On/off ratings confuse the player with the lineup
On/off net rating has become the go-to number for people who want to sound advanced. Instead of counting a player’s points, it compares how the team performs with him on the floor against how it performs when he sits, and reports the difference in net points per hundred possessions. The framing is seductive because it feels like a controlled experiment: same team, same season, player in versus player out. But it is not a controlled experiment, and the number it produces is not the player’s impact. On/off measures the performance of two different five-man lineups against two different sets of opponents, and then hands the entire gap to the one player whose name is on the split. It confuses the player with the lineups he happens to share the floor with, and in the samples people quote it, that confusion is most of the signal.
What on/off actually records
On/off takes the team’s net rating during the minutes a player is on the court and subtracts its net rating during the minutes he is off it. A positive on/off means the team outscored opponents by more with him playing; a negative one means it did better without him. The number is presented as a measure of the player, but every possession in it involved four other players, and the split says nothing about which of the five did the work.
The stat is a difference between two team results, not a measurement of an individual. It cannot see who set the screen, who made the rotation, or who missed the open shot. It sees only the scoreboard margin while a particular player was among the five on the floor, and it attributes the whole of that margin to him by default.
The bench-unit problem
The single biggest distortion in on/off is who a player shares his minutes with. A star who plays his on-court minutes beside four other good players and whose off-court minutes are covered by a weak bench will post a fantastic on/off — but the split is partly measuring how bad the backups are, not how good he is. Flip the roster construction and the same player’s on/off can crater.
This is why on/off numbers for players on lopsided rosters are almost uninterpretable in isolation. A deep team with two strong units will show smaller on/off splits for everyone, because the team barely dips when any one player sits, while a top-heavy team will show enormous splits that say more about the absence of a competent backup than about the presence of a great starter. The stat rewards being surrounded by weakness during your rest minutes.
Teammate overlap muddies everything
Because starters tend to play together and rest together, on/off splits are heavily correlated across a rotation. Two players who share almost all of their minutes will have nearly identical on/off numbers, and the split cannot tell you which of them is driving the result. Their fates are entangled, and the stat has no way to separate the passenger from the engine.
Untangling that overlap is exactly what more sophisticated plus-minus models — the regression-based adjusted versions — were built to attempt, precisely because raw on/off cannot do it. Raw on/off treats every player on a lineup as fully responsible for its results, which means it double-counts, triple-counts, and quintuple-counts the same possessions across five different players’ splits.
Opponent and schedule bias
On/off assumes the competition a player faces on the floor is the same as the competition faced when he sits, and it usually is not. Stars play more of their minutes against opposing starters, while their off-court minutes come disproportionately against second units. A player can post a mediocre on/off simply because his on-court time is spent against tougher lineups, and that has nothing to do with his ability.
Garbage time compounds it. Blowouts inflate or deflate off-court minutes with possessions that bear no resemblance to competitive basketball, and a starter who sits out the fourth quarter of routs accumulates off-court data against deep-bench players in low-effort situations. The split treats those possessions as equal to championship-leverage minutes, and they are not remotely the same thing.
Coaching patterns bake the bias in still deeper. Rotations are built so that stars rest together and return together, which means a player’s off-court minutes are not a random sample of the game but a specific, often weaker, slice of it — the stretches a coach deliberately chose to run without his best players. The comparison on/off pretends is apples-to-apples is, in practice, the strongest lineups against the toughest minutes versus the weakest lineups against the softest.
Sample size and noise
On/off is a difference between two noisy estimates, which makes it one of the least stable numbers in a box score over anything short of a full season. Early in a year, on/off leaderboards are dominated by small-sample flukes, and individual-game or ten-game on/off figures are essentially random. It takes thousands of possessions on both sides of the split before the number settles into something worth reading.
This matters because on/off gets cited weekly and even nightly, as if a single game’s split were evidence of anything. It is not. A player can post a plus-thirty on/off one night and a minus-twenty the next without changing anything about how he played, because a few lineup-wide hot or cold stretches swing the margin entirely.
What to read instead
The honest use of on/off is as a starting question, checked against the tools built to answer it. Adjusted and regularized plus-minus models attempt to strip out teammate and opponent effects, and while they are imperfect, they at least try to isolate the player rather than the lineup. On/off paired with the specific lineup data — who a player plays with and against — turns a misleading split into a source of hypotheses.
Read on/off beside the box-score and tracking numbers that describe what a player actually did: his scoring efficiency, his passing, his defensive activity. When the on/off and the individual production agree, the split is probably picking up something real. When they disagree wildly, the split is usually telling you about the roster around him, not about him.
The honest read
On/off net rating is a real measurement of a real thing — how the team performed in two sets of minutes — and at the extremes, over a full season, a consistently large split is a genuine clue worth chasing. Elite players do tend to have positive on/off numbers, and a starter whose team is repeatedly better without him is telling you something.
But it is not an individual impact metric, and quoting it as one credits a single player for a lineup, a bench, and a schedule he did not choose. Read on/off as a prompt to investigate, not a verdict to cite — a number that only means something once you know who was on the floor beside him and who was on the floor when he rested.