Everyone else sells you patterns. We test them against 281,192 real player-games across 11 seasons (2015–present) and show the receipts — real, not real, or not yet knowable.
281,192
player-games analyzed
8
beliefs put on trial
6
turned out false
1
not yet knowable
Signal or Noise
Does any of it actually repeat?
For a player “trait” to be real, his edge has to carry from one season to the next. Here's how much five popular ones actually do — on a scale from pure chance to a rock-solid skill.
A real, repeating skill →
“Hot hand” momentum
“Owns this matchup”
“Wears down on back-to-backs”
“Road warrior / home hero”
“Clutch gene”
Five separate beliefs. All five huddle in the noise. Not one of these “traits” carries season to season the way a real skill would — the whole left-hand cluster is, statistically, a coin flip.
Honesty note: at a quarter-million games, a few of these correlations turn statistically nonzero — but at r ≤ 0.09 they remain useless for any single prediction. Detectable isn't the same as usable.
Show
Showing 8 of 8 verdicts.
Real
The signal+1.22 shots · t = 33.4robust
Torchbearer
“When the top option sits, his teammates get more shots.”
Real, and mechanical — the top option's shots get redistributed to whoever's left.
When a team's top-usage player sits, teammates take +1.22 more field-goal attempts per 36 minutes at flat efficiency — and it survives the injury-wave confound, still holding in clean solo-absence games (+1.20, t = 8.8).
Extra shot attempts when the star sits.
per 36 min · bars show the 95% interval
Both clear zero with room to spare — the effect survives the injury-wave confound. What we don't show: a per-player “who benefits most” ranking. A single pair's 95% interval (±2.0 shots) is wider than the effect, so that number would be fiction.
Not real
The signal0.02 / 1.0 repeatabilitynoise
Heat Check
“He's heating up — ride the hot hand.”
Not real — a scoring streak tells you almost nothing his last game didn't.
A five-game hot streak predicts an above-average next game just 47.6% of the time — worse than simply asking 'was his last game good?' Once you know that, the streak adds ~0.05% of the variance.
Why people believe itScoring is noisy and right-skewed, so ordinary randomness throws off streaky-looking runs that memory turns into trends.
A real 15-game stretch vs a pure coin-flip with the same average.
Sequence
his actual games, 2026-03-16 to 2026-04-10
Real data: Kevin Durant's actual last 15 representative games of 2025-26, against his real season average (26) and game-to-game noise (±6.7 pts over 78 games). A ±1.6σ band contains ~89% of games, so 1–2 of any 15 stray outside it by definition. The coin-flip sequence is explicitly synthetic — same average, same noise, no basketball — and it streaks anyway. In our tests a five-game scoring streak predicts an above-average next game 47.6% of the time and adds ~0.05% once you know his last game. Streaks say where he's been, not where he's going.
Not real
The signal0.02 / 1.0 repeatabilitynoise
Nemesis Index
“He owns this matchup — he always torches them.”
Not real — 'owning' a team one season doesn't carry to the next.
72% of player-opponent pairs have two games or fewer, and whether a player 'owns' a team barely carries season to season. Across 44,000 pairs it measures r = 0.02 — detectable at that scale, but 0.06% of the variance: real, and useless for calling any single matchup.
Why people believe itOne memorable night against a rival sticks in the mind, and small samples naturally produce wide spreads that look like a signal.
Season-one edge vs season-two edge — a repeat would be a diagonal.
Show
r = +0.02no relationship — last season's “ownership” tells you nothing about the next
Shown: a random 60 of all 44,198 player-opponent season-pairs (2015–present). The r is computed on the full population, not the sample.
Real data, both views: matchup edge = points vs a specific opponent minus own season average, for every player-opponent pair with 2+ meetings in consecutive seasons across 11 seasons (2015–present) — 44,198 pairs, r = +0.02. The comparison trait is the same players' scoring averages, season vs next (3,186 season-pairs, r = +0.88) — that diagonal is what a real, stable quality looks like. "Owning a team" doesn't make that shape.
Not real
The signal0.06 / 1.0 repeatabilitynoise
Fatigue Shadow
“Some guys just wear down on back-to-backs.”
Not real as a personal trait — back-to-back scoring holds up league-wide.
How an individual responds to a back-to-back barely repeats season to season (0.06 — noise). League-wide there's only a small efficiency dip, and it's smaller than the folklore claims.
“He's a road warrior — or he only shows up at home.”
Not real as a trait — the real home edge is efficiency, and it belongs to no one in particular.
Individual home/road splits are ~96% noise season to season (0.04). The shared, league-level truth is small: about +1.02pp of true shooting at home, with volume and minutes essentially flat.
The signalelevation n.s. · team quality r = −0.49confounded
High Ground Tax
“The altitude in Denver and Utah wears visiting teams down.”
Not real — how hard an arena plays follows the home team's quality; elevation itself predicts nothing.
Across all 30 arenas and 11 seasons, how much visitors score tracks the home team's quality, not its elevation — team win% is the dominant correlate (r = −0.49), while elevation itself explains nothing (r = −0.22, not significant). Denver, the true 5,280-ft outlier, shows no altitude effect in any season.
Why people believe itAltitude's effect on endurance is genuinely real in sports science — the clearest proof that outside research generates hypotheses while only our own data issues verdicts.
All 30 arenas: visitor scoring against elevation, then against home-team quality.
Explain visitor scoring by
vs home win% · r = -0.49 | vs elevation · r = -0.22 (n.s.)
no relationship — across 11 seasons, elevation predicts nothing about visitor scoring (r = −0.22, not significant)
one dot = one arena (30)ringed: the arenas the folklore is about
Tested on visitor scoring: every visiting player's points vs his own sea-level road average, 2015-16 through 2025-26 (within 281,192 player-games played; ~4,200 per arena). Denver — the real 5,280 ft outlier — shows no altitude effect in any season. Which arena looks “toughest” shifts era to era with team quality; elevation stays flat throughout.
Not real
The signal0.01 / 1.0 skill-controlled repeatabilitynoise
Clutch Gene
“He's got the clutch gene — he's a different player when the game is on the line.”
Not real — remove ordinary scoring skill and clutch over-performance repeats at 0.01, the same noise as the hot hand.
Clutch shooting minus a player's own baseline appears to repeat at 0.21 season to season — but that is exactly what a world with NO clutch gene produces: clutch TS% rests on so few attempts that subtracting full-game skill over-corrects and smuggles skill back in. Control for scoring skill properly and clutch-specific repeatability is 0.01 across 1,176 consecutive-season pairs. The 'most clutch' seasons on record belong to role players on tiny samples — Okogie, Sefolosha, Thaddeus Young — and the top-15 cohort collapses from +3.45 to −0.13 the next year. What IS stable is who TAKES the clutch shots (0.39) — the closer's role repeats; making them doesn't.
Why people believe itGame-winners are the most replayed moments in the sport, and a handful of makes on a tiny sample is all memory needs to issue a permanent label.
Untestable
The signalnot in the box scoreuntestable
Gravity
“He bends the defense — his gravity makes everyone around him better.”
Untestable with box scores — gravity is spatial, and a box score is the wrong instrument.
Gravity is about where defenders stand and who they leave open — none of which a box score records. Our indirect read through teammate shooting ran backwards, consistently across both seasons.
What would settle itPlayer-tracking data — positions, off-ball movement, and defensive matchups.