Player projections

The method

Two steps. First we project minutes from a player's minutes per game this season, over his last 10 games and over his last 5 (the last 5 count the most), plus home court and the second night of a back-to-back. Then each stat is his per-minute rate (season, last 10 and last 5) times those minutes, adjusted for the opponent and the pace.

The opponent adjustment compares how many points, rebounds, assists and threes the opponent allows per 100 possessions to his position group (guards, forwards or centers) with the league. Early in a season it leans on last season's number. An opponent that allows 10% more points to his position adds about 2.1% to his points. A game 3% faster than average adds about 5.5%.

The weights were fit by least squares on 24,874 player games from 2025-26. The 80% ranges come from how far real results landed from the projection in testing, grouped by projected minutes.

How we tested it

We replayed 2025-26 from Mon, Dec 1 to Sun, Apr 12, refitting on the first of each month using only earlier games, and projected 19,968 player games. The average miss on points was 4.58, against 4.69 for his season average and 4.66 for his last-10 average. The gain is small, but it held for every stat, and 80% of real point totals landed inside the 80% range.

A gradient boosting model on the same inputs missed by 4.65 points, so we kept the simpler model. It runs inside our database every morning.

What it cannot see

Injuries, rest days, who starts and trades after the morning update. It projects every active player as if he plays, and it is graded only on players who did. Minutes are the biggest source of error: had we known each player's real minutes, the average miss on points would have been 3.74 instead of 4.58. Blowouts cut starters' minutes and are not predicted.

Until a player has 5 games in a season, his projection uses last season's per-minute numbers and his expected minutes on his current roster (last season's minutes handed out down the depth chart). That early-season mode could not be tested, because our data starts in 2025-26. Rookies without NBA games get no projection, and neither do players with fewer than 20 games last season (a few games is too small a sample) until they play this season (Decision #87).

Updated every morning for the next seven days of games. Once a game tips off its projections are frozen, and those frozen numbers are what we grade.