Lineup Data and the Five-Man Unit: Reading On/Off and Net Ratings

The box score is a record of individuals. It tells you what each of five players did, then quietly assumes the team is the sum of those five lines. But anyone who has watched basketball knows the sum is the whole point — that two players can make each other better or worse, that a spacing wing unlocks a slow center, that a ball-dominant guard and a ball-dominant forward can cancel each other out. None of that chemistry lives in the box score. It lives in lineup data, and reading that data well is one of the most useful — and most abused — skills in modern analysis.

What the box score can't see

Individual stats are additive by construction. A player's points, rebounds, and assists are his alone, tallied as if he produced them in a vacuum. But fit is interactive, not additive. The value a shooter creates by dragging a defender to the arc shows up nowhere on his own line — it shows up as a teammate's easier drive. The damage a non-shooter does by letting his man sag into the paint doesn't appear in his stats either; it appears as the offense's clogged spacing. The box score is blind to all of it because it never asks what happens between players. Lineup data asks exactly that question by changing the unit of analysis from the player to the group on the floor.

Net rating, applied to a unit instead of a team

The core tool is the same offensive and defensive rating you'd apply to a whole team, just pointed at the possessions a specific group shares the floor. For any five-man unit, you can compute how many points it scores and allows per 100 possessions and take the difference:

Lineup Net RatingNet = ORtg − DRtg = (points scored − points allowed) per 100 possessions, over only the possessions those five share the floor

The same math scales down. Two- and three-man combinations ask how a team performs whenever a particular pair or trio is on the court together, regardless of who fills the other spots — the natural way to test whether a backcourt duo actually works or whether a center and a stretch four genuinely fit. And on/off splits go the other direction, comparing a team's net rating with a given player on the floor versus off it. Stack these views and you can start to see the texture the box score misses: which groupings outscore opponents, which pairings sag, and where a single player's presence flips a unit from good to bad.

How coaches actually use it

This is not just an analyst's toy; it's a rotation tool. A coaching staff will look at which lineups have outscored opponents over a meaningful stretch and lean into them, and which combinations have been consistently underwater and break them up. If a particular trio has quietly been the team's best three-man group all year, that's an argument to keep them on the floor together in close games. If a starter's on/off split is ugly — the team better with him on the bench — that's at least a flag worth investigating, even if it isn't a verdict. Lineup data is how staffs translate the vague language of "fit" and "chemistry" into something they can stagger substitutions around.

The caveat that swallows everything: sample size

Here is the part that gets ignored, and it's the most important sentence in this article: the vast majority of five-man lineups play only a handful of minutes together, and their net ratings are mostly noise. A team runs through dozens upon dozens of distinct five-man combinations over a season — injuries, foul trouble, matchup tweaks, garbage time — and the long tail of those lineups logs only a few minutes each. A lineup that has played five minutes has seen maybe ten or twelve possessions on each end. Over a sample that small, a couple of lucky threes or one blown defensive rotation completely dominates the per-100 rate.

Consider a clearly-illustrative example, with the numbers invented purely to show the mechanic. Suppose a bench-heavy five-man unit checks in together for one short burst, plays about four minutes, and happens to ring up a 14–2 run because two contested threes drop and the opponent coughs up a couple of turnovers. Scaled to 100 possessions, that lineup's net rating might read as something absurd like +60 — a figure that, taken at face value, would make it the greatest lineup in league history by a mile. It is, of course, nothing of the sort. It's a four-minute fluke wearing the costume of a statistic. Sort any team's lineup page by net rating and the top of the list will be exactly these tiny-sample mirages, while the genuinely good high-minute lineups sit lower with sane, believable numbers.

So the discipline is simple to state and hard to follow: before you quote a lineup's net rating, look at the possession count. A rule of thumb I'd defend is that you want to see a lineup accumulate into the hundreds of possessions — not minutes, possessions — before the number is stable enough to mean much, and even then you read it with the error bars in mind. The lineups that clear that bar are almost always the starters and the most-used rotation groups, precisely because they're the ones that play enough for the noise to wash out. Everything below that threshold is a hypothesis, not a finding.

Why adjusted plus-minus exists

This noise problem is not a footnote; it's the reason an entire branch of analytics exists. Raw lineup and on/off numbers are confounded as well as noisy. A lineup's net rating is the joint product of all five players plus whoever they happened to be playing against, so a glowing five-man number might be telling you about one great player, or about soft opponents, or about the four teammates — you genuinely can't tell from the lineup figure alone. On/off has the same disease: a player's on-court number is entangled with whichever teammates usually share his minutes.

Adjusted plus-minus and its regularized cousin RAPM were invented precisely to cut through this. Instead of reading raw lineup splits, they treat every stint as data and use regression to solve for each individual player's contribution while holding teammates and opponents constant — the formal fix for exactly the confounding that makes raw lineup data so treacherous. The full story of why raw plus-minus lies and how regularization tames it is its own piece: plus-minus, RAPM, and noise. The short version for our purposes is that if you ever find yourself tempted to trust a small-sample lineup rating, the existence of RAPM is the field's way of admitting that you shouldn't.

The takeaway

Lineup data fills a real gap the box score leaves open: it's the only way to see chemistry and fit, to ask how groups perform rather than how individuals tally. Five-man net ratings, two- and three-man combinations, and on/off splits are genuinely useful for understanding rotations — when you respect the sample. The catch is that most lineups never accumulate enough possessions to say anything, so their net ratings are noise dressed up as insight, and the absurd numbers at the top of any lineup leaderboard are almost always mirages. Check the possession count first, demand hundreds before you believe a lineup figure, and remember that the confounding which makes raw lineup data so slippery is the exact reason adjusted plus-minus had to be invented.