Calibration · 2025-26
dots on the diagonal = estimates mean what they sayRecency-weighted, smoothed (used on the site)Bootstrap of 3-game averages
Every player-game of the last full season was replayed in order: before each game, each method estimated the chance of going over several lines using only earlier games, then we checked what happened. A well-calibrated method’s 70% calls come true about 70% of the time.
Lines in the test are synthetic (around each player’s running average), not bookmaker lines. Past calibration does not guarantee future results. Estimates are for information and entertainment only — not betting advice.
| Method | Brier score | Log loss | Estimates |
|---|---|---|---|
| Recency-weighted, smoothed (used on the site) | 0.1928 | 0.5691 | 438,145 |
| Recency-weighted | 0.1933 | 0.5734 | 438,145 |
| Plain share of past games | 0.1944 | 0.5750 | 438,145 |
| Bootstrap of 3-game averages | 0.2058 | 0.6452 | 438,145 |