<!-- canonical: https://blog.scrymore.com/blog/lineup-win-rate-board/ | published: 2026-11-08 -->

# Lineup Win Rate Board: five-man units, and what a model makes of them

![The Lineup Win Rate Board for 2026-27: ten five-man units, each with a won, lost and tied bar and a win percentage.](/blog-images/lineup-board/board.jpg)

Box scores tell you how a player did over a night. They don't tell you how a specific five did together. We sliced the league's public play-by-play into stints, each one a stretch with the same ten players on the floor, and counted how each five-man unit did. This post covers games through 7 November 2026 (139 games) and what a model trained on the previous five seasons makes of the results.

All counts are stints, not games. A stint of zero game-clock seconds is not counted. Win % is stints won divided by stints won plus lost, with ties left out.

## The board

We took the ten lineups with more than 50 minutes that have the best plus/minus in the league, then ordered them by how often they win a stint.

| # | Team | Lineup | W-L-D | Win % | Min | +/- | Net per 100 |
|---|---|---|---|---:|---:|---:|---:|
| 1 | SAS | Fox, Vassell, Wembanyama, Castle, Harper | 23-10-4 | 69.7 | 105 | +70 | +32.5 |
| 2 | MIN | Edwards, Ball, McDaniels, Kuminga, Gobert | 13-7-3 | 65.0 | 56 | +31 | +28.5 |
| 3 | MIN | Ball, McDaniels, Kuminga, Dosunmu, Gobert | 18-10-4 | 64.3 | 70 | +20 | +11.3 |
| 4 | LAL | Dončić, Thybulle, Williams, Reaves, Kessler | 10-6-6 | 62.5 | 58 | +17 | +17.6 |
| 5 | ATL | Landale, Alexander-Walker, Dort, Wiggins, Flemings | 17-12-3 | 58.6 | 54 | +14 | +11.1 |
| 6 | LAL | Dončić, Thybulle, Reaves, Kessler, LaRavia | 8-6-2 | 57.1 | 53 | +33 | +32.1 |
| 7 | ATL | Alexander-Walker, Okongwu, Johnson, Daniels, McCollum | 19-15-4 | 55.9 | 100 | +18 | +7.2 |
| 8 | DEN | Murray, C. Johnson, Braun, Gordon, Jokić | 17-14-4 | 54.8 | 76 | +20 | +12.8 |
| 9 | MIL | Turner, Kuzma, Herro, Rollins, Jaquez | 12-11-4 | 52.2 | 76 | +37 | +20.2 |
| 10 | MEM | Jerome, Edey, Wells, Coward, Boozer | 9-10-6 | 47.4 | 73 | +18 | +12.1 |

Net per 100 is points scored minus points allowed per 100 possessions. Possessions are counted from play-by-play and credited to the stint where they end, which runs about 4% under the league's own count. These lineups have far fewer than 550 possessions each, so read the ratings as early.

Two things stand out. The Spurs' starting five leads the board by a wide margin. The Lakers appear twice, both with Dončić. And a unit can have a large plus/minus on a record close to even, as Milwaukee's +37 on 12-11-4 shows: its margin came from a few big stints.

## What a model expected

We built a model that predicts the result of each stint from the ten players on the floor. It is a set model with a learned vector for each player, plus each player's season-to-date box score numbers, trained on 2021-22 through 2024-25 and tuned on 2025-26. It sees the stint's opponents and updates its box-score features before every stint.

Here is how it did on these ten lineups, and across the 148 lineups with at least ten stints.

:::stats
0.38 | Rank correlation | Between the model's expected net rating and the actual one, 148 lineups.
4 of 10 | Board lineups in the model's top 25 | Spurs starters, both Dončić units and the Nuggets.
23% vs 19.5% | Ties predicted vs actual | The raw model over-predicts ties.
54.7% | Spurs starters' expected win rate | They won 69.7%.
:::

- **The direction is right.** Split the 148 lineups into four equal groups by expected net points. The top group averaged +8.1 actual points, the bottom group -4.4. A rank correlation of 0.38 over 148 lineups is far from chance, but it is a loose match, not a ranking you could bet on.
- **The size is not.** The model's best lineups are expected to win only 54 to 57% of stints. Its predictions are compressed toward even, so most of the lineups on this board did better than it expected, by a lot for some. It expected +16 points from the Spurs starters, who actually scored +70.
- **Some misses are large.** It saw Minnesota's two lineups as ordinary, about 50% each, and they won 64 to 65%. It also rated Milwaukee's lineup below even (48.5%), and it won 52.2% of decided stints with a +20.2 net rating.

Most of this is luck. These lineups have 10 to 40 stints each, and a lineup with a true 55% win rate goes 1-7 in eight decided stints about 2 times in 100. We looked at many lineups, so extreme records are expected. Picking the best ten by plus/minus also picks the luckiest ones.

## Fixing the tie rate

The model's raw tie share was too high, 23.0% against 19.5% actual. We fitted a correction on 2024-25 and scored it on 2026-27.

| | Raw | Fixed |
|---|---:|---:|
| Stint log-loss (lower is better) | 1.0351 | 1.0149 |
| Predicted ties | 23.0% | 19.5% |
| 90% band coverage, 148 lineups | 88.5% | 91.2% |

Most of the miscalibration was the tie share. The model needed only about 12% more sharpening on wins versus losses. The correction does not change the rank order of lineups, so the 0.38 stays at 0.38.

## What it can't do

We tried swapping players into the Spurs' starting five. Replacing one starter with Kornet or Champagnie changed the expected win rate by up to about 2 percentage points, with Kornet for Vassell the best case. But the same method run on 4,903 real one-player swaps in the last two seasons predicted the change with a correlation of 0.04. Treat swap results as tendencies, not forecasts. The model scores players individually and does not know how a specific five fit together.

## Caveats

- Samples are small. Most lineups here have 10 to 40 stints.
- Starters are the five players on the floor at the start of a team's first stint in a game.
- The model has no information on injuries, rest or opponent pairing beyond who is on the floor.
- We counted stints, not games, so a lineup's record is not a game record.

Source: NBA play-by-play from the league's public S3 feed, sliced by lineup. Names and teams appear as text only.
