How team race forecasts work
Add up three boats' ratings, a school term and a home edge, then take the logistic of the difference. Held out on two spring seasons the forecast calls 78.8% of matches, and the stated chances match what happened.
A team race forecast adds up what three boats bring, then asks how often that beats what the other three bring. Fleet ratings do most of the work; the school does the rest. Held out on two spring seasons, the forecast calls 78.8% of matches and its probabilities are honest.
The model
Every ICSA team race match since 2013 with a full line-up: 31,338 matches, 6,452 sailors, 130 schools. A side's strength is the sum over its three boats of the skipper's fleet rating (scaled by a fitted 0.72), a small per-sailor team race offset for the skipper and the crew, the school's team race term, and a home-water edge when the school hosts. The probability that side A beats side B is the logistic of the difference. Offsets are shrunk toward zero so a sailor with ten matches cannot acquire a large one; the fit uses a two-year half-life so old matches count less.
P(A beats B) = σ( strength(A) − strength(B) ), strength = Σ3 boats [ 0.72·rating + skipper offset + crew offset ] + school + home
What each piece is worth
Knowing only which schools are racing gets to 0.562 and 70.4% of matches called. The three skippers' fleet ratings, nothing else, get to 0.454 and 77.9%. Adding the school term, the sailor offsets and the home edge brings it to 0.438 and 78.8%. The sailor offsets themselves are tiny: their fitted spread is 0.030 for skippers and 0.026 for crews, against a school spread of 0.22. Individual team race skill beyond the fleet rating is barely identifiable from results; what persists is the programme.
The school term
| School | Team race term | Matches | Win share |
|---|---|---|---|
| Miami | +0.56 ± 0.16 | 520 | 45% |
| St Mary's | +0.48 ± 0.12 | 1,860 | 59% |
| Stanford | +0.46 ± 0.12 | 1,909 | 78% |
| Roger Williams | +0.45 ± 0.08 | 2,994 | 61% |
| Charleston | +0.45 ± 0.10 | 2,014 | 65% |
| Dartmouth | +0.41 ± 0.10 | 1,773 | 55% |
| Boston College | +0.36 ± 0.10 | 2,381 | 64% |
| Tufts | +0.36 ± 0.08 | 3,040 | 51% |
Schools with at least 500 matches, top of 35. The term is what the school adds to the sum of its three boats, in the same units as a rating: St. Mary's, Stanford, Roger Williams and Charleston win more team races than their fleet ratings alone would predict, season after season. A per-season version of the term overfits and hurts the forecast, so the site uses one persistent number per school.
Are the probabilities honest?
| Stated chance for the favourite | Matches | Mean stated | Favourite won |
|---|---|---|---|
| 50–60% | 954 | 54.9% | 53.6% |
| 60–70% | 1,017 | 65.2% | 65.3% |
| 70–80% | 1,086 | 75.1% | 73.5% |
| 80–90% | 1,282 | 85.3% | 86.2% |
| 90–100% | 1,996 | 95.8% | 95.9% |
Held-out matches grouped by the stated chance. The best temperature rescaling is 1.00: the numbers can be read as they are.
What did not help
- Finishing positions as extra training signal. Fitting the six places as a ranking, alongside the win, made the win forecast clearly worse. Team race finishes are tactical; a 2-3-4 says nothing about who was fastest.
- Refitting between rounds of the same regatta. Form within an event adds almost nothing once the pre-event numbers are in.
- Pair and trio familiarity. Line-ups are stable (the median pair has sailed 13 matches together), so there is nothing to learn from.
- High school. The same model on ISSA data reaches 0.493 on 2,138 held-out matches (75.8% called), with crews often missing from the line-ups.
On the site: every team race regatta page has a Ratings & Forecast tab with the pre-event boat values, the forecast for each match and who beat it; sailor pages show a team race rating and a regatta-by-regatta surprisal; the power rankings have a Team Race selector.
Every number on this page is produced by analysis/team_race/model.py, evaluate.py and fit.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.