Measuring Player Impact: Box Score, On/Off and Minutes
No single stat captures a basketball player's value. Box scores miss most defence; plus-minus is noisy and depends on teammates. We combine both, and weight them by what actually predicts future results.
The inputs
- Per-100-possession box production: points, shot attempts, free throws, threes, offensive and defensive rebounds, assists, steals, blocks, turnovers and fouls.
- On-court margin: the team's scoring margin per 100 possessions while he plays.
- On/off margin: on-court margin minus the team's margin while he sits.
All are time-decayed and shrunk toward a replacement-level player until the sample is large enough.
The weights come from prediction
Rather than hand-picking weights, we let the game model learn them: the weights that best predict future game margins from each side's expected rotation. A player's impact is then the points per 48 minutes his rates add, versus a replacement-level player, under those weights.
How to read it
+6 or more is MVP-level; +2 to +4 is a quality starter; around 0 is a rotation player. Impact is a rate, so a +5 player in 20 minutes adds less to a game than a +3 player in 36. Browse the player table.
📚 More Articles
How We Backtest an NBA Model Without Fooling Ourselves
Walk-forward testing, frozen settings and four untouched seasons: the rules that keep an NBA projection model honest.
What a 70% Win Probability Really Means
A 70% favourite loses three games in ten. Calibration, not bold calls, is what makes a probability useful.
How NBA Totals Are Projected: Pace First, Then Efficiency
Possessions times points per possession: the two numbers behind every projected NBA total, and why league scoring trends matter.
Offensive, Defensive and Net Rating, Explained
Why points per 100 possessions beat points per game, and how opponent adjustment turns raw ratings into team strength.