Lineup Chemistry
Do five-man lineups play better than the sum of their players? Each lineup is compared with what last season's player ratings predicted. Ratings are LEBRON, a public all-in-one rating from BBall Index in points per 100 possessions: 0 is an average player, about -2 is replacement level, +2 or more is a clear starter. Lineup results are net rating, points scored minus points allowed per 100 possessions. A Ghost Screen Analytics model, not an NBA stat. The tests were written before the model was run.
How to read it
- Expected
- What the five players' ratings from the season before add up to, turned into a net rating. The dark tick on each range.
- Actual
- What the lineup really did. The hollow circle.
- Gap
- Actual minus expected. Green means the lineup beat its parts, red means it fell short.
- 80% range
- Where about four of every five lineups land. It is very wide: a lineup with a few hundred possessions swings a lot on luck alone.
Who is left out. Only lineups with 100+ possessions are scored. Coaches keep playing lineups that work, so the lineups that reach that mark skew good, and the biggest gaps below are partly selection and partly luck. A player with no rating the year before is set to replacement level (-2).
Does keeping players together help?
| Season | Familiar lineups minus new ones |
|---|
Points per 100 possessions by which lineups with 4 or 5 returning players beat their expected rating, compared with lineups with 0 to 2 returning players. Positive in 7 of 9 seasons. Description of bar L7.
The lineups
What it is built on
Sample: lineup-seasons tested, 2017-18 to 2025-26, from the NBA's lineup tables (regular season, five-man lineups with 100+ possessions), scored against LEBRON ratings from the season before. Each test season used only earlier seasons for training. Inputs to the chemistry model: sum of ratings, best player, spread of ratings, offense and defense sums, centers, guards, returning players and average age.
The rules, code, data and results are in the lineup-model folder on GitHub. Four of seven bars passed, and the model is not better than simply adding up the parts. The 2026-27 season is not forecast lineup by lineup because rotations are not settled.