Home court advantage in college basketball

By George Boyle · Published 2026-09-26 · 6 min read

Our projections use one home-court number for every true home game. A look at where that number goes wrong turned up a real pattern — and a model built to fix it that still didn't make the cut, for an honest reason worth explaining.

Bias, Nov–Dec (pinned model)−1.57 ptshome teams underrated early
Bias, Jan 1 on (pinned model)+1.33 ptshome teams overrated later
Conference home edge (fitted)+2.2 pts/100 possvs. +4.7 non-conference
Crowd effect (fitted)+1.3 pts/100 possper SD of pregame attendance

A bias that flips sign with the calendar

The single home-court number our projections use — worth 2.877 points per 100 possessions to whichever team is truly at home — is a season-long average. Splitting the held-out 2025-26 season by month shows it isn't a great average for either half: in November and December the model underrates home teams by about a point and a half, and from January on it overrates them by about the same amount in the other direction.

The same pattern shows up on the three training seasons used to build the model, which is what made it worth investigating rather than filing away as one season of noise.

The hypothesis: it's about the crowd

The working theory was that home-court advantage isn't really about being 'at home' — it's about a building being full. November and December carry the sport's non-conference schedule, including 'buy games' where a high-major team hosts a heavy underdog in front of a half-empty arena; conference play, later in the season, reliably fills seats.

Splitting the training data by game type instead of by calendar month makes the pattern explicit: home underdogs in November and December non-conference games are underrated by nearly two points, while conference home games — in any month — are close to unbiased.

Model bias by game type, training seasons only (positive = model overrates the home side)
WindowTrue home, conferenceTrue home, non-conferenceNeutral site
Nov–Dec+1.00 pts (n 646)−1.93 pts (n 4,773)−0.87 pts (n 1,178)
Jan 1 on+0.98 pts (n 9,357)−0.22 pts (n 181)−0.90 pts (n 1,063)

Six ways to model home court, tested on training data only

Six versions of the home-court term were compared, from the single constant currently in production to a version with a separate number for conference games, non-conference games and neutral sites, to versions that add a crowd-size term (the standardized log of a team's average attendance in its home games so far that season, falling back to last season and then the league median for a team with none yet). Every candidate was scored on the same training seasons the production model was fit on, before the 2025-26 season was looked at — using the fitting objective the production model itself was chosen by (the average of spread error and total error).

Training-season results, 2022-23 through 2024-25
VersionFitting objective (lower = better)Nov–Dec biasJan+ bias
Pinned (production)22.2533−1.45 pts+0.77 pts
By game type (conference / non-conference / neutral)22.2740+0.07 pts+0.05 pts
By calendar month22.2721−0.11 pts−0.07 pts
Crowd size only22.2458−1.45 pts+0.72 pts
Game type + crowd size22.2515+0.07 pts+0.02 pts
Game type + a non-conference "buy game" stretch22.2711+0.65 pts+0.06 pts
None of the versions that removed the bias improved the fitting objective by the margin needed to replace the production model — a bar set before any candidate was run. A roughly one-point bias, it turns out, barely moves an error measure on games that miss by eleven points on average: objective and bias are not the same test, and a rule built for one doesn't catch problems in the other.

A second rule, and the one held-out look

With sign-off to write a second rule aimed specifically at the bias, the winner was picked on bias alone instead: qualify within three-quarters of a point in both halves of the season, then take the most accurate qualifier. That picked the game-type-plus-crowd version — conference home worth 2.2 points per 100 possessions, non-conference 4.7, a neutral-site home designation 0.9, plus a crowd term worth about 1.3 points per 100 possessions per standard deviation of attendance. That model, and only that model, then got the one look at the 2025-26 season the pre-registration allows.

The one held-out look, 2025-26 (n = 5,736 graded games with a Pinnacle close)
ModelSpread errorTotal errorNov–Dec biasJan+ bias
Pinned (production)9.322 pts13.757 pts−1.57 pts+1.33 pts
Game type + crowd9.362 pts13.757 pts (no better)−0.15 pts+0.47 pts
Verdict: the pinned model stayed in production. The game-type-plus-crowd version fixed the bias — both halves of the season landed inside the target — but its held-out spread error was slightly worse and its total error was no better, and the rule required the replacement to be no worse on both. A version that reads better on paper still has to clear the same bar the pinned one already cleared.

What held up anyway

The crowd effect itself is real and did not go away when the model that used it lost: home teams gain about 1.3 points per 100 possessions for every standard deviation of pregame attendance, a stable number across every version that included it. Conference home games really are worth less than non-conference ones in the fitted numbers — roughly half as much. Both findings describe the sport honestly; neither was strong enough, on its own, to earn a place in the live projection.

Method

Every candidate kept all of the pinned model's other constants and only changed how the home-court term was computed. Coefficients were estimated the same way the pinned home-court number was: an unregularized term in each training season's full-season ridge fit, averaged over 2022-23 through 2024-25. The crowd term uses only games played before the one being projected — never the game's own attendance.

The selection rules and their thresholds were written down before any candidate was run. The second rule was written after seeing the first rule's training results but before the held-out season was looked at a second time — a disclosed exception, approved in advance, that still only allowed the one held-out look this whole amendment gets.

What this does not claim

This is not a claim that the pinned model has no home-court bias — the split above shows exactly where it does. It is a record of a specific, disciplined attempt to fix it that came up short of its own bar, and of why a visibly smaller bias isn't, by itself, a reason to change a live projection. If the pinned model is still live after 1,000 graded 2026-27 games, its bias by part of season gets reported again, honestly, whatever it says.

Written by George Boyle, who builds The Sport Stack — the models, the ratings and these write-ups. Corrections and questions: hello@thesportstack.io. Who runs this.

See it live