How our forecasts work

Last updated: 15 October 2026.

TL;DR

Every spot gets a star rating for each forecast hour, from 0 (FLAT) to 5 (EPIC). It takes three steps:

1. Size. The wave height, adjusted for swell direction, gives a size score.
2. Quality. Four scores adjust the size score: wind, chop, period and tide. They are blended with weights, not simply multiplied together.
3. Rounding. The result is rounded to the nearest half star and kept between 1 and 5. A height under half a foot, after the direction adjustment, is FLAT: 0 stars.

stars = size_score × wind^0.35 × chop^0.25 × period^0.25 × tide^0.15

How the stars are built

Size

The size score comes from the wave height after the swell-direction adjustment (see the swell window, below). It's the model's height at every spot: at the 141 calibrated California spots the page shows a calibrated height, but the rating uses the model's, as everywhere else.

Between rows, the score moves in a straight line. Under half a foot, the spot is FLAT whatever else is going on.

Quality

Four scores adjust the size score. Each is close to 1 when that part of the conditions is good, and lower when it isn't. Wind and period can go a little above 1.

The four are blended with these weights, which add up to 1:

Technically it's a weighted geometric mean: each score is raised to the power of its weight, and the results are multiplied. In practice:

• If all four scores are 1, you get exactly the size score, before rounding.
• Wind moves the rating most, and tide least.
• A few middling scores don't pile up the way they would if you multiplied them straight.

Example. Say a swell comes straight into the spot at 4 ft, which scores 3.0 for size. The wind is offshore (1.20), there's some chop (0.85), the period is 10 s (0.85) and the tide is wrong for the spot (0.70). Weighted, those four come to 0.93. So the rating is 3.0 × 0.93 = 2.8, which rounds to 3 stars. Multiplied straight, the same scores would give 1.8.

Where the data comes from

Where the wave height comes from

The model behind the height depends on the spot and the hour:

• Most spots: NWPS. NOAA runs NWPS separately for each coastal forecast office, on a much finer grid than the global models. We read a grid point just offshore of the spot, for each hour the latest run covers.
• 48 California spots: CDIP MOP, for past hours only. MOP is the Coastal Data Information Program's model of points along the California coast, just outside the surf zone. We read its nowcast, and the nowcast ends before the current hour, even right after a forecast run. So at the 48 MOP spots, the hours you can see, the current one included, are handled like any other hour, below.
• Any other hour: WAVEWATCH III. The WW3 swells are combined as described below. If WW3 has nothing usable for that hour either, we fall back to NWPS's own height.

Calibration at 141 California spots. For these spots we compared our height with MOP's over four weeks, 8 September to 6 October 2026, and now divide the height we show by each spot's typical ratio. The star rating isn't divided: it comes from the model's height, the same input every other spot is rated on.

What the number is. The height we publish is labelled nearshore swell height: the significant height of the waves just outside the surf zone, roughly the average of the bigger waves. It is not the face of a breaking wave, which is often bigger. The label holds at the 141 calibrated California spots, which are scaled so their typical height matches MOP's. Everywhere else, the model height gets a boost for longer-period swell and isn't corrected, so it can read higher than nearshore swell height. That includes all the hours you can see at the 48 MOP spots: MOP's own height, which carries no boost, never reaches the page. On the site, each spot's current height is marked with which it is: "Calibrated to CDIP's nearshore model" or "Model estimate".

Swell direction and period

WAVEWATCH III splits the sea at each grid point into up to three swells, each with its own height, period and direction, plus a separate local wind sea. We read it in 3-hour steps, and each forecast hour uses the nearest step.

When the height comes from WW3, the swells are combined as energy rather than added up:

combined height = √(sum over swells of height² × direction gain)

The local wind sea is left out, and a swell 90° or more outside the spot's window adds nothing.

The period and direction we show come from one swell, the dominant one. We pick it by height² × direction gain × period score², so long-period swell counts extra. From the same direction, a 2 ft 14 s swell beats a 3 ft 7 s swell: 2² × 1.02² ≈ 4.1 against 3² × 0.60² ≈ 3.2.

Other sources step in when WW3 can't:

• At the MOP spots, for MOP's hours, direction and period come from MOP.
• If WW3 has no usable swell for an hour, we use NWPS's own direction and period.
• At some spots and hours, the nearest buoy's latest reading stands in when the models give no swell direction. It's a snapshot: the same reading is used for each hour of the forecast.

The swell window

Every spot has a swell window: the range of directions it can get swell from, given the headlands and seabed around it. The first version of the rater scored any swell outside the window as zero. A refracted NW swell wrapping into a south-facing point like Steamer Lane counted for nothing, even when the wave model said the energy was there.

Now each swell gets a direction gain:

Inside the window, the offset is the angle between the swell and the spot's best swell direction. A swell straight down that line gets 1.00, and one 60° off it gets 0.75. Outside the window, the angle is measured from the window's edge. Swell from outside the window isn't impossible: it wraps in at reduced strength.

Wind

The first version of Stormy Petrel used the wind that comes with the NWPS wave forecast. Users told us it said offshore when it was clearly onshore. HRRR is a 3 km model with hourly steps, and it picks up sea breezes and terrain-driven offshore winds much better.

We use HRRR for the first 48 hours of each run, and NWPS wind after that. HRRR only covers the continental US, so Hawaii and Puerto Rico use NWPS wind throughout. Each spot takes its nearest HRRR grid point.

The wind score depends on the wind's angle to the spot's offshore direction: offshore scores best, onshore worst. Light wind matters less, strong offshore wind loses its bonus, and very strong wind costs extra.

Period

Long-period swell usually makes better surf than short-period wind swell of the same height. The period score:

Between rows it's a straight line, so 14 s gives about 1.02.

Chop

When local wind sea is a big share of the wave height, the lineup is bumpy even if the swell lines up well. We measure that share:

chop_ratio = (total_hs - swell_hs) / total_hs

The chop score is a straight line between these points, not a step, so a chop_ratio of 0.3 gives 0.9250, not 0.85:

chop_ratio is left empty when the inputs can't support a ratio: no total height, no swell height, or a swell height above the total. An unknown scores the neutral 1.00 rather than being treated as zero.

Heavy chop also takes away the offshore-wind bonus. When more than 40% of the height is wind sea, a wind score that would be above neutral is held to 0.80.

Tide

The tide score compares the tide at that hour with the tide the spot prefers: low, mid, high, or a range in between. It's full marks when they match and lower when they don't.

A preferred tide only counts once it's verified. We use it only when a person has reviewed it, or when it comes from a page about that spot that we cite: a source we researched, or the spot's own surf-forecast.com page. Any other preference scores neutral, the same as a spot that works at all tides. So does a spot with no preference, and an hour with no tide data.

Why: every preferred tide we held, 444 of them, was an estimate written by a language model that had been told to give an answer even when it didn't know one, and 365 of the 444 said "mid". The tide score can only lower a rating, never raise it. When we replayed the forecast published on 15 October 2026 with those estimates set aside, 6,574 of 83,359 spot-hours (7.9%), across 426 of our spots, came out half a star higher, and none came out lower. A guess shouldn't cost a spot half a star, so for now no spot's preferred tide affects its rating. Each one will count again once it's verified.

Surface conditions

The Clean / Mixed / Choppy / Blown out word on each spot page is separate from the chop score above. It comes from the wind, not from chop_ratio.

That's deliberate. "Blown out" is a wind condition: the Encyclopedia of Surfing defines it as an "ocean surface condition created by a moderate-to-strong onshore wind, which, by degrees, produces chopped-up, crumbly, messy surf." chop_ratio describes the water offshore, which is a different thing. When we measured it in August 2026, across 84,774 spot-hours, a label based on chop_ratio barely tracked the wind: 57.3% of offshore-wind hours were labelled "Blown out".

The label now comes from the wind's angle off the spot's straight-offshore direction (off_angle, 0–180°) and its speed:

Speed is checked first: a 1 m/s straight-onshore hour is glassy, not choppy. On the same August 2026 hours, this gives Clean 43.5%, Mixed 23.8%, Choppy 28.3% and Blown out 4.4%.

chop_ratio still appears on the spot page, as the "swell mix" figure under the word.

What we're still working on

1. Local geology. A reef pass that focuses energy, or a sandbar that moves with the seasons, isn't in any model. The calibration at 141 California spots corrects each one's typical height; other spots have no local height correction.
2. Crowds aren't modelled. Two spots with the same conditions can be very different sessions.
3. Tide preference. No spot's preferred tide is verified yet, so for now the tide doesn't change any rating (see Tide). We're researching them spot by spot, each from a source we cite.
4. Height calibration drifts. Each spot's factor comes from a single four-week window, and it moves with the swell mix: between two of the earlier two-week windows, about a third of spots shifted by more than 10%.
5. Uncalibrated heights can run high. Outside the 141 calibrated California spots, heights aren't corrected, and that includes all the hours you can see at the 48 MOP spots. They include the period boost and can read higher than nearshore swell height, which is why they're marked as model estimates. At the California spots where we measured it, the uncorrected height was typically about 1.6 times MOP's.
6. Calibrated spots used to rate lower. We used to work out stars at the calibrated spots from the calibrated height. The calibration in use then covered 130 California spots, measured from 18 August to 1 September 2026, and at all but two of them that height was smaller than the uncalibrated one, so in the same swell they rated lower than everywhere else. In the forecast published on 4 October 2026, 0.4% of those spots' daylight hours rated 3 stars or more, against 9.1% at uncalibrated spots; rated on the model's height instead, they came to 10.3%. Now the calibration corrects only the swell height we show, and the rating uses the same input at every spot.
7. Refresh delay. Forecasts are updated several times a day, not continuously. If a swell builds faster than that, the rating lags until the next update. Buoy readings arrive throughout the day, and we show the latest one next to each spot.

The code is on GitHub. The rating is in pipeline/interpret.py, and the NWPS, MOP and calibration steps are in nwps_nearshore.py, mop.py and face_correction.py. File an issue if you spot a bug, or send a PR if you want to fix one.