How good are these forecasts?

For every SailGP event: each team's chance of winning each race, reaching the final and winning the event, updated after every race while the event is on. For the season: each team's projected points and chance of reaching the Grand Final and winning the championship, and a first look at next season. Ratings for every team and driver.

Wind is classed relative to SailGP's own history rather than fixed knots, because weather records are coarse over water: light air is the calmest third of race days (under 7 kn in the records), strong wind the windiest third (over 11 kn), and choppy water the roughest third of wave heights (over 0.9 m).

The tidal and ocean current on each course comes from Open-Meteo's marine model (the nearest sea cell, about 8 km across, so a current squeezed through a harbour mouth reads low), averaged over the racing hours. It goes back to 2022 and is forecast about a week ahead; lakes have none. A race day counts as a strong current in the top third of SailGP's current readings (over 0.47 kn in the model).

It is a model input in two ways: each team has its own strong-current adjustment (like its light-air and strong-wind ones), and a weight of -0.30 fitted on the tuning seasons sharpens the forecast in a strong current (the better crews made more of it). Both stay in the model whatever the numbers say, because the current is a real part of every course.

Honestly, so far: on the tuning seasons the current improved the log loss by 0.0006 on the races (264 races, 174 with current data); on the held-out 2026 races it made the log loss worse by 0.0009 on them (1.9161 without it, 1.9170 with it, over 94 races). With only a few seasons of current data this is a small, noisy test, and it is re-run after every event.

Each boat has a strength. In a race, the chance a boat wins is its strength divided by the combined strength of the fleet; second place is then decided the same way among the boats left, and so on down the order. This is a Plackett–Luce model, and it uses every finishing place, not just who won. Boats that don't finish count as beaten by every boat that did.

A boat's strength adds up these pieces, each estimated from all races since 2019:

Older races count for less (a race's weight halves every four years), and each piece is pulled gently towards average so a team or driver with few races doesn't get an extreme rating. Ratings are shown on a scale where the fleet averages 1500 and 100 points means about a 64–36 edge head to head.

We sail each event 20,000 times in simulation, following the real rules: 10, 9, 8… points per fleet race (or 5, 4, 3… in each split-fleet group, seeded 1-2-2-1 from the previous event's order), a winner-takes-all final for the top three (or top two of each group), ties broken on the last race, and season points of 10 down to 1 by event place. After the last event the top three race a single Grand Final for the title. Each simulated race day draws its wind from the forecast or the venue's history.

During an event, races already sailed are fixed and only the rest are simulated, which is how the odds move race by race. The ratings themselves don't change mid-event.

Next season is simulated with the published calendar and current drivers. Every team's strength also gets a random change of the size teams have typically seen between seasons (about 55 rating points), so those odds are wide on purpose.

Every past event is re-forecast using only races sailed before it. The model's settings were chosen on 2021–22, 2022–23, 2023–24, 2025 and the 2026 season was kept aside as a test it never saw, and is re-scored after every event. On that test it is 6% better than treating every team as equal at picking race winners, and its favourite won 3 of 11 events.

It is not clearly better than the simplest thing that works. Ranking teams purely on their average finishing place over their last 30 races scores 1.9138 against the full model's 1.9177, a gap far smaller than the margin of error on 94 races. We report this rather than bury it: the model earns its place by also simulating events and seasons, not by beating that baseline race by race.

Not everything we tried made it in. A separate strength for each team in each season looked good on the test season but not on the tuning seasons, so it was left out rather than chosen after seeing the test.

SailGP Insights publishes every boat's telemetry and the race-control log, from 2023 on (37 events, 192 races). The log records each penalty, who gave it and why.

Since 2026, when the log started naming deliberate breaches, 17 of the 76 penalties in light-air races were for deliberately breaking a rule (0.07 per boat per race), against 0 of 157 in medium and strong wind. It is a light-air tactic: a boat about to drop off its foils can lose less by fouling and taking the penalty than by losing the foil.

We tested these as model inputs, each fitted on 172 races from the tuning seasons and scored on 94 held-out races. Recent penalty rate is the one that earned a place: fewer penalties recently means a slightly stronger boat. It saved 0.0084 on the tuning races and 0.0060 on the held-out races. (recent boat speed, recent VMG fitted better but the wrong way round, e.g. faster boats doing worse, so they measure something other than what their name says.) So a team's penalty record over its last 20 races is part of every forecast; the others are shown as context only.

SailGP is very random: even the best boats win only about a quarter of races, and one final race decides each event. Expect favourites to lose often.

It doesn't know about boat damage, injuries, new wing designs or crew changes other than the driver until they show up in results.

Weather records are coarse, so a race's conditions are an estimate, and races within a day share the same conditions.

Of the SailGP Insights data only the penalty record is a model input: boat speed, VMG and time foiling didn't improve the forecast when tested. Insights has no data before 2023.

No historical betting odds exist for SailGP, so we can't yet check the model against the market.