Part 8 of a series on the NBA’s fading home-court advantage; start with Home Court Advantage Is Fading.
The central claim of this series is easy to state and easy to get wrong: NBA home court advantage has faded, from about 65% to 55% of regular-season games and from about 68% to 58% in the playoffs, and a handful of real-world changes narrowed it. A finding like that is easy to believe, which is exactly the problem. Pick the right years, fit a model to a story you already believe, run enough tests, and something will always look real. So every headline claim in these articles was put through a battery of checks built to make it fail, not just the top-line number that the advantage fell. The same tests defend the account behind it: which of the four factors drove the decline, and how much each one did and didn’t. They held, with a handful of adverse or complicating results down there with the rest rather than quietly dropped. Here is what was tried, in plain terms.
The explanation was frozen on old seasons and made to predict the rest. A fair worry about any after-the-fact story is that it was fit to the whole history, so of course it lines up. To test that, the four-factor box-score model was trained only on seasons through 2013, then used unchanged to predict each later season’s home win rate. Its error on seasons it had never seen is about 1 percentage point in the regular season, well under the 1.45 of a straight-line extrapolation and against more than 5 for a flat guess, and it even catches the 2021 dip the trend line misses. The same frozen model beats both baselines in the playoffs too, though on a fifteenth as many games it gets the level much closer than the slope. The mechanism predicts data it was not built from.
📊 Actual playoff home win % across all seasons with the four-factor model forecast and a…
The decline was re-fit every way the defensible choices allow. Every analysis makes calls that could have gone otherwise: where to start the clock, which model to fit, what to do with the pandemic seasons. So the regular-season slope was re-estimated under every combination of those three, 12 fits in all, and every single one still shows a decline. They run from -0.25 to -0.16 points of home win rate per year, around a median of -0.23. Only one of the three calls moves the number much. Starting the clock in 1995 rather than 1984 clips off the steepest stretch of the fall and gives the shallowest slopes in the set. That is worth knowing when reading any single figure here, and it is also the point: no combination of these choices makes the decline go away.
The slowdown was dated two ways, and both agreed. The decline bent, gently, in the late 1990s, then kept drifting. A bend like that can be an illusion of where you draw the line, so it was located with two methods built on different math: one that checks every possible break year for the sharpest change, and one that weighs how many bends the record needs. Their single-bend answers match: 1999, inside a range of roughly 1992 to 2003. A third check, one that watches the trend for drift as seasons pile up, never leaves its stability bounds, which is what a gentle bend rather than a sudden jump would look like: that check is built to catch level jumps and has little power against a slope that merely eases. When independent methods agree, the bend is in the data, not in the choice of line.
The second method has one more thing to say, and it complicates the picture rather than tidying it. Offered up to three bends, it does not think one is enough: a single bend gets only 19.3% of its confidence, with almost all of the rest going to two or three. Both of those richer fits then put their most recent bend in the same season, the one ending in 2020, when the pandemic cut the schedule short and finished the playoffs on a neutral floor. Hold that one loosely, and hold the single-bend year loosely too: neither richer fit’s earlier break lands on it either, so agreement on 1999 depends on restricting the model to one bend. The pandemic seasons are unlike anything else in four decades of the record, and a method hunting for bends will find one at a stretch that strange whether or not it marks a lasting change. The specification curve above is the answer to whether it matters: dropping those seasons outright barely moves the slope.
All of this is the regular season: each postseason rests on so few games that the break test cannot single out a year in it at all. This is the shape Home Court Advantage Is Fading describes.
Fake rule changes were planted where nothing happened. A favorite way to fool yourself is to find an effect at a moment when nothing actually changed. So the same test was re-run at every year from 1987 to 2010, to see whether the method would invent an effect where the story says there was none. The raw test does light up several years clustered around the real 1994–95 break, which is exactly what a genuine one-time shift does to the seasons leading up to it, and a second run of years in the 2000s, which the post-1999 slope moderation produces rather than any rule change. Ask instead which era boundary shows a one-time level shift once every era is in the model at once, and only 1994–95 survives, a level shift of about 2.6 points: the real hand-checking crackdown. That is a regular-season result; the playoffs have too few games to show any era step at all. That is the fingerprint Home Court Advantage Is Fading dates the first drop by, and the reason The Usual Suspects can clear every other rule change.
The breakdown held up under a very different model. The split of the decline across the four factors, the one The Three-Point Suspect lays out, comes from a simple model that adds the four up in a straight line. Would a method that assumes nothing about how they add up blame the factors differently? A flexible model, one that lets the factors bend and feed off each other instead of adding up cleanly, re-split the same decline, and it agreed on three of the four: rebounding, turnovers, and fouls all landed within a few points of their earlier shares. That is a consistency check the breakdown passed, not a second, independent measurement: both methods read the same four box-score edges off the same games, so they can agree without either one confirming the other. Shooting is the one they split on: the flexible model gives it 34% of the decline against 21% before, which puts shooting at the top of its own ranking. That ordering is not a rival verdict to argue down: the extra credit it hands shooting is credit for a fade the three-point shift already explains. Rebounding is the largest driver on the straight-line split, with a cause of its own, the lead The Mystery on the Glass rests on. The only thing to hold loosely is shooting’s exact share. The playoffs are where the two methods part company hardest: the flexible model puts shooting first at 38% of the decline, with rebounding close behind, where the straight-line breakdown gives shooting only 12% and can’t tell that share from zero. Both call rebounding large, and on a fifteenth as many games neither method’s share is pinned down.
📊 How the home-court decline splits across the four box-score factors when a flexible win…
The word “largest” was put on the clock. Calling rebounding the biggest single driver is a ranking, and a ranking only has to slip once for the headline to change. So the whole breakdown was re-run over and over on re-drawn seasons, counting which factor came out on top each time. Rebounding wins 76% of those redraws, turnovers 21%, and shooting, the factor its own flexible-model ranking puts first, just 2%. In the playoffs, on a fifteenth as many games, rebounding leads 64% of them. So the lead is real and it is not a runaway: roughly one redraw in four hands the title to a different factor, and almost every time that factor is turnovers, the one The Mystery on the Glass already treats as rebounding’s near-equal.
The drivers were tested against a hidden cause. Maybe a factor is just standing in for something the box score never recorded. There is a standard way to ask how strong such a hidden cause would have to be to explain a link away. For the shooting link that The Three-Point Suspect rests on, a hidden cause would have to explain at least 60.5% of everything left unexplained in both that gap and who wins; for turnovers, at least 40.7%; for rebounding, at least 36.8%; for the foul link, the lowest of the four, at least 28.9%. Those are demanding thresholds, so the strongest links are hard to wish away.
A test that could have sunk the three-point case was run anyway, and it came back against us. If the move to the arc were pushing home court down season by season, then a year when threes jumped ought to help predict the drop that followed. It doesn’t. Across 42 year-over-year changes in the regular season, last season’s three-point rate adds nothing to a forecast of this season’s home win rate beyond what home court’s own history already supplies. That result is here because it points away from The Three-Point Suspect, not in spite of it: a checks article that only lists the tests that passed is not a checks article. It is also weak evidence, and for a reason worth naming: four decades give only a few dozen seasons, which is a very short series for a test of this kind, so it cannot tell “no effect” apart from “not enough seasons to see one.” The three-point case does not rest on it either way. That case rests on comparing high-three-point games against low-three-point games within the same rules era, tens of thousands of them, rather than one dot per season.
Running many tests was corrected for. Test enough things and a few will look real by chance alone. After the standard correction for having run a whole battery of tests at once, the central results clear the bar: both decline trends, the within-era three-point effect, rest, altitude, and the 1994–95 era shift. What drops out is what this series already treats as too small or too unclear to lean on: time zones, pace, and the playoff era step. The central findings are not the lucky few.
It wasn’t just which teams happened to host. The era decline could be a trick of composition, if the stronger home teams happened to host more in the early years. Even when the model is told exactly which teams are playing (home and away team identity), the era-by-era decline barely moves: the biggest change in any era’s estimated size is 0.5 pp. It is not an artifact of which teams played at home in which decade, which is why the fade reads as league-wide rather than the work of a few franchises.
None of these checks proves a mechanism on its own, and where the data can only propose a cause, these articles say so. What they establish is that the decline is real, its regular-season slowdown is real, though exactly where to date it depends on how many bends the model is allowed, and its main drivers are not artifacts of how the numbers were sliced. For every test above with its full numbers and the range each result could shift within, see The Investigation. If you would rather start from a sentence you doubt than from a test, Every Claim, and What Backs It lists every checked claim these articles make and the number sitting under each one. It also records that each of those claims has been made to fail on a deliberately wrong number, so none of them rests on a check that passes no matter what.
That closes the case.
Back to the series hub: The Disappearing Home-Court Advantage
How this was made: the writing here is mostly AI/LLMs working from my analysis and direction; the numbers are all Python on public NBA data. Every claim is backed up by data. The data, analysis, and conclusions are trustworthy. The full note is on the series hub.
Data: NBA.com Stats, pulled with the open-source nba_api tool; attendance records from Basketball-Reference.