slightly from member to member.
ECMWF runs 51 members, NOAA’s GEFS runs 31, and Germany’s ICON ensemble runs 40. Each member is a complete, physically consistent forecast. What matters is the collection.
Historically, ensemble members ran at lower resolution than the main forecast to save computing time. That gap has closed at ECMWF, where all 51 members now run at about 9 km.
Reading the mean and the spread
The ensemble mean is the average of all members. It is smoother than any single run because the unpredictable details cancel out, and beyond a few days it is, on average, more accurate than the control. The cost is that it blurs extremes: the mean of twenty sharp storms in twenty slightly different places is one broad, weak storm that no member actually predicts.
The spread is how much the members differ, and it is the most valuable thing an ensemble gives you. Small spread means the forecast is not sensitive to the starting errors, so confidence is high. Large spread means the outcome is undecided.
Counting members gives a probability. If 40 of 51 members bring rain to your location on Saturday, the chance of rain is roughly 80 percent. Most precipitation probabilities in weather apps are produced this way.
Common ensemble charts
A few chart types turn up again and again, and each answers a different question.
- Plume or meteogram: every member’s forecast for one place plotted as a line over time. The lines start together and fan out; where they fan out is where the forecast stops being reliable.
- Spaghetti plot: one contour or one storm track drawn for every member on the same map. Tight bundles mean agreement, tangles mean uncertainty.
- Probability map: the share of members exceeding a threshold, such as 25 mm of rain or gale-force wind.
- Box-and-whisker meteogram: the median, the middle half and the extremes of the members at each time step, in one compact symbol.
Ensembles versus comparing different models
An ensemble varies the starting conditions of one model. Comparing different models — ECMWF against the GFS against ICON — varies everything: the starting analysis, the grid, the physics, the institution. The two approaches answer related but different questions.
A single-model ensemble shares that model’s systematic biases. If the model is too slow with a certain type of storm, all its members tend to be too slow, and the ensemble can be confidently wrong. A multi-model comparison is sometimes called a poor man’s ensemble: it has far fewer members, but they are independent, so their agreement is strong evidence.
MeteoModels takes the second approach. The certainty mask measures the spread across the separate models you select, not across the members of one ensemble. It is not a calibrated probability, but it is a fast and transparent answer to the practical question: do the world’s models agree here or not?
Common mistakes when reading ensembles
Three errors are worth avoiding.
- Treating the mean as a forecast of an event. The mean is a best guess for an average value, not a realistic map of a storm.
- Picking one dramatic member. A single member with a blizzard is a possibility worth a few percent, not a prediction.
- Assuming small spread means certainty. Ensembles tend to be somewhat overconfident, especially for surface weather, and all members share the same model’s blind spots.
Frequently asked questions
What is an ensemble forecast?
An ensemble forecast is a set of forecasts from the same model, each started from slightly different initial conditions. The differences between them show how uncertain the forecast is.
What is the ensemble mean?
The ensemble mean is the average of all members. It is usually more accurate than a single run beyond a few days, but it smooths out extremes and should not be read as a realistic picture of an individual storm.
How many members does the ECMWF ensemble have?
The ECMWF ensemble has 51 members: one control run and 50 perturbed members, all at about 9 km resolution.
What is the difference between an ensemble and a deterministic forecast?
A deterministic forecast is a single run that gives one answer. An ensemble is many runs that give a range of answers and therefore an estimate of probability and confidence.
Is a multi-model comparison the same as an ensemble?
No. An ensemble varies the starting conditions of one model. A multi-model comparison uses separate models built by different centres. The models are fewer but independent, so their agreement is a strong sign of confidence.