The shape of normal. Most NBA games are decided by single digits. A handful are decided by 50+. Here's the histogram.
Every distribution has tails. These are the tails.
Every game of every team, broken out by season. Wins green, losses red, read left-to-right. Win-loss record is on the right. The number after that — the one that matters most — is point differential per game. That's the closest thing to a true talent metric you can compute from final scores alone.
Ranked by surprise. The algorithmic detector found these patterns by scanning every game in the dataset for streaks, blowouts, nail-biters, and statistical anomalies.
The v2 ratchet model — the same one that beat baseline by 45% Brier on a held-out test set — applied to upcoming NBA games. Track record updates as games complete. Picks are reasoned, not vibes. Confidence is honest.
The ratchet loop: hypothesize, test, keep or revert. Each iteration is a new rule added on top of the previous. Every version is scored on a TEST set the model has never seen — games from 2024-25 and 2025-26. No future leakage, no overfitting. Bootstrap confidence intervals, not point estimates.
Six leagues, hundreds of athletes, the leaders in every category we track. Volume scorers, efficiency monsters, defensive anchors, the overlooked specialists. Each row is a real player having a real season.