How a Surf Forecast Actually Works

When you check the forecast at 6 a.m. and see “head-high, 3.2 stars” for tomorrow, you’re reading the output of a 24-hour computational relay race. A real storm is brewing thousands of miles away. Overnight, a supercomputer simulated where its winds will go and how strong they’ll be. Those predicted winds got fed into a separate model that grew waves on a global grid, then sent them traveling across the ocean. The energy was bent and shoaled by your local bathymetry. Near the end, a forecaster who knows your break nudged the rating up or down.

This is the final piece in our forecasting series. It ties together the period, direction, wind, tides, and buoys articles into a single picture. The big idea: a forecast is a stack of nested models, with each layer feeding the next. Knowing how they fit together tells you which ones to trust and where errors come from.

The five layers

Every surf forecast on every app runs through roughly the same chain. Here’s what each layer does.

Layer 1: The weather model

Everything starts with wind. Until you know what the wind has been doing across an entire ocean basin for the past several days, you can’t predict any waves.

The big two are GFS (run by NOAA) and ECMWF IFS (run by the European Centre for Medium-Range Weather Forecasts). Both run on supercomputers four times a day at horizontal resolutions of around 9 to 13 kilometers globally. They solve the same physical equations (how air moves on a rotating planet, how heat and moisture get carried around) but with different numerical schemes and different initial data. ECMWF has been consistently more accurate than GFS for most of the last two decades.

For coastal wind detail in the US, there’s also HRRR (High-Resolution Rapid Refresh), a 3-km model that updates every hour. It captures sea breezes, marine layers, and Santa Ana outflows that the bigger global models smooth away. HRRR is what tells you whether your dawn-patrol offshore will hold until 10 a.m. or get shredded by 8.

If the model gets the wind wrong by even a small amount, the wave forecast ends up significantly off. A wind error of just a couple miles per hour can translate to a 10 to 15 percent error in predicted wave height. Most forecast busts trace back to this layer first.

Layer 2: The global wave model

The wind forecast gets fed into a global wave model. The standard in the US is WAVEWATCH III (developed at NOAA). The European equivalent is WAM.

These don’t predict “wave height.” They predict the full wave spectrum, meaning how much energy exists at every wave period and every direction, at every point in the ocean. Significant wave height is just a summary of the spectrum.

Wave models track four things: waves being generated by wind, waves traveling across the ocean, waves redistributing energy among themselves through nonlinear interactions, and waves losing energy to whitecaps and friction. Global wave models work on a grid where each cell is roughly 10 to 15 miles across, which is way too coarse to see your specific surf break. They give you the open-ocean signal.

This is the same model output that powers the buoy partition algorithms from the buoy article. When a buoy reports “4 ft of 14-second NW swell + 2 ft of 6-second wind sea,” those are partitions of a model spectrum.

Layer 3: The nearshore transformation

A global wave model knows what’s happening offshore, but it doesn’t know about your point break, your reef, or your sandbar. To get from “open-ocean spectrum” to “wave height at your spot,” you need a nearshore transformation.

The standard tool is SWAN, developed at Delft University. SWAN takes the offshore spectrum and propagates it through high-resolution bathymetry, meaning the actual underwater geography of your coastline. It applies refraction (waves bending toward shallower water), shoaling (waves slowing and steepening as they hit the shelf), island shadowing, and depth-induced breaking.

The result is spot-specific. Not “this ocean basin has 6 ft of 14-second swell” but “your break should see 4 ft.” This is where two spots ten miles apart can get totally different forecasts. Same offshore signal, different bathymetry.

For California, the CDIP MOP system at Scripps does this at very high resolution (100m × 100m grid), using actual buoy measurements as the initial condition rather than just modeled wind. CDIP forecasts are often more accurate than the standard chain because they start from observed reality.

Layer 4: Wind + tide blend

The wave forecast at this point is just the wave field. Real surf depends on wind quality and tide too. The forecast app blends:

• Local wind forecast (HRRR in the US, or ECMWF for global)
• Tide predictions (deterministic, computed decades in advance)
• Spot-specific corrections for things like wind shadowing or breeze timing

This is where you start seeing the “wind quality” and “tide stage” annotations next to the wave height.

Layer 5: The human overlay

The last step is interpretation. Surfline’s in-house model (called LOTUS) takes the chain’s output and runs spot-specific corrections. But the star rating you actually see has typically been adjusted by:

• Spot-specific calibration based on years of “what the model said vs what showed up”
• Forecaster judgment (“the model is overestimating the south swell this morning”)
• Increasingly, machine-learning corrections trained on historical forecast errors

Free apps (Windy, surf-forecast.com, etc.) typically skip the human overlay and present model output directly. The paid services with human forecasters tend to be more accurate at well-known spots. Free apps can be more accurate for obscure ones where there isn’t enough observation data to train spot-specific corrections.

What star ratings actually mean

A star rating is not a wave-height measurement. It’s a composite score that combines several factors:

• Wave height relative to the spot’s potential, not absolute size
• Swell period, where longer is better with diminishing returns past 16 to 18 seconds
• Swell direction, meaning how well it aligns with the spot’s optimal window
• Wind, where offshore is good, side-shore is neutral, onshore is bad, and strength matters
• Tide, meaning alignment with the spot’s preferred window

A 3-star at Lower Trestles means something completely different than a 3-star at Mavericks. The score is normalized to each spot’s own potential. A 5-star at Trestles is “perfect Trestles,” meaning head-high, organized, offshore, ideal tide. A 5-star at Mavericks is double-overhead-plus with a long-period west swell.

The same offshore conditions can produce wildly different ratings at different spots, because the rating reflects how well those conditions match what that specific spot needs.

Why forecasts get worse with lead time

Forecast skill is bounded by chaos. This is the same effect Edward Lorenz discovered in the 1960s, where tiny errors in the initial conditions grow exponentially over time. The skill ceiling for atmospheric forecasts in the medium range is roughly 14 days. Beyond that, no amount of computing power can tell you reliably what tomorrow’s wind will be.

For surf forecasts in practice:

• 24 hours out: very reliable. Treat as fact.
• 72 hours (3 days): solid. Treat as a strong working hypothesis.
• 5 days: pattern recognition. The forecast captures the rough shape, like a swell coming from a certain direction. But timing might slip by 12 to 24 hours and size might be off by 20 to 30 percent.
• 7 to 10 days: vibes only. The model can see large-scale patterns (“a storm is going to develop in the North Pacific”) but can’t predict exactly where or how strong.
• Past 14 days: no skill. Don’t trust it.

A note on AI

Since 2023, AI weather models (DeepMind’s GraphCast, ECMWF’s AIFS, Huawei’s Pangu-Weather) have started matching or beating traditional numerical weather prediction at standard skill metrics, at a fraction of the compute cost. Wave-model AI analogues are coming. Surfline has been incorporating machine learning into LOTUS since 2021.

Nobody yet knows how much further AI will push the skill ceiling. What’s certain is that the physics of swell propagation across an ocean, which is the most predictable part, will keep working the same way it always has. The wind forecasts driving those waves just keep getting better.

How to use this

1. Look at ensemble spread when you can. Most apps show only the deterministic run, but the full picture comes from running the model many times with slightly different starting conditions. If all the runs show the same thing, trust it. If they disagree, the model is uncertain.
2. Check the buoys as a swell approaches. Once a swell is within 12 to 24 hours of arrival, the deep-water buoy will see it. That’s a reality check on the model.
3. Treat star ratings as a starting point. The model can tell you everything except whether the session will be worth it for you. That last call is still yours.

The forecast is a stack of models. Outer shell: a global atmospheric model. Inside that, a global wave model. Inside that, a nearshore transformation. Inside that, a wind blend and a tide model. And at the very center, a human who has surfed your break in winter and summer, in offshores and onshores, and who is nudging the score based on something the model can’t quite see.

Knowing this lets you read a forecast critically, and spot errors when they happen.