Play 1, Play 2, Play 4: how team races are won
Winning combinations named the way team racers name them, across 29,869 college matches: Play 1 is most of team racing; college winners hold first less in a breeze and high school winners sweep more; a star with two weak boats wins the same way as three even boats; and six ratings predict the finish set a tenth of a nat better than the win probability alone.
Team racers name their winning combinations: Play 1 is 1-2-anything, Play 2 is 2-3-4, Play 4 is 1-4-5. Here is how often each actually happens, how that shifts with the breeze and the level, whether a star with two weak boats wins differently (no), and whether a model can predict the whole finish set from six ratings (somewhat).
The twenty sets
A side holds three of six places; there are twenty ways to do that and ten of them win (places summing to ten or less). Team racers group the winning ones: Play 1, 1-2-x, where the two lead boats control the race; Play 2, 2-3-4, with 2-3-5 the same play held one place looser; Play 4, 1-4-5, where the leader holds and the fourth and fifth boats keep the last opponent back; and the 1-3 sets, which win but are unstable, the usual aim being to pass back the opponent in second and convert them into Play 1. 29,869 ICSA matches with full line-ups and a pre-event value for each of the six boats.
Play 1 is most of team racing: the winner holds first place in 88% of light-air wins and 84% of heavy-air wins. As the breeze builds, college winners hold first less and win through the back-heavy sets more: Play 2 goes from 12% of wins to 16%, Play 4 from 3.2% to 6.3%. Only 631 matches were sailed in what the scorer called heavy air, so the heavy column is the noisy one, but the gradient runs through fresh as well.
By level and breeze
The breeze effect lives in the middle and top thirds. Weak fleets win by 1-2-3 sweeps in any wind; in the middle third, Play 2 climbs from 13% of wins in light air to 17% in fresh air. Strong, even fleets in a breeze produce the fewest Play 1 wins.
High school goes the other way. In a breeze, high school winners sweep 1-2-3 more and the winner holds first in 89% of fresh-air wins. One reading: college teams in a breeze are even enough that the race is decided by mark traps in the pack; high school fleets are uneven enough that breeze simply stretches them out.
Does a star with two weak boats win differently?
No. Within each strength third, the balanced and lopsided halves win through the same mix of plays, to within a point. The spread also adds nothing to the win forecast: held-out log-loss 0.4358 on the sum alone, 0.4358 with the spread added. Among 526 held-out matches where the two sides had nearly equal totals and clearly different spreads, the more lopsided side won 54%. Three boats of 3.0 and one of 4.0 with two of 2.5 are, as far as results can tell, the same side.
Predicting the whole finish set
Can six ratings predict not only who wins but how? A twenty-way multinomial on the six sorted boat values, the gap, each side's spread, best against best, worst against worst, the school term and the home flags, trained on seasons before spring 2025 and scored on spring 2025 and 2026 (6,309 matches, each seen from both sides).
| Model | Finish-set NLL | Implied win log-loss | Shape given the winner |
|---|---|---|---|
| Combination frequencies alone | 2.758 | — | — |
| Production win probability x combination shape | 2.497 | 0.4309 | 2.121 |
| Multinomial on the six boat values | 2.400 | 0.4328 | 2.022 |
| Small network (64 hidden, early stopped) | 2.398 | 0.4331 | 2.020 |
Negative log-likelihood of the actual set per match, in nats, lower is better. The multinomial beats the obvious two-stage baseline, the production win probability times the overall frequency of each set, by about a tenth of a nat; the gain is in the shape of the win, not in who wins (its implied win log-loss is a touch worse than the production forecast's). A small network does no better than the linear model, so there is no hidden non-linearity in six numbers. What the model has learned is simple: the stronger the side, the more its wins are Play 1 and the fewer are 1-3 sets and Play 2.
Caveats
- Breeze is one word per event, from the scorer's daily summary; a two-day regatta with one light day and one fresh day is "moderate". Heavy-air events are few and cluster at a handful of venues.
- Boat values are the pre-event forecast values, so nothing looks ahead; regattas forecast from fewer than 1,000 prior matches are left out.
- These are finish sets, not the plays as sailed. A race that ends 1-2-6 may have been Play 4 until the last beat.
Every number on this page is produced by analysis/team_race/sequence_model.py and export_combinations.py; numbers via analysis/blog/team_race_post_data.py in the project repository, from the same database that powers the rest of the site. Corrections welcome: quinnbrighton2005@gmail.com.