The Expected Score Chapter 7 · From Ratings to a Score

Chapter 7

From Ratings to a Score



We have a rating for every player, a level for the league, and terms for home court and the playoffs. To predict a game we need two more things: how many minutes each player will play, and how many possessions the game will have. This chapter adds both, and then predicts one game from start to finish: Game 5 of the 2026 Finals, as the model saw it on the morning it was played.

7.1 A team is its minutes

A team’s strength in a game is its players’ ratings weighted by how much each is expected to play. Five players are on the floor at a time, so over a game a team’s players play 240 minutes between them.

Definition 7.1 (Team strength). If player i is expected to play m_i minutes for a team, with \sum_i m_i = 240, the team’s offensive and defensive strengths are

\begin{aligned} O &= 5 \sum_i \frac{m_i}{240}\, o_i , \\ D &= 5 \sum_i \frac{m_i}{240}\, d_i . \end{aligned}(7.1)

The factor of 5 makes this the average over the team’s possessions of the sum of the five ratings on the floor, which is what Equation (3.1) adds up. A team that played the same five players all 48 minutes would have O equal to the sum of their offensive ratings.

Definition 7.2 (Expected score). In a game between A, at home, and B, with P expected possessions each, the expected points are

\begin{aligned} \hat S_A &= \tfrac{P}{100}\left(\ell + \pi + \eta + O_A - D_B\right), \\ \hat S_B &= \tfrac{P}{100}\left(\ell + \pi - \eta + O_B - D_A\right), \end{aligned}(7.2)

where \ell is the league’s level (Definition 5.4), \pi the playoff term if it’s a playoff game, and \eta the home-court term (Section 4.5), which is zero at a neutral site.

This is Equation (1.1) with each efficiency predicted from the model. All that remains is m_i and P.

7.2 Minutes when playing

How long will a player play, if they play? Rotations change slowly, so the best guess is recent history.

Definition 7.3 (Minutes when playing). A player’s projected minutes when playing are an average of their minutes in their recent games, with the kth most recent game weighted by 2^{-k/5}: a half life of five games. Each game’s minutes are first scaled so the team’s add to 240, so overtime doesn’t count. A player who started their last game, and doesn’t usually start, gets up to five minutes more, and one who usually starts but didn’t, up to five fewer. A player with no games gets 10 minutes.

The half life is counted in the player’s own games, not in days, so a player back from a month out with an injury is projected from the games before it rather than from nothing. And the player’s team doesn’t enter: a player traded yesterday brings their minutes along, which is usually about right. The five-game half life and the five-minute start bonus were tuned on how well they predicted minutes. Over the 2021–2024 seasons, the projections missed each player’s minutes by 4.7 a game on average when we knew who would play, and the predicted margins were barely sensitive to the choice: within 0.03 points over the whole grid.

7.3 Who plays

The harder question is who will play at all. The NBA publishes injury reports, but nba_data doesn’t have them yet, so the model guesses from each player’s recent pattern.

Definition 7.4 (Chance of playing). A team’s roster is the players who played in one of its last 10 games and haven’t played for another team since. For each player on it, let g be the number of the team’s games since they last played and p the number of the last 10 they played in. The chance they play tonight is the share of players in the same spot (g, p) who played their team’s next game, over the previous three seasons: \Pr(\text{plays}) = \frac{\text{played} + \tfrac12}{\text{in that spot} + 1} .

The \tfrac12 and 1 are one last application of Theorem 4.2: a spot seen rarely is pulled toward even odds. Figure 7.1 shows the table as of the 2026 Finals. A player who has played all of their team’s last 10 games plays the next one 94% of the time. Missing just the last game, with 9 of 10 played, drops it to 49%: a player who sat last night is usually still hurt. And a player out for a week or more is a long shot.

0 1 2 3 4 5 6 7 8 9 0 0.25 0.5 0.75 1 team games since the player last played chance of playing played 9 of 10 played 7 of 10 played 4 of 10 played 1 of 10
Figure 7.1. The chance a player on a team’s roster plays its next game, by the number of team games since they last played, for four rates of recent appearance. Fit on the 2023–2025 seasons, as of the 2026 Finals.

A player’s expected minutes are then their chance of playing times their minutes when playing, and the team’s are scaled to add to 240, with no player above 42; minutes over the cap go to the others in proportion.

This is the weakest part of the model, and the scoreboard shows it. When the harness is allowed to know who actually played, standing in for an injury report, the margin RMSE on 2021–2024 is 13.58; guessing, as it must for real predictions, it’s 13.71. Knowing who plays is worth more than the box-score prior.

7.4 How many possessions

Pace is modeled the way Chapter 1 suggested, as a property of the game that both teams contribute to, and it’s fit on team games rather than segments.

Definition 7.5 (Adjusted tempo). For a game between teams A and B with P possessions in 48 minutes, \log P = c + \pi_P + t_A + t_B + \varepsilon, where c is the league’s tempo, \pi_P a playoff term and t_A, t_B the teams’ effects, each pulled toward zero by a small ridge penalty, and games older than about two months count half (a half life of 60 days). The predicted possessions are e^{c + \pi_P + t_A + t_B}, times the league’s average game length in units of 48 minutes, to allow for overtime.

The logarithm makes the effects multiply: a team 3% faster than average playing a team 2% slower gives a game about 1% faster than average. In June 2026 the league’s tempo was 100.0 possessions per 48 minutes, and the playoff term was -2.7%: playoff games are slower.

7.5 Game 5, predicted

Here is everything at once, for the game in Figure 2.1: San Antonio at home against New York, fit on every game through June 12.

Table 7.1. The model’s prediction of Game 5 of the 2026 Finals, made from games before June 13, 2026, built up term by term from Equation (7.2).
Spurs Knicks
league level, \ell 105.38 105.38
playoff term, \pi -3.93 -3.93
home court, \pm\eta +0.93 -0.93
own offense, O +14.22 +13.73
opponent’s defense, -D -0.84 -5.06
points per 100 possessions 115.76 109.19
expected possessions, P 96.2 96.2
expected points 111.3 105.0
actual points 90 94

The Spurs’ offense rates slightly better than the Knicks’, and their defense much better: Wembanyama alone, projected for 37 minutes with a defensive rating of +6.3, contributes 5 \times \tfrac{37.1}{240} \times 6.3 \approx 4.9 points per 100 to it. With home court, the model made San Antonio a 6.3-point favorite, and a 68% chance to win (Chapter 8 explains the conversion).

Remark 7.6. The individual terms in Table 7.1 are less stable than their sum. While this chapter was being written, nba_data added four earlier seasons of play-by-play, and the box-score prior learned from them. Refit, the league level for this game fell from 107.8 to 105.4, while both teams’ offensive strengths rose by about 0.6 and their defensive strengths fell by about 1.8, which together make up the 2.4. The predicted scores changed by a hundredth of a point. This is Proposition 3.3 at work: shifting the ratings one way and the level the other predicts the same scores, and the data only pins down the combination. Here the prior, not the data, decided where the level sits.

The Knicks won by four. The game was slower than predicted, about 90 possessions to 96, and both offenses far worse: the Spurs scored 100 per 100 possessions against a prediction of 116. A 68% favorite loses about one time in three, and this was one of those times. The question Chapter 8 takes up is whether, over thousands of games, the favorites lose as often as the model says they should, and no more.

The projected minutes are worth a look too. For the Knicks, Brunson was projected for 36 minutes and played 41. OG Anunoby, who had missed two games of the second round in May and so had played 8 of the Knicks’ last 10, was given an 87% chance of playing rather than 94%, and played 33 minutes. In a Finals elimination game, starters play more than their season pattern says, which the model doesn’t know. Playoff rotations are one of the open experiments in the project.

In the code. The expected score, Equation (7.2), is PlayerRatings.predict, and team strength is strength, which takes each team’s projected minutes and normalizes them, so it works with any total. Minutes are ProjectedMinutes: minutes when playing are one SQL query with a window function numbering each player’s games (when_playing), the roster and the (g, p) spots come from rosters and availability, and allocate scales a team’s minutes to 240 under the cap, iterating because capping one player changes everyone else’s share. Tempo is Tempo, a small dense weighted ridge regression with one column per team.

Exercises

Exercise 7.1. A team plays five players 48 minutes each, with offensive ratings +4, +2, +1, 0 and -1. What is O? If the second player is out and a player rated -2 takes their minutes, how many points per game does the team’s expected margin fall, at 100 possessions?

Exercise 7.2. Using Table 7.1, how much of the Spurs’ 6.3-point expected margin came from home court? From the difference in defense? If Wembanyama had been ruled out that morning, and the 37 projected minutes had gone to players with an average defensive rating of 0 and offensive rating of +1, what would the predicted margin have been?

Exercise 7.3. Show that under Definition 7.5, the predicted possessions of A against B are the league’s tempo times (1 + a)(1 + b) approximately, where a and b are the teams’ percentages above average, written as fractions. What’s the error of the approximation for a = b = 0.05?

Exercise 7.4. yarapm.minutes has each player’s projected minutes and chance of playing for every upcoming game. Pick tonight’s games and find the players with the lowest play_prob who are still projected for more than 10 minutes. What does that say about the model’s view of them?