A projected score for every NBA game, built from each team's points scored and allowed per 100 possessions and its pace, adjusted for who it played and where. Recent games count more, back-to-backs and missing rotation players are priced in, and early in the season each rating leans on a preseason estimate built from the roster. Open any game for the full breakdown.
Projections are model output, not betting advice. Every number here comes from the same model that produced the score; nothing is hand-entered except the market lines.
A projected line for every rotation player: minutes, points, rebounds, assists and threes. Where a sportsbook line was captured, the model gives the chance of the over and of the under, each at the number actually on offer for that side, and leans to the likelier one. Every player is in his game's breakdown above; these are the leans the model is most sure of.
The projection assumes a player who missed his team's last game sits again and that everyone else plays their usual minutes. It has not read the injury report; the sportsbook has. Check who is playing before reading anything into a lean.
Every game that has a market line when the board is built is logged with the line and the side the model leans to. Entries are frozen when first captured and graded when the final score arrives.
Every game from six past seasons was projected using only games played before that day. The settings were chosen on four seasons, using those seasons' closing lines as the target because a line is a far steadier guide than one night's score. The other two seasons were held back and never seen while tuning, so those are the fair measure. Each projection was then compared with the closing line for that game.
Points per 100 possessions against an average team, relative to the league average. On defence negative is good: fewer points allowed. Net is offence minus defence. Returning and Acquired are the shares of this season's minutes the model expected from last year's players and from players who were on another team.
Possessions. Each box score gives possessions as shots taken, minus offensive rebounds, plus turnovers, plus 0.44 for each free throw. Points divided by possessions is efficiency, which takes pace out of the comparison.
Ratings. One regression per day solves for every team's offence, defence and pace at once, so a team is judged against the opponents it actually played. A home-court term is fitted in the same step, with an extra amount in the play-in and playoffs, where home teams have done better.
Early season. Each team starts from a preseason estimate built from its roster. Players are traced by id through last season's box scores, so the estimate knows what share of minutes returned, whether the returners were the team's better players, and how much each player acquired by trade or free agency played and for how good a team. The rating is pulled toward that estimate, and the pull fades as the season goes on.
Who is playing. Each rotation player is valued from this season's box scores. Games a team played short-handed are corrected back to full strength before rating, and a player who missed the team's latest game is expected to sit again, more firmly if he has missed two or more in a row. This starts once a team has played four games.
Rest. A team on the second night of a back-to-back is marked down by a fixed amount, and past games played on no rest are corrected before rating.
Shooting luck. A hot or cold night from three or the line is mostly noise, so part of each game's over- or under-shooting is removed before the game counts toward a rating.
The projection. Expected possessions for the matchup, times each side's expected points per possession against this defence, plus home court. The total carries a small allowance for overtime, larger when the game is projected to be close. The league-wide scoring level and pace follow the last week or two of games, and play-in and playoff games are marked down because they are slower and lower scoring. The win probability treats the final margin as landing around the projection with a spread measured in the backtest.
Player props. A player's line is his expected minutes times his production per minute. Minutes blend his last few games with his longer-run average, and are nudged up when the players expected to dress add up to fewer than 240. Rates use this season and last, recent games counting more, pulled toward the league rate until he has played enough. Scoring stats then move with the points the game model projects for his team, rebounds with projected pace. The chance of an over treats the result as a lopsided count around the projection, because a big night sits further above the average than a bad one sits below.
What it does not know. The injury report. It only learns a player is out after he has missed a game, it cannot tell rest from injury, and a player who has been out since opening night is never counted as missing. It does not see rookies before they play, trades until the player appears for his new team, travel or motivation. A sportsbook line knows all of those. A large gap between the model and the market is far more often the model missing news than the market being wrong.