A self-built statistics project on 26 seasons of college sailing
Ranking college sailors with a model that actually gets validated
Existing sailing rankings are mostly eyeballed win totals. This project fits
a real statistical model — a temporal Plackett–Luce
ranking — to 1.5M+ scraped boat-race results, and checks it against races it never saw.
Quinn Brighton · Applied Math–CS, Brown University · Class of 2027EmailLinkedInGitHub
100K+
Training races
24K+
Sailors modeled
13.7M
Parameters
70%
Held‑out accuracy
This week's predictions
Live output from the model above — win probability from 5,000 Monte Carlo race simulations per regatta, built on the predicted lineups. Check back after results post to see how it did.
Quinn's race this week
Match Race NE Champs
Maine Maritime Academy · 10/03/2026 · 8 schools · 5,000 simulated races
1Brown University20% win
2Dartmouth College16% win
3Yale University16% win
4U. S. Coast Guard Academy14% win
5Bowdoin College12% win
6Tufts University10% win
Rose Bowl
USC · 01/02/2027
Boston College — 82% win
WACC Round 1B
Bowdoin College · 10/03/2026
2
27%
Brown University
1
41%
Harvard University
3
17%
Roger Williams University
WACC Round 1A
Dartmouth College · 10/03/2026
2
21%
Tufts University
1
42%
Yale University
3
15%
Dartmouth College
Danmark Trophy
U. S. Coast Guard Academy · 10/03/2026
2
16%
Harvard University
1
24%
Brown University
3
15%
Boston College
Smith Trophy
Massachusetts Institute of Technology · 10/03/2026
Real output from the live model — weekly top‑3‑sailor average rating
for four ICSA programs, June 2024–April 2026. Ratings drift gradually
week to week instead of swinging on a single regatta, because that drift is
exactly what the model's smoothness prior constrains.
How it works
Every sailor gets a latent skill value θ[sailor, week]
for every ISO week they've been active — a full trajectory, not one static number.
It's fit by maximum likelihood under the Plackett–Luce distribution (the standard
way to model a full finishing order, not just win/loss), with an L2 anchor plus
log–cosh penalties on week‑to‑week velocity and acceleration so skill can
drift realistically instead of chasing noise — trained in PyTorch on GPU, batched
by fleet size, over roughly 13.7 million parameters.
A second tool, the Boat Speed Analyzer, is built on the same data:
it compares a school's fleet-race performance across two time windows and flags
whether the difference is statistically significant, so a real speed change isn't
confused with normal race-to-race noise.
Validated, not just fit
The model is trained on 100,000+ races across 24,000+
sailors, and checked against a random 1-in-20 sample of regattas held
out before fitting — races it never saw during training.
70%
of the time, the model correctly predicts which of two sailors
finished ahead — on races it never trained on.
Why I built this
College sailing rankings are usually eyeballed or based on raw win totals — I wanted something that could actually be validated against held-out results.
As Club Captain of Brown Sailing, I use this data to make real decisions about lineups and scheduling, not just as a portfolio piece.
It's the project that convinced me I want to work somewhere that rewards building real statistical models against real data.
