How are weather forecasts made?

A wind forecast is not anyone's guess. It is calculated on a computer by a weather model, which takes thousands of observations as its starting data. This guide explains how a forecast is made, why the same model can show different figures in different weather services, and how all this helps you on the water.

From observations to forecast

Every forecast goes through the same chain:

1. Observations. Weather stations, buoys and ships measure wind, temperature and pressure. Weather balloons and aircraft measure the air higher up. Weather satellites and weather radars see clouds, moisture and rain.
2. Initial state. Observations are scattered and contain errors. So they are combined with the previous forecast into a single overall picture of the current state of the atmosphere. This is called data assimilation. According to a 2023 description, HARMONIE does this every three hours and waits about an hour and a quarter for observations (the observation cut-off) before the calculation begins (Eresmaa 2023).
3. Model. The atmosphere is divided into a grid, with boxes side by side and layers stacked on top of each other. Using the equations of physics, a computer calculates how wind, temperature and moisture change in every box, step by step, days ahead. One such calculation is called a run, and a run is named after its start time, for example "the 06 UTC run". UTC is universal time, the world's reference clock. In Finnish time (EET, UTC+2 in winter; EEST, UTC+3 in summer), the 06 UTC run therefore starts at 08:00 in winter.
4. Post-processing. The model output is turned into figures that suit people, such as the 10-metre wind and gusts. Some services correct the figures using observations and statistics, or combine several models. At the Finnish Meteorological Institute (FMI), meteorologists also edit the forecast by hand (FMI 2024).
5. Service. The data is passed to weather services such as Yr, Windy or VamoSpot, which draw it as a map or a table.

On the coast, grid size makes the difference

For each grid box, the model knows only one value: the average wind, the average terrain, and whether the box is land or water. The size of the grid boxes is also called the grid spacing.

• With 2.5 km grid boxes (HARMONIE), the coastline of the Gulf of Finland, large islands and bays can be told apart. The model can calculate how the land warms up during the day and how the sea breeze develops along the coast.
• With 9 km grid boxes (ECMWF), a single box can contain a headland, a bay and islands all at once. To the model, the coast is a blurred boundary, and the sea breeze comes out weaker or in the wrong place.

In an Australian comparison, higher-resolution models depicted the sea-breeze front in more detail than coarse ones (Brown et al. 2026). According to FMI, MEPS is generally better than ECMWF for wind and gusts (FMI 2024). Even so, neither model can see an individual headland, jetty or forest edge. Those only show up on site and on a wind gauge.

There is more about the sea breeze in the wind forecast guide.

Models used in Finland

HARMONIE is the name of the model, and MEPS is the name of the system the Nordic and Baltic weather services use to run the model together. FMI's open data and VamoSpot use the name HARMONIE; MET Norway and Yr use MEPS. It is the same model. The model domain covers Scandinavia, Finland and the Baltic states (Eresmaa et al. 2026).

European weather services commonly use ECMWF for forecasts more than two days ahead. FMI's own forecasts are based on it from two days ahead onwards at the latest (FMI 2024). ECMWF also runs an AI model, AIFS, which is covered in more detail in the wind forecast guide.

VamoSpot shows HARMONIE for the next few days and ECMWF up to ten days ahead. VamoSpot's ECMWF map is coarser than ECMWF's own model, because the data is fetched at half-degree intervals so that the page loads quickly.

Ensemble forecasts: many calculations of the same weather

The current state of the atmosphere is never known exactly. Even a small error in the initial state can grow so large within a day that the sea breeze arrives hours late or not at all. That is why forecasting centres calculate the same forecast many times. This is called an ensemble forecast, and each calculation is an ensemble member.

The members differ from one another because the starting data of each one has been deliberately changed slightly. These changes are called perturbations. In MEPS, the following are perturbed (Frogner 2022):

• the initial state, i.e. the estimate of the current state of the atmosphere
• the data coming in from the edges of the model domain (the lateral boundary conditions), because MEPS calculates only northern Europe and gets the surrounding weather from another model
• the initial state of the land and sea surface
• values in the model's physics that are not known exactly, for example in how clouds and the boundary layer are represented

The control member is the member whose starting data has not been changed. The perturbations to the other members are random, but their size is set to match the real uncertainty in the observations and the model.

MEPS: five new members every hour

Every hour, MEPS calculates five new members, each reaching 66 hours ahead. The members from the last six hours are combined into a 30-member ensemble, which therefore reaches about 61 hours ahead (MET Norway, NWPdocs). Every third hour, one of the five new members is the control. So the ensemble always contains fresh calculations alongside ones a few hours old; this is known as a lagged ensemble. The ECMWF ensemble has 51 members: the control and 50 perturbed members (ECMWF, IFS Cycle 50r1).

How to read an ensemble

• Members close together: the forecast is more certain. For example, when nearly all 30 members show 7–9 m/s for the afternoon, you will probably get those conditions.
• Members spread out: the situation is uncertain. If 12 members show 4 m/s and 18 members show 10 m/s, you cannot rely on the average of 7.6 m/s. The wind will be either light or strong, and is unlikely to be anything in between.
• The proportion gives the probability: if 24 of the 30 members show more than 8 m/s, the probability of wind above 8 m/s is roughly 80%. The figure is only indicative, because ensembles are often overconfident.

According to ECMWF, a good ensemble forecast is reliable: when it gives something a 70% probability, it happens about seven times out of ten (ECMWF 2017).

VamoSpot currently shows the control member of the MEPS ensemble, i.e. a single calculation. VamoSpot does not yet show the ensemble spread.

Same model, different update times

A model run is not ready at its start time. First the system waits for observations, then it calculates, and finally the data is sent out for distribution. Every intermediary adds its own delay. That is why two apps can show the same model but a different run. The delay is counted from the run's start time, for example from 06 UTC.

HARMONIE through three distributors

Here the same run is compared across different sources. A new main run, i.e. the control, arrives every 3 hours, so the run you see is always at least as old as the delay and at most the delay plus 3 hours.

MET Norway also publishes the ensemble's five new members every hour. Measured on 8 November and 13–14 November 2026. The figures are based on a few days of observations, not service guarantees.

ECMWF

ECMWF waits longer for observations and calculates the whole globe, so its run is ready later than HARMONIE's. According to ECMWF's schedule, the control of the 00 UTC run is disseminated to weather services from 05:45 UTC and the full ensemble at 06:40–07:40 UTC. Open data is published at the end of the schedule. On Open-Meteo, where VamoSpot gets it from, a new run was available about 7–8.5 hours after the run started, depending on the run (measured on 13–14 November 2026). It reaches VamoSpot about 5–15 minutes later. A new run arrives every 6 hours, so the ECMWF run you see is about 7–14 hours old.

Other services

Yr, FMI's online forecasts, Foreca and Windy do not publish their delays. Yr's MEPS forecast is updated once an hour (MET Norway), and FMI's online forecasts several times a day (FMI).

Example: when a friend sees a different wind in another app, the reason may be that they have the 06 UTC run and you have the 09 UTC run.

What this means for you

• Check the run time. In VamoSpot it is shown in the Layers menu under Forecast model, for example "updated 12:48 · run 11:00" (in your device's local time). If a friend sees a different figure in another app, first check together whether it is the same run.
• Compare models. If HARMONIE and ECMWF show the same wind for tomorrow, the forecast is more certain. If they differ a lot, the situation is uncertain.
• Watch whether the forecast changes. If the same window of good conditions holds over three consecutive runs, you can trust it more than a single run.
• Look at the spread when you can. If the ensemble members spread out, be prepared for both light and strong wind.
• On the day, trust the detailed model and the wind gauge. For a few hours ahead, the latest HARMONIE run and the wind gauge tell you more than the previous evening's forecast.

See also Which wind forecast works on the water? and Spot settings for your sport and skill level.

Sources

Figures checked on 14 November 2026. Delays measured on 8 November and 13–14 November 2026.

Studies and presentations

• Brown, A., Vincent, C. and Short, E. (2026): Identifying sea breezes from atmospheric model output (sea_breeze v1.1). Geoscientific Model Development 19, 933–953.
• Eresmaa, R. (2023): The use of satellite data in limited-area NWP models of MetCoOp. EUMETRAIN, presentation (PDF).
• Eresmaa, R. et al. (2026): The Operational Forecast Process at MetCoOp.
• Frogner, I.-L. et al. (2022): Model Uncertainty Representation in a Convection-Permitting Ensemble – SPP and SPPT in HarmonEPS. ECMWF Annual Seminar 2022 (PDF).

Weather services and other sources

• MET Norway: MEPS dataset, MetCoOp, Locationforecast data model and open data catalogue
• Finnish Meteorological Institute (FMI): Numerical weather prediction, open data forecast models, Weather services on the internet and Application and verification of ECMWF products 2024 (PDF)
• ECMWF: IFS Cycle 50r1, dissemination schedule, Open data and Fact sheet: Ensemble weather forecasting (2017)
• DWD: NWP forecast data
• NOAA NCEP: GFS products
• Open-Meteo: Model updates and the models' run and publication times from the API metadata