Method
How the ratings are built, how they are tested, and how the data are checked.
The rating: SiPM
SiPM rates a player's effect on his team's scoring margin, in points per 100 possessions. Zero is an average player; offence plus defence equals the net rating.
It has two parts. The first is a lineup rating: a regularised plus-minus over every lineup, with a prior that decays over past seasons and an adjustment for shooting luck. The second is a box-score estimate, adjusted for the strength of the opponents. A final layer combines the two, with weights fitted on how well they predict the next season's lineup results, and adds an age curve that differs by player quality.
Projections
This season. The 2026-27 rating starts from his SiPM rating last season, moved one year forward: where players with his rating, age and minutes stood a season later. After each game, a game-level rating (a fixed map from one game's box line to SiPM's offence and defence) updates a running estimate, and the projection moves with it. Rookies start from an estimate based on draft slot.
Five-season outlook. The mean rating over 2026-27 to 2030-31, with 10th and 90th percentiles, if he plays. It reflects age, current level and how players like him have moved. Tested on 2,330 player forecasts made in 2013-14 to 2020-21 (fitted only on earlier seasons), its typical error is 1.16 points per 100, against 1.51 for assuming his rating stays where it is; the 80% range held 82% of outcomes.
Best season ahead. His expected rating in his best season of the next five, the age he is likely to reach it, and the chance that some later season beats this one. A typical path can be flat while the best season still lies ahead: careers zigzag around the trend. All three are fitted on what happened next to players with the same rating, age and minutes since 2007-08.
Projected stats. Usage and minutes come from a trajectory model that includes a role-change block: who left his team, who arrived, and whether the top option is gone. Shooting and per-75 stats come from a short-term box-score forecast that weights recent games more.
How it is tested
Every score below is out of sample: the forecasts are scored on seasons and dates the model never saw when it was fitted.
Against the leading public models. Ratings from SiPM, DARKO and EPM were frozen on each date and scored on the next 30 days of game margins. SiPM is level with both: every interval includes zero (6 of 6 comparisons).
| Comparison | Dates | Difference90% interval | Reading |
|---|---|---|---|
| SiPM minus DARKO (every game day, 2025-26) | 135 | +0.14 (−0.17 to +0.38) | level |
| SiPM minus DARKO (every week, 2025-26) | 21 | +0.19 (−0.16 to +0.39) | level |
| SiPM minus DARKO (every week, 2024-25) | 21 | −0.25 (−0.56 to +0.12) | level |
| SiPM minus DARKO (every week, 2023-24) | 21 | −0.01 (−0.37 to +0.25) | level |
| SiPM minus EPM (game days with an archived EPM, 2025-26) | 22 | +0.07 (−0.10 to +0.24) | level |
| SiPM minus EPM (weeks with an archived EPM, 2025-26) | 3 | +0.12 (−0.09 to +0.30) | level |
Difference: percentage points of forecast error removed, SiPM minus the other model; above 0 favours SiPM. EPM rows use the copies archived on each date.
Data checks
Every source is audited for errors before use. The audit has traced 76 root causes: 52 in the providers' data and 24 in our own processing. 58 are fixed, 9 excluded, 9 documented, and 0 open. Full audit.
Providers sometimes change data without notice. After our September pull, NBA.com re-issued its 2024-25 player-tracking data on its newer Hawk-Eye definitions. A routine check caught the jump; the 24 affected files were re-pulled and the results built on them re-run. No headline result flipped: the shot-making rating's correlation with the NBA's Hawk-Eye shot-making moved from 0.71 to 0.70.