Why there's more wind than forecast at the water's edge

Your phone says 12 knots, you get to the beach and it's blowing 18. That isn't bad luck, and it isn't "the model is rubbish" either. It's a problem of geometry — and it can be measured.

4 September 2026 · 6 min read

The number that started all this

On a spot on the French Atlantic coast, we queried a forecast model at two points 1.1 kilometres apart. At the spot's own coordinates: 9.3 knots. One kilometre further west: 18.1 knots.

Same model. Same hour. Same day. Nearly double.

Neither number is wrong. They simply describe two different places — and one of them is forest.

A grid cell is not a place

A weather model does not compute the wind "at Lacanau". It cuts the world into squares and computes one average value per square. The finest models now get down to around 1.3 km a side; global models work at more like 11 to 27 km.

On the open ocean this doesn't matter: the sea is uniform over tens of kilometres. On a coastline it's another story. Inside an 11 km square, the model has to fit the beach, the dune, the pine forest and the water — four surfaces whose roughness slows the wind by a factor that can reach two.

The model has to pick: it assigns the cell a dominant roughness. If your pin falls on the wrong side of that boundary, you read the wind of the forest and then go sailing on water.

We quantified the effect over 1,624 hours and a dozen coastal points: between two points 2 km apart, the forecast wind differs by 3.15 km/h at the median, 8 km/h at the ninth decile, and up to 24.7 km/h in the worst case. Moving a pin two kilometres could change the forecast by thirteen knots.

And the error is worse when there's little wind

Against wind actually measured by coastal sensors, the best model we tested is off by roughly 3.3 km/h on average. But that error is not spread evenly.

Relative to the mean wind, it is 45% in the 0-10 km/h band, 21% between 10 and 20, and 17% between 20 and 30. In other words: the harder it blows, the more reliable the forecast is, in relative terms.

That's counter-intuitive, and it's rather good news: the days when the call is hard — the marginal days, at the edge of your gear — are also the days when uncertainty matters most. On the days it really blows, the model roughly knows what to tell you.

"Gust": one word for two different things

Here is the sneakiest trap in coastal wind, and it has nothing to do with geography.

When a model reports a gust, it gives the maximum of the hour. When a weather station reports a gust, it usually gives a ten-minute average. Two different quantities carrying the same name, and they are not comparable.

We measured it. The forecast gust-to-mean ratio is 1.72 at the median when it genuinely is gusty… and 1.56 when it isn't. Less than 0.2 separates the two regimes — while the thresholds commonly used to declare "it's gusty" sit right between them.

What we found auditing our own app

This paragraph costs us something, and it is the reason this page exists.

Tide It displayed a "gusty" badge, set at a ratio threshold of 1.55. We just saw that the median of the non-gusty regime sits at 1.56. The threshold was therefore below both regimes.

Measured result: the app shouted "gusty" on 80.6% of its windy hours, of which 53% were provably false — contradicted by a station's reading at the same place and the same hour. And since that badge also weighs on the session score, it was quietly docking 11.3 points from perfectly sailable slots.

A warning that fires four times out of five stops warning about anything. You end up reading it the way you read a dashboard light that never goes out, and on the day it's telling the truth, you don't see it. That is the opposite of safety.

Fixed: two separate sets of thresholds, one for forecasts and one for measurements — because they are not the same quantity. Plus a 10 km/h margin, the typical model error on gusts, before any hard safety cut-off: without it, the high thresholds fired wrongly between 38% and 51% of the time.

What we took from it: measurement confirms, it doesn't correct

The general lesson of all this is that you don't repair a forecast with a measurement. You put them side by side.

• The forecast says what the wind will do. One model — the best-resolved one that covers the area — with speed, gust and direction all from the same model, because averaging several models manufactures a wind direction none of them predicted, when onshore versus offshore is what decides a session.
• The measurement says what it is doing, with its age always on screen. It confirms the present moment. It does not rewrite the coming seven days.
• The spread between models says how much they agree. We checked that this agreement does predict the error — but only up to a point: beyond it, ranking 0.96 above 0.88 is deciding on noise.

We also refused things that looked like good ideas. Correcting the forecast using a station's learned bias: measured, no out-of-sample gain, and a sheltered station dragged down the forecast for an exposed spot. Shifting the sampling point two kilometres offshore to "catch the sea cell": measured, and the app then announced 8 to 9 knots more than everyone else. A badly placed spot is fixed in the catalogue, never in the way you query it.

The problem nobody mentions: there is no station

Everything above assumes you can check. We measured how often you actually can: across 284 surf spots in 64 countries, only 34.2% have a usable anemometer within 15 kilometres.

And the distribution is unkind: 57% in North America, 52% in France, 47% in Oceania, 16% in Asia — which also happens to hold the largest share of the spots.

Why 15 km and not 30? Because beyond that you start importing airfield sensors sitting inland. On one spot near Arcachon, a station too far away made the app show "real 9 knots" against a forecast of 17. Both numbers were right where they stood; putting them side by side meant nothing. Widening the radius to 25 km would lift the displayed coverage by fourteen points with a single setting — that would be a number, not a measurement.

The method, and what it does not prove

The figures quoted here come from comparing forecast models against wind actually observed by public sensor networks: first 17 coastal stations and 2,800 measured hours, then a broader audit of 93 stations across 14 countries and 26 spots on every continent, in summer and in winter.

We'd rather say so. A measurement presented without its limits is an opinion dressed as a number.

What you can do with this

• Check where the pin is before blaming the model. On a coastal spot, a few hundred metres are worth several knots.
• Distrust a "gusty" badge that is always on. Ask what quantity it is set against — an hourly maximum and a ten-minute average do not compare.
• Look for a real reading within 15 km. Beyond that you're comparing two places, not two methods.
• In light wind, doubt more. That's where the relative error is largest — and precisely where you hesitate.