ECMWF, GFS or ICON: which weather model should sailors trust?

In this piece

1. Who runs ECMWF, GFS and ICON
2. Resolution is not accuracy
3. ECMWF vs GFS vs ICON: which is most accurate?
4. Why GFS is still worth a look
5. What none of them see near the coast
6. How to compare weather models before a passage
7. How far ahead can you trust a forecast?
8. GRIB files and the official forecast
9. Keep score in your own waters

If you can only look at one, the best weather model for sailing is ECMWF. In the medium range, from about two days out to ten, it has led the published verification scores for most of its history, and that is the range where most passage decisions get made.

Then look at the other two anyway. Put GFS and ICON beside it for the hours you care about. When all three tell the same story, you can plan on it. When they disagree, the disagreement is the forecast: the atmosphere has not made up its mind, and the plan you want is the one that still works if the worst of the three comes true. For the next couple of days close to the coast in Europe, give the finer grid of ICON-EU a proper look. And read the official marine forecast and its warnings before you leave, because a forecaster has already done this comparison with local knowledge you do not have.

That is our routine. Here is why.

Who runs ECMWF, GFS and ICON

All three are global numerical weather prediction models. Each one takes millions of observations from satellites, weather balloons, aircraft, ships, buoys and land stations, builds a picture of the whole atmosphere as it is now, then steps that picture forward in time on a three-dimensional grid of boxes. They differ in grid, in physics and above all in how good that starting picture is.

The figures are approximate, and they move: every few years each centre upgrades its model, usually with a finer grid.

ECMWF

The European Centre for Medium-Range Weather Forecasts is an intergovernmental organisation backed by more than thirty European states, with its headquarters in Reading, in England, and its supercomputers in Bologna. Its model is properly called the IFS, but everyone says ECMWF. Its reputation rests less on the grid than on data assimilation, the step that turns millions of scattered, imperfect observations into one consistent starting state, which it has done better than almost anyone for decades. Its forecasts used to be sold rather than given away, which is why for years the free apps and GRIB services showed GFS. In the last few years ECMWF has made much of its real-time data free.

GFS

The Global Forecast System is run by the US National Weather Service, part of NOAA. US government data is free, and GFS has been free to download for decades, so a whole generation of GRIB tools, routing software and apps was built on it. It runs four times a day, reaches 16 days ahead, and usually trails ECMWF in the medium-range scores.

ICON and ICON-EU

ICON belongs to the Deutscher Wetterdienst, Germany’s national weather service, and was developed with the Max Planck Institute for Meteorology. Its grid is built from triangles rather than boxes of latitude and longitude, hence the name: icosahedral nonhydrostatic. The global run reaches about a week ahead. Over Europe, including the whole Mediterranean, DWD runs a finer nest called ICON-EU at about half the spacing, out to about five days. In the Aegean, the Adriatic or the Balearics over the next two or three days, that finer grid has the best chance of the three of drawing the larger islands and gulfs.

Resolution is not accuracy

Grid spacing is the size of the boxes the model works in. A model needs several boxes to draw any feature properly, so a model with a 9 km grid sees a 10 km island as a bump and a 2 km channel not at all. Mountains are smoothed too. A ridge that rises 1,000 m straight out of the sea may stand at a few hundred metres in the model, which is one reason models underplay what that ridge does to the wind.

So a finer grid helps, up to a point. A 6.5 km model that started from a slightly wrong picture of the atmosphere draws a beautifully detailed wrong answer. Much of ECMWF’s lead comes from its starting state and its physics rather than its grid, and that is why it tends to beat finer models once you look beyond the first day or two.

One practical trap: many GRIB services and apps hand these models out on a coarser grid than they were run on, often a quarter of a degree, which is about 28 km north to south. The detail has been thinned before it reaches you. If you download GRIBs, check which grid you are actually getting.

ECMWF vs GFS vs ICON: which is most accurate?

Over the whole globe and over a season, ECMWF. In the standard scores its lead over the other global models has typically been worth something like half a day to a day of forecast range: its five day forecast is often about as good as the others’ four day forecast. That gap matters most between days three and seven, when you are deciding whether to leave on Thursday or wait for Saturday.

Averages hide a lot, though. Any of the three can be the one that got it right on a given day, and for the next 24 hours in a particular bay, the regional models that many national weather services run at 1 to 3 km can beat all three. The scores are a good reason for ECMWF to be the first tab you open, and no reason to close the others.

Why GFS is still worth a look

Sailors who have just discovered ECMWF tend to write GFS off. It deserves better. Having been free for so long, it is everywhere, including the email and satellite GRIB services that offshore crews rely on. Its 16 days are no basis for planning a passage, but they hint at whether a big change in the pattern is coming in the second week.

Above all, it is independent: different data assimilation, different physics, a different team on a different continent. When GFS and ECMWF reach the same answer by different routes, that is worth far more than one model saying the same thing twice.

What none of them see near the coast

Every global model is weakest where you spend most of your time, within a few miles of land.

Acceleration zones

Wind squeezed between two islands or bent round a headland speeds up, often by a force or more. In our waters the Kafireas strait between Evia and Andros in a meltemi is the classic case (the meltemi guide has more on it), and every coast has its own; the Strait of Bonifacio is another famous one. The gap is narrower than the grid, so the model averages the fast wind in it with the slower wind either side.

Gusts off high land

On the lee side of a high island, wind comes over the ridge and down the slopes in bursts far stronger than the wind over open water. At night, cold air sliding off high ground (a katabatic wind) can pour down a valley onto an anchorage that was calm at dusk. Both depend on the real height and shape of the land, which the models have smoothed away.

Sea breezes

On a hot day the land heats, the air over it rises, and cooler air flows in off the sea to replace it. On a coast with little gradient wind, that afternoon breeze is most of the wind there is. Global models do produce sea breezes, but often too weak or at the wrong time for a given bay. A forecast of 8 knots that blows 18 from 14:00 is common enough that we no longer find it surprising.

Squalls and thunderstorms

Individual shower clouds are far smaller than the grid. A model can tell you that thunderstorms are likely in a region; it cannot tell you which bay one will cross. Under a squall the gusts can be double the wind around it.

Which brings us to gusts in general. Models do not simulate individual gusts. They estimate them from the mean wind and how turbulent the air is likely to be. Over open water that estimate is usually fair; in the lee of high ground, under showers and along cliffs it is usually too low. We read the forecast gust as the least we should expect in those places. Our guide to reading a marine forecast goes further into gust factors.

How to compare weather models before a passage

None of this needs special software, just a few minutes and the habit of looking at more than one tab. It is the weather half of passage planning.

1. Four or five days out, look at the pattern. ECMWF and GFS, and the ensembles if you can get them. The questions are big ones: is a front coming, does the meltemi set in, is there a window. Ignore the hourly numbers for now.
2. Two days out, open all three. Look at the hours you would be sailing. Do they agree on direction, on the peak, on the timing of any change? Agreement within a few knots and a couple of hours is real confidence.
3. Watch the runs as well as the models. A forecast that holds steady through three or four runs is more believable than one that jumps each time. A low that moves 100 nm between last night’s run and this morning’s is telling you it is not pinned down.
4. Plan for the strongest plausible answer. If ECMWF says 18 knots and GFS says 26, the plan has to work in 26. Often that just means leaving three hours earlier or putting the reef in at the dock.
5. Read the official forecast that morning. The national marine bulletin and any warnings, online, on VHF or on Navtex. Where it is stronger than the models, take its word. Then add what no model sees: the channels, the capes and the lee of high land.

How far ahead can you trust a forecast?

Roughly: the first two days are good enough for hour by hour decisions, though the timing of a change can still be a few hours out. Days three to five are good for the pattern, for whether something happens more than exactly when. Beyond five days, any single model run is one possibility among many. Beyond a week, the hourly numbers in a 16 day forecast are close to noise, and the trend is all you should take from them.

That is the job of ensembles. Instead of one run, the centre runs its model dozens of times from slightly different starting states: ECMWF’s ensemble has about 50 members, the GFS ensemble about 30. When the members bunch together, confidence is high. When their lines fan out, nobody knows, and the honest forecast for that day is anything from a calm to a blow. Plenty of free websites show ensemble plumes for a single point, and they are the best tool there is for deciding whether to wait until Saturday.

GRIB files and the official forecast

A GRIB file (gridded binary, a format standardised by the World Meteorological Organization) is a model’s raw numbers on a grid. It is small enough to download over a satellite connection, and offshore it may be all you have. But nobody has looked at it: no forecaster’s judgement, no warning, no “locally 8 near the capes”. Treat it as raw material.

The official marine forecast is the other end of the scale, written by people who have seen every model and the latest observations and who know where their coast misbehaves. In the Aegean that means the Hellenic National Meteorological Service on the Greek side and the Turkish State Meteorological Service on the other. When it and the models disagree, give it the benefit of the doubt.

Keep score in your own waters

The published scores are averages over the whole planet. What you need to know is how each model behaves in the few hundred square miles where you sail, and the only way to find out is to keep notes. Before a day sail, write down what each model gave for the afternoon. That evening, write down what you actually got. By the end of a season you will know which model runs light in your channel and which one is always late with the sea breeze, and no verification score will tell you that.