October 8, 2026 · data through October 4, 2026 · All notes

Which venues get orderly in a breeze?

Sort 30 venues by how results respond to wind after the fleet is taken out: river venues get orderly in a breeze, a few open-water ones get messier, the Chesapeake likes a southerly. Small effects, carefully isolated.

Some venues get more orderly when it blows, a few get less, most barely care — and at a handful of venues the favourites are reliably over- or under-rated. Fleet first, then wind, then venue; small effects, carefully isolated.

Surprisal and temperature

A boat's surprisal is the derivative of the finishing order's log-likelihood with respect to its rating (score g over information h): how much better it did than its rating predicted. An event's temperature β is the same thing for the whole fleet — refit the ratings' scale to the event; β < 1 is a surprising event, > 1 an orderly one. Temperatures vary more than noise allows (mean s²/I = 1.43; true spread 0.21) and persist within a day (odd vs even races r = +0.47).

The wind labels

Scorers' daily summaries, scraped and parsed by a language model: 8,520 regattas, 7,580 regatta-days with a stated wind speed and 8,704 with a direction. Reanalysis wind fills the gaps after calibration (r = 0.62; stated ≈ 0.81 × reanalysis + 3.7).

The fleet first

What a day's entry list says about its orderliness, before any wind (CV R² = 0.16). Stronger fleets are more surprising; a standout top boat makes an event orderly; newcomers make it unpredictable.

Everything below is the residual: temperature minus what the fleet predicted. Once the fleet is out, a venue-only model has no held-out skill (R² = -0.027), and removing skippers' individual breeze preferences changes the residuals by nothing (r = 0.992; no out-of-sample gain, log-loss -0.00003).

Then the wind

By stated wind speed, 3,425 labelled days: breeze helps, about +0.05 per 10 knots.

By direction, venue-demeaned (se ≈ ±0.007): onshore sea-breeze directions orderly, the post-frontal north-westerly surprising.

Venue	Days	p	Most orderly sector	Least
UC Santa Barbara	46	0.002	S +0.18	NW -0.02
Rhode Island	47	0.003	W +0.32	SE -0.14
Navy	114	0.010	S +0.13	NW -0.06
Brown	78	0.049	S +0.10	NW -0.19
Eckerd	55	0.049	S +0.10	E +0.01
Mass Maritime	46	0.127	NE +0.09	NW -0.14
Yale	69	0.128	S +0.25	SW +0.01
Charleston	69	0.155	W +0.15	NW -0.14

Direction effects per venue, permutation-tested: 5 of 22 venues beyond p < 0.05 where one would be by chance. Navy (114 days) is the robust case.

Favourites' venues and underdogs' venues

Mean residual orderliness per venue (30+ labelled days). Yale and Wisconsin reward the favourites beyond what their fleets explain; at Connecticut College and Cornell the fleet compresses. Six of 30 venues sit beyond two standard errors (1.5 expected).

Clustering the venues

Venues on the first two principal components of their wind-response profiles, coloured by cluster

Each venue's profile (overall residual, eight direction sectors, four speed bands, shrunk) on its first two principal components; Ward clusters. PC1 (24%) is the speed response, left "breeze brings order" to right "breeze brings chaos"; direction is PC2 (14%).

Cluster profiles by wind-speed band.

Group	Venues	Slope / 10 kn	By speed band	Members
breeze helps	13	+0.15	<6kn: -0.03, 6-10: -0.04, 10-14: +0.04, 14+: +0.04	South Florida (75), Cornell (70), Charleston (69), Harvard (59), Coast Guard (57), Boston University (56), Connecticut College (51), Mass Maritime (46), Salve Regina (43), Wisconsin (41), Dartmouth (34), NY Maritime (32), Hobart & William Sm (30)
wind barely matters	5	-0.03	<6kn: +0.04, 6-10: -0.02, 10-14: -0.02, 14+: +0.01	MIT (112), Old Dominion (76), Tufts (60), Tulane (38), Bowdoin (33)
light air orderly, S clean	8	-0.07	<6kn: +0.06, 6-10: -0.02, 10-14: +0.01, 14+: -0.02	Navy (114), Kings Point (80), Brown (78), St. Mary's (70), Yale (69), Eckerd (55), Stanford (48), Norfolk Yacht and Country Club (32)
breeze hurts	3	-0.13	<6kn: +0.04, 6-10: +0.03, 10-14: -0.06, 14+: -0.01	Rhode Island (47), UC Santa Barbara (46), Summit (33)
breeze hurts (Washington)	1	-0.17	<6kn: -0.01, 6-10: +0.02, 10-14: +0.04, 14+: -0.21	Washington (37)

The two ends are dependable; the middle reshuffles when inputs change and should not be read as assigned.

Is it a good day to be an underdog? Explore every venue's map, two at a time →

Navy, Brown, Yale, Tufts

Navy: residual orderliness by wind direction and speed

Navy, 114 days, fleet removed: a southerly up the Chesapeake is the favourites' breeze, a north-westerly off the Academy the underdogs'. Left: nearby-day average; right: fitted surface.

Brown: residual orderliness by wind direction and speed

Brown, 78 days: the breezy westerly corner is the messy one.

Yale (+0.091, 69 days, 67% above zero) and Tufts (-0.049, 60 days), 150 km apart, same conferences, opposite ends of the scale. Compare them interactively →

Validation, in one list

Every model cross-validated with whole events held out; venue direction tested by within-venue permutation; the clustering rerun on breeze-adjusted residuals (ends survive, middle reshuffles); stated winds checked against reanalysis (0.62).
Caveats: venue cells hold 3–20 days; race times are approximate; the wind effects are a few hundredths of temperature against a daily spread of 0.21, and none improve race forecasts.

Every number on this page is produced by analysis/blog/venue_post_data.py, orderliness_wind.py, venue_clusters.py and the wind pipeline in the project repository, from the same database that powers the rest of the site. Corrections welcome: quinnbrighton2005@gmail.com.