Frequently asked questions
Short answers to the questions people ask most about weather models, each linking to the full explanation.
The basics
What is a weather model?
A computer simulation of the atmosphere, built on the laws of physics. It starts from today’s observations and is stepped forward in time. Its output is numbers on a grid (wind, pressure, rain, waves) which forecasters and apps turn into forecasts.
What types of weather models are there?
Global models cover the whole Earth; regional models zoom in on one area; ensembles run a model many times to show the range of outcomes; wave and ocean models handle the sea; AI models are a newer family. Nowcasting covers the next few hours only.
How is a forecast calculated?
The atmosphere is cut into millions of boxes and the equations for motion, heat and moisture are solved for each box, one short time step after another. Things smaller than a box, such as thunderstorms and turbulence, are approximated. Halving the box size makes a run about eight times more expensive.
Where does the data come from?
Satellites supply most observations by number; the rest come from weather balloons, aircraft, ships, buoys, land stations and radar. A process called data assimilation blends them with the previous short forecast into one complete starting picture.
Does a model look up similar weather from the past?
No. A physics-based model solves the equations from the current state. Past data is used to build and test models, to correct known errors, and to train AI models, but not as a lookup table.
Which model, and how good
GFS or ECMWF: which is more accurate?
In published comparisons ECMWF’s model has generally scored higher at medium range, but it varies by place, season and weather situation, and the other can win in a single case. Look at both, and at an ensemble, rather than relying on one.
Is a model that updates more often better?
Not by itself. A model that runs more often gives you newer information, but accuracy depends on the starting data, the physics and the resolution. Compare the models and always use the latest run of each.
Which model is best for sailing in my area?
There is no single best one. Use a global model for the big picture, the finest regional model that covers your coast for the next one to three days, and an ensemble for confidence. Far offshore, grid spacing matters little.
How good are AI weather models?
Several match or beat the leading physics-based models on common measures at medium range, and they run in minutes. They tend to give smooth fields and can be weaker for extremes and local detail, and most still start from a physics-based analysis. Use them as another opinion.
Why do different apps show different forecasts?
They may use different models, or a blend; show different runs; pick your point from the grid in different ways; correct the model differently; and show different wind (mean or gust). Set the same model, time zone, units and wind type before comparing.
Resolution and scale
What does “13 km” or “2.5 km” on a model mean?
It is the grid spacing: the distance between neighbouring points where the model keeps a value. One nautical mile is 1.852 km, so 13 km is about 7 NM and 2.5 km is about 1.35 NM.
Does a finer grid always mean a better forecast?
No. It helps most near coasts, islands and mountains and for squalls, but a fine model with poor starting data can be sharply wrong, and fine models cover small areas for only a day or two. Far from land, 9 to 13 km is often just as good.
Why can’t a model see a sea breeze or a narrow strait?
A model shows features well only if they span about five to seven grid spacings. A 13 km model blurs anything smaller than roughly 65 to 90 km, and narrow straits, fjords and small islands can disappear from its land-sea map.
Disagreement and uncertainty
Why do two models with the same resolution disagree?
Grid spacing is only one setting. Models start from different analyses, use different rules for clouds, convection and turbulence, different grids, land-sea masks and numerics, and different ways to compute the output. Small differences also grow with time.
What are ensembles and spaghetti plots?
An ensemble is one model run many times from slightly different starting points. Where the runs agree, confidence is high; where they spread out, it is low. Counting the runs gives a probability, such as the chance of exceeding your wind limit. A spaghetti plot draws one feature, such as a storm track, for every run.
How many days ahead can I trust a forecast?
Detail is reliable for about two or three days, trends for up to about a week, and after that you only get a range of possibilities. The limit comes from chaos in the atmosphere, not from computing power.
Why are forecasts so often wrong near the coast?
Land friction, sea breezes, headlands and mountains change the wind on scales of a few to a few tens of kilometres, smaller than the boxes of a coarse model, and the land-sea boundary falls inside one box. Use the finest local model and add a margin.
Wind, waves and storms
What is the difference between mean wind and gusts?
The forecast wind is an average at 10 m height over about ten minutes. A gust is a brief peak, usually averaged over about three seconds, and is commonly 30 to 60% higher at sea. Reef for the gusts, not only the average.
What is significant wave height?
The average height of the highest third of the waves, which matches what an observer estimates by eye. The largest single wave is typically about twice that. Read it together with the period and direction.
Why are the waves worse than forecast?
The wave forecast inherits any error in the wind forecast; swell from far away may not be in your local wind; shallow water steepens waves; and wind against current makes them shorter and steeper. Look at the period, not only the height.
Why are storm tracks forecast better than storm strength?
The track is steered by the large-scale flow, which models handle well. The strength depends on a core of intense wind only tens of kilometres across, smaller than a global model’s boxes, and on small changes in ocean heat and wind shear.
What does the cone of uncertainty show?
Where the centre of the storm may go, based on past errors; the centre stays inside it about two times in three. It does not show how big the storm is or how strong it will be: wind and waves reach far outside the cone.
Timing, files and checking
When does a new run appear, and why the delay?
A run is named after its start time (00Z, 06Z, 12Z, 18Z in UTC) but the files appear hours later, because observations must be collected, the analysis built, the forecast computed and the output uploaded. The early forecast hours appear before the later ones.
What is a GRIB file, and how big is it?
The standard compact format for gridded weather data: each record holds one parameter at one time over a grid. A coarse file for a sea area can be tens of kilobytes; a fine one a few megabytes, which is why it works over satellite and email.
How can I check a forecast against reality?
Write down the forecast before the event, compare it with a nearby station, buoy or your own log for the same time and the same quantity, and repeat for ten to twenty cases. The average miss (bias) shows which way the model leans in your waters.