How team races are won, and how that moves with the ratings

Every match ends with one side holding three of the six places. Ten of the twenty possible sets win. This page shows which sets teams win through, how the mix moves as a side gets stronger or its strength is concentrated in one boat, and what the breeze and the level do to it.

Plays as team racers name them: Play 1 is 1-2-anything (the two lead boats control the race); Play 2 is 2-3-4, with 2-3-5 the same play held one place looser; Play 4 is 1-4-5 (the leader holds, boats 4 and 5 keep the last opponent back). The 1-3 sets (1-3-4, 1-3-5, 1-3-6) win but are unstable: the usual aim is to convert them into Play 1 by passing back the opponent in 2nd. Losing sets are labelled by the play the opponent won through. WindCheck: know your combination · BUSA ten-minute guide.

1. The mix as one side gets stronger

Left: the real matches, binned by the gap in forecast strength between side X and side Y (sum of three boat values). Right: the multinomial model's prediction for the same gap with three even boats a side. Each band is one finish set for X; the blue-green bands win, the warm bands lose. The line is the share of matches X won.

2. Movement inside the line-up

Start from two even sides and change one thing. The trace shows every step of the sweep; the marker and the bars show where you are. Press play to watch it move.

3. Does concentrating the strength matter? The data says barely

Observed matches with near-equal totals (gap under 0.75), binned by how unevenly side X's strength is spread across its three boats. If a star with two weak boats won differently, the bands would move left to right.

4. Breeze and level

Winning sets only, as a share of all wins. Breeze is the scorer's word for the day (one word per event). Level is thirds by the mean strength of the two sides.

5. Predicting the whole finish set

Trained on seasons before spring 2025, scored on spring 2025 and 2026: negative log-likelihood of the actual set per match (nats, lower is better), each match seen from both sides. Win log-loss is the implied probability that the side won; shape is the likelihood of the set given who won.

Notes

• Boat values are the pre-event forecast values, so nothing here looks ahead. Only regattas forecast from 1,000 or more prior matches are used.
• The model is the multinomial of section 5 (the small network scores the same and is harder to audit); its inputs are the six sorted values, the gap, each side's spread, best-vs-best, worst-vs-worst, the school term and home flags.
• Breeze is an event-level word; heavy-air events are few and cluster at a handful of venues.
• Spread adds nothing to the win model: held-out log-loss 0.4358 with or without it; the more lopsided side wins 54% of equal-total matches (526 held out).
• Source: analysis/team_race/sequence_model.py, export_combinations.py.