Chapter 9
Ratings Through Time
Every morning the model is fit on every game before that day. Kept, the fits make a record: each player’s rating as of each day, back as far as the data allows. This chapter is about that record, what it is and isn’t, and the engineering that keeps it consistent.
9.1 What we would have said
Definition 9.1 (Daily
ratings). A player’s rating as of day t is their rating in the model fit on
the games before t. The published
dataset ratings_daily has one row per player and game day
of each season, and one for the day after each season’s last game, with
the offensive and defensive ratings, the box-score prior and the
possessions played so far that season. Each day’s ratings are relative
to that day’s average player: weighted by their expected minutes in the
day’s lineups, they average 0.
Because each day’s fit only sees the past, a player’s ratings over time are what the model would have said about them that morning: the predictions of Chapter 8 were made from exactly these numbers. That makes them different from season-end ratings, which are fit to the whole season at once and describe it in hindsight. A daily rating in November rests on a few weeks of this season and a discounted memory of the last ones. Season-end ratings know how the season turned out.
The history starts in 2012. The box-score prior learns from earlier seasons of plus-minus (Chapter 6), so a season can only be rated once there are two seasons of play-by-play before it, and nba_data’s play-by-play currently starts in 2010. The first seasons rest on less than the later ones: a prior learned from two seasons rather than ten, and a fit with fewer years of games behind it.
9.2 A career
Figure 9.1 follows Paul George through sixteen seasons and five teams. In Indiana, George’s rating ended each season between +3.4 and +4.9, with defense rated higher than offense in all but the last. After the trade to Oklahoma City it ended 2018 at +3.0 and 2019 at +6.1. Through five seasons with the Clippers it stayed between +4.7 and +6.1, most of it on offense. In Philadelphia, with 41 games played in 2025 and 49 in 2026, it fell to +2.8 and then +2.4, and this season, in Boston, it started near +2.3.
The figure shows the model’s habits as well as the player’s. Within a season the rating moves in small steps, a game’s worth of evidence at a time, and most at the start, when each game is a larger share of what’s new. Between seasons it can jump, because the box-score formula is refit on opening night (Chapter 6). And the defensive line is the flatter one, which is true of the ratings generally: among 2026’s regulars, the published defensive ratings have a standard deviation of 1.5 against offense’s 1.8. Defense leaves less of a trace in the box score and in the scoring margin, so the data moves defensive ratings less far from their priors.
9.3 Teams at full strength, and tonight
The same ratings and minutes of Chapter 7 rate teams. Two ratings are published for each team and day, both relative to the league’s average team:
- Full strength: the players who’ve played for the team this season, with 240 minutes given out in order of their minutes when playing, as if everyone were healthy.
- Tonight: the expected lineup the predictions use, with each player’s chance of playing (Definition 7.4).
The gap between them is the team’s injury report, in points per 100 possessions. Figure 9.2 returns to the Warriors of Chapter 1.
At full strength the Warriors averaged +5.9 points per 100 possessions over the season, a good team. With the lineup expected each day they averaged +3.2, and the gap follows Curry’s absences: on March 20, in the middle of Curry’s longest stretch out, they were +7.8 at full strength and +1.2 as they would actually take the floor. A team rating fit to results, Chapter 1’s baseline, sees one number somewhere in between and can’t tell which team is playing tonight.
9.4 One model for all of history
The record has to be consistent to be worth charting. A model that changed in March would draw a jump in every player’s line in March, which would look like something about the players. So every season of the published history comes from the same model.
Definition 9.2 (Model version). The model version is a hash of the program’s code, its locked dependencies, and the first season of play-by-play. Every row of the daily datasets records the version that computed it, and every row of a season has the same one.
When the version changes, every season is refit from scratch, about 2,900 days of fits. The first season of play-by-play is in the hash because of Chapter 6: adding earlier seasons changes what the box formula learns, and so every rating after them, even though no code has changed. This happened while this book was being written. On November 5, 2026, nba_data added play-by-play for 2012 through 2015, the version changed, and the publishing job refit thirteen seasons, starting the history in 2014 instead of 2018, without anyone doing anything. Days later, play-by-play for 2010 and 2011 arrived, and it started in 2012.
In the code. The datasets are built by yarapm/publish.py: build
writes the current season every time nba_data updates, refitting only
the new days if the season’s files were built by this version (at about
0.2 to 0.9 seconds a day), and backfill refits finished seasons whose
files are stale, which takes 45 to 75 minutes for all of them in GitHub
Actions. The version is model_version. The full-strength lineups
are in yarapm/strength.py.
Each season is one Parquet file, sorted by player and written in row groups of 20,480 rows. A reader filtering on one player can skip every row group whose statistics say the player isn’t in it, so one player’s career, over sixteen files, is about a megabyte over HTTP, little enough to chart in a browser with DuckDB’s WebAssembly build:
SELECT r.season, r.as_of_date, r.o_rating, r.d_rating, r.rating
FROM yarapm.ratings_daily r JOIN nba.players p USING (player_id)
WHERE p.name = 'Paul George'
ORDER BY r.as_of_date;9.5 What isn’t there yet
The record has no error bars. Proposition 3.5 gave the variance of the least squares estimate, and the same argument for ridge gives a posterior variance of roughly \sigma^2 (X^\top W X + \lambda I)^{-1} (Exercise 9.3), so every rating could carry a band, narrow for a veteran and wide for a rookie. Computing the diagonal of that inverse for 2,700 unknowns every day is the hard part; it’s one of the project’s open experiments. Until then, read small differences between players, a point or less, as noise.
Exercises
Exercise 9.1. Chart any player’s career
from yarapm.ratings_daily with the query above. Find a
player whose offensive and defensive ratings moved in opposite
directions over a few seasons.
Exercise 9.2. From
yarapm.team_ratings_daily, find the team and day in 2026
with the largest gap between full_net and net.
Using yarapm.team_players_daily, which players account for
it?
Exercise 9.3. Under the assumptions of Proposition 4.3, show that the posterior covariance of the ratings is \sigma^2 (X^\top W X + \lambda I)^{-1}. For the one player of Theorem 4.2, what is the posterior standard deviation with n possessions? With \sigma = 110 and \lambda = 3{,}000, how many possessions does it take to get it below 1 point per 100?
Exercise 9.4. nba.openrapm
has season-end ratings fit to whole seasons. For 2026, compare a few
players’ OpenRAPM ratings with their ratings_daily ratings
on the day after the season’s last game, and on the first day of the
season. Which is closer to OpenRAPM, and why should it be?