How Statheum works.
Everyone else sells you patterns. We test them. Statheum takes the stories basketball tells itself — hot hands, revenge games, altitude, clutch genes — and checks them against every box score of the modern NBA. What survives gets called real. What doesn't gets called what it is.
The data
Game-level box scores from the BallDontLie API: 281,192 = player-games played — a player appeared in the game — across 11 seasons, 2015–present. That count is the corpus, and it is the one number every Statheum surface uses. Verdict metrics are then computed on its representative subset (252,420 player-games): appearances of at least max(5, half the player's own median) minutes that season — garbage-time cameos are filtered before anything is computed, because a two-minute appearance isn't evidence of anything.
The window is deliberate. Rule changes and load management make older eras a different sport; testing a modern belief against 1994 data would answer a question nobody asked. The base refreshes as new games land.
How a verdict is earned
Every check runs the same road. A hypothesis comes from folklore — something people actually say. It gets tested league-wide on the full window, never on a cherry-picked slice. Minutes are decomposed from efficiency before we believe any scoring change, because points-per-game mixes playing time with performance. And the bar for calling anything a traitis repeatability: a player's edge has to carry from one season to the next, or it's noise wearing a story.
Every published verdict shows its sample sizes, its intervals, and its limits. The verdicts live in the Hypogeum →
What we refuse to publish
Per-player rankings the data can't resolve. When a single pair's error bar is wider than the effect itself, a “who benefits most” list would be fiction — so we don't print one.
Significance manufactured from big samples. At a quarter-million games some correlations turn statistically nonzero — at r ≤ 0.09 they remain useless for any single prediction. Detectable isn't usable.
Anything phrased as betting advice. Statheum is not a betting product. No odds, no picks, no edges for sale — evidence, not anecdote.
Where it stops
Box scores can't see everything. Clutch-time performance and defensive gravity are real questions the box score is structurally blind to — those verdicts stay UNTESTABLE until better data exists, and we say so rather than guess.