Skip to contents

Women's College Basketball RAPM: League-Wide and Within-Team Player Impact

Saiem Gilani

Source: vignettes/articles/womens-college-basketball-rapm-model.Rmd
womens-college-basketball-rapm-model.Rmd

Regularized adjusted plus-minus (RAPM) estimates how many points per 100 possessions a player adds to her team’s margin, holding fixed the other nine players on the court. wehoop publishes two RAPM estimates for NCAA Division I women’s basketball, and they answer different questions. League-wide RAPM (load_ncaa_wbb_rapm()) puts every Division I player on one scale. Within-team RAPM (load_ncaa_wbb_rapm_within_team()) splits one team’s performance among its own players and cannot be compared across teams.

This article explains both, how they are built, how they are checked, and their limits. It condenses the official model card in the ncaa-wbb-hoops-data repository; numbers quoted from that card are labelled as such. The last section shows how to fit RAPM yourself on WNBA possessions with wnba_rapm().

What the model estimates

Plain plus-minus credits a player with her team’s point margin while she is on the court. That rewards a bench player who happens to share the floor with the best starter. RAPM fixes this by estimating every player’s effect at once, so each player’s number is her effect with her teammates and opponents accounted for.

League-wide (load_ncaa_wbb_rapm()) Within-team (load_ncaa_wbb_rapm_within_team())
Question How much does this player add, on a scale shared by all of Division I? How is this team’s performance divided among its players?
Fit One ridge regression over every Division I possession in a season One fit per team, on that team’s possessions
Comparable across teams Yes No, by construction
Columns orapm, drapm, rapm_net, off_poss, def_poss, estimand rapm_off, rapm_def, rapm_net, team_off_poss, num_players
Coverage 2011 onward 2010 onward

In the league-wide file rapm_net is offense plus defense, in points per 100 possessions, and a positive defensive value is good; the within-team file signs defense the other way (see below). The estimand column ("league") is on every league-wide row so the two estimates are never confused.

Loading the ratings 

rapm <- load_ncaa_wbb_rapm(seasons = 2026)

nrow(rapm)
#> [1] 4596
count(rapm, estimand)
#> # A tibble: 1 × 2
#>   estimand     n
#>   <chr>    <int>
#> 1 league    4596

The top of the 2025-26 season among players with at least 800 possessions, offense and defense combined (the floor the model card uses for its leaderboard):

rapm |>
  filter(off_poss + def_poss >= 800) |>
  arrange(desc(rapm_net)) |>
  select(player, team, off_poss, orapm, drapm, rapm_net) |>
  mutate(across(where(is.double), ~ round(.x, 2))) |>
  head(15)
#> # A tibble: 15 × 6
#>    player                team           off_poss orapm drapm rapm_net
#>    <chr>                 <chr>             <int> <dbl> <dbl>    <dbl>
#>  1 SARAH.STRONG          UConn              1954 11.2   8.44     19.7
#>  2 MADISON.BOOKER        Texas              2328  8.53  9.21     17.7
#>  3 AZZI.FUDD             UConn              2088  8.04  8.95     17.0
#>  4 JOYCE.EDWARDS         South Carolina     2270  8.33  7.52     15.8
#>  5 GABRIELA.JAQUEZ       UCLA               1998  8.72  6.03     14.8
#>  6 HANNAH.HIDALGO        Notre Dame         2316  6.84  7.68     14.5
#>  7 MADINA.OKOT           South Carolina     1678  8.93  4.96     13.9
#>  8 JORDAN.LEE            Texas              2260  8.86  4.61     13.5
#>  9 MILAYSIA.FULWILEY     LSU                1567  6.47  6.95     13.4
#> 10 ASHLYNN.SHADE         UConn              1934  4.69  8.4      13.1
#> 11 AUBREY.GALVAN         Vanderbilt         2171  9.87  3.12     13.0
#> 12 GIANNA.KNEEPKENS      UCLA               1865  8.22  4.68     12.9
#> 13 CHARLISSE.LEGERWALKER UCLA               1871  7.7   5.02     12.7
#> 14 RAEGAN.BEERS          Oklahoma           1688  7.15  5.55     12.7
#> 15 TESSA.JOHNSON         South Carolina     1952  8.46  4.23     12.7

Data: possessions, lineups and stints

RAPM needs to know who was on the court for every possession. The NCAA engine reconstructs that from stats.ncaa.org play-by-play by tracking substitutions, and wehoop publishes each layer:

stints <- load_ncaa_wbb_matchup_stints(seasons = 2026)

nrow(stints)
#> [1] 208582

stints |>
  select(home, away, period, duration_seconds, n_possessions,
         home_lineup, home_pts, away_pts) |>
  head(3)
#> # A tibble: 3 × 8
#>   home          away  period duration_seconds n_possessions home_lineup home_pts
#>   <chr>         <chr>  <int>            <int>         <int> <chr>          <int>
#> 1 George Washi… Dist…      1              262            14 EMMA.THEOD…       13
#> 2 George Washi… Dist…      1               66             4 EMMA.THEOD…        2
#> 3 George Washi… Dist…      1               45             4 CAIA.LOVIN…        5
#> # ℹ 1 more variable: away_pts <int>

As reported in the model card, the inputs are the published possessions, team rosters and a name-change table that keeps a player in one column across transfers and spelling variants. Seasons whose usable-possession share falls below a gate are excluded; 2010 fails it, which is why league-wide RAPM starts in 2011. The card reports between about 76% and 80% of possessions usable in the early seasons.

How the model works

Each possession is one row of a regression. Each player has an offensive column and a defensive column, switched on when she is on the court for that side. The model also has an intercept (the league’s average points per 100 possessions) and a home-court term. The outcome is the points scored.

Ordinary least squares would fail: players who share most of their minutes are nearly impossible to separate, and players with a handful of possessions would get wild estimates. Ridge regression adds a penalty on the size of every coefficient, which shrinks players toward zero (league average) unless the data pull them away:

β̂=arg⁡minβ∑iwi(yi−xi⊤β)2+λ‖β‖2 \hat\beta = \arg\min_\beta \; \sum_i w_i (y_i - x_i^\top \beta)^2 + \lambda \lVert \beta \rVert^2

As reported in the model card, the league-wide fit uses λ=1000\lambda = 1000, fixed by a validation program and checked at run time so a silently unregularized fit cannot publish. The card reports intercepts between about 87 and 93 points per 100 possessions and home-court advantage between about 2.4 and 2.8 points per 100 across seasons, and uses both as scale checks: a normalization bug moves them out of band even when the player ordering survives.

Shrinkage is visible in the published ratings: the fewer possessions a player has, the closer her estimate sits to zero.

ggplot(rapm, aes(off_poss, abs(rapm_net))) +
  geom_point(alpha = 0.15, color = "#221A4D") +
  geom_smooth(method = "loess", se = FALSE, color = "#b298dc", formula = y ~ x) +
  labs(
    title = "Shrinkage: low-possession players are pulled toward zero",
    x = "Offensive possessions", y = "|RAPM| (points per 100)",
    caption = "Data: wehoop::load_ncaa_wbb_rapm(seasons = 2026)"
  ) +
  theme_minimal()

Scatter plot of 2025-26 NCAA women's league-wide RAPM magnitude against offensive possessions; low-possession players cluster near zero.

Uncertainty

The model card describes ridge posterior standard errors for every player (orapm_se, drapm_se, rapm_net_se) and a split-half calibration check: the season’s games are split in two, both halves are refit, and each player’s two estimates are compared with the standard errors. As reported in the card, the published (posterior) standard errors are conservative by about 2.5 times relative to how much an estimate moves between two halves of a season, and a typical player’s standard error is about 4 points per 100. The card’s advice: read rapm_net plus or minus two standard errors as a cautious band, and treat overlapping bands as a tier, not a ranking.

intersect(c("orapm_se", "drapm_se", "rapm_net_se"), names(rapm))
#> character(0)

At the time this page was built the release file that load_ncaa_wbb_rapm() reads did not include the standard error columns (the result above is empty when they are absent). The practical point stands either way: at roughly 4 points per 100 of uncertainty, the top of any RAPM leaderboard is a group of players, not an order.

Within-team RAPM

Within-team RAPM fits each team separately. It answers “who drives this team’s results” rather than “who is best in the country”.

within <- load_ncaa_wbb_rapm_within_team(seasons = 2026)

within |>
  filter(team == "UConn") |>
  arrange(desc(rapm_net)) |>
  select(team, player_code, rapm_off, rapm_def, rapm_net, team_off_poss, num_players) |>
  mutate(across(where(is.double), ~ round(.x, 2))) |>
  head(8)
#> # A tibble: 8 × 7
#>   team  player_code        rapm_off rapm_def rapm_net team_off_poss num_players
#>   <chr> <chr>                 <dbl>    <dbl>    <dbl>         <dbl>       <int>
#> 1 UConn Fudd, Azzi             6.17    -6.17    12.3           2472          12
#> 2 UConn Arnold, KK             4.63    -4.43     9.06          2472          12
#> 3 UConn Shade, Ashlynn         3.02    -4.33     7.34          2472          12
#> 4 UConn Heckel, Kayleigh       4.11    -2.09     6.2           2472          12
#> 5 UConn Strong, Sarah          4.43    -1.68     6.11          2472          12
#> 6 UConn Ziebell, Allie         4.52    -1.59     6.11          2472          12
#> 7 UConn Quinonez, Blanca       1.99    -3.94     5.93          2472          12
#> 8 UConn Ducharme, Caroline     2.45    -1.61     4.06          2472          12

Mind the defensive sign. The two files use different conventions. In the league-wide file a positive drapm is good and rapm_net = orapm + drapm. In the within-team file rapm_def is measured in points allowed, so a negative value is good and rapm_net = rapm_off - rapm_def. Check it on the data before combining them:

c(
  league_net_is_o_plus_d = isTRUE(all.equal(rapm$rapm_net, rapm$orapm + rapm$drapm,
                                            tolerance = 1e-6)),
  within_net_is_o_minus_d = isTRUE(all.equal(within$rapm_net, within$rapm_off - within$rapm_def,
                                             tolerance = 1e-6))
)
#>  league_net_is_o_plus_d within_net_is_o_minus_d 
#>                    TRUE                    TRUE

Comparing the two net estimates for the same players:

stopifnot(identical(class(rapm$player_id), class(within$player_id)))

both <- inner_join(
  rapm |> select(player_id, league_net = rapm_net, off_poss, def_poss),
  within |> select(player_id, within_net = rapm_net),
  by = "player_id"
)

both |>
  filter(off_poss + def_poss >= 800) |>
  summarise(players = n(), correlation = cor(league_net, within_net))
#> # A tibble: 1 × 2
#>   players correlation
#>     <int>       <dbl>
#> 1    3162       0.845

They agree broadly but not exactly. A player on a deep team shares credit with good teammates in the within-team fit, while the league-wide fit also credits her for how her team performs against the rest of Division I. Use within-team RAPM to compare teammates, and league-wide RAPM for everything else.

How well does it work?

As reported in the model card, the league-wide model is gated before publishing on every season:

  • the usable-possession share must be at least 0.65;
  • the intercept must sit in an 83 to 98 band, and home-court advantage in a 1 to 4 band, per 100 possessions;
  • the standard error checks must pass (standard errors fall as possessions rise, and the split-half calibration coverage stays in band).

Externally, a team’s possession-weighted average of its players’ RAPM is compared with Bart Torvik’s adjusted efficiency margin. As reported in the card, the rank correlation (Spearman) is about 0.98 for 2023 through 2026 and 0.90 for 2021; Torvik has no women’s ratings before 2021, so earlier seasons carry no external check. This validates the team-level aggregate, not the ordering of individual players.

Fit your own WNBA RAPM with wnba_rapm()

wehoop exports a ridge RAPM you can run yourself. wnba_rapm() takes any possession table with off_player_1 to off_player_5, def_player_1 to def_player_5 (player ids) and points, and returns one row per player with o_rapm, d_rapm, rapm, off_poss and def_poss. Under the hood it:

  • builds a sparse design with one offensive and one defensive indicator column per player, dropping possessions with an incomplete lineup;
  • fits a ridge regression with glmnet (no intercept), choosing the penalty by 10-fold cross-validation with fixed, deterministic folds, so repeated calls give identical results without setting a seed;
  • multiplies by 100 and flips the sign of the defensive coefficient so that good defense is positive.

It needs the glmnet and Matrix packages. The function’s documentation notes that a full season of possessions is needed for stable estimates; a single game is for illustration only.

load_wnba_stats_possessions() publishes WNBA possessions with lineups from a GitHub release, so a full season fits in seconds without calling the WNBA Stats API:

poss <- load_wnba_stats_possessions(seasons = 2025) |>
  # regular-season WNBA Stats game ids start with "102"
  filter(substr(game_id, 1, 3) == "102", count_as_possession == TRUE)

nrow(poss)
#> [1] 44528

my_rapm <- wnba_rapm(poss)

nrow(my_rapm)
#> [1] 181

How close is this to the published suite, which uses its own penalty and design? Compare with the regular-season RAPM in load_wnba_player_impact():

published <- load_wnba_player_impact(seasons = 2025) |>
  filter(season_type == "Regular Season") |>
  select(player_id, player_name, pub_rapm = rapm, pub_o = o_rapm, pub_d = d_rapm)

stopifnot(identical(class(my_rapm$player_id), class(published$player_id)))

compare <- inner_join(my_rapm, published, by = "player_id") |>
  filter(off_poss >= 1000)

compare |>
  summarise(
    players = n(),
    r_offense = cor(o_rapm, pub_o),
    r_defense = cor(d_rapm, pub_d),
    r_total = cor(rapm, pub_rapm),
    sd_wnba_rapm = sd(rapm),
    sd_published = sd(pub_rapm)
  )
#> # A tibble: 1 × 6
#>   players r_offense r_defense r_total sd_wnba_rapm sd_published
#>     <int>     <dbl>     <dbl>   <dbl>        <dbl>        <dbl>
#> 1     102     0.949     0.951   0.946         3.83         2.31
ggplot(compare, aes(pub_rapm, rapm)) +
  geom_abline(linetype = "dashed", color = "grey60") +
  geom_point(alpha = 0.7, color = "#221A4D") +
  labs(
    title = "2025 WNBA: wnba_rapm() vs the published suite",
    subtitle = "Players with 1,000+ offensive possessions",
    x = "Published RAPM (load_wnba_player_impact)", y = "wnba_rapm()",
    caption = "Data: wehoop::load_wnba_stats_possessions(), wehoop::load_wnba_player_impact()"
  ) +
  theme_minimal()

Scatter plot comparing RAPM fitted with wnba_rapm() against the published WNBA player impact RAPM for 2025 players with at least 1,000 offensive possessions.

For regular-rotation players the two rank players very similarly. Compare the two standard deviations above for the difference in spread: the cross-validated penalty in wnba_rapm() is chosen per fit, while the suite uses its own settings, and a stronger penalty pulls every estimate closer to zero.

To build possessions for a single game yourself, wnba_possession_lineups() calls the WNBA Stats API play-by-play (with on-court players) and returns the same possession layout. It is a live call to stats.wnba.com, so it is not run here.

one_game <- wnba_possession_lineups(game_id = "1022500001")
wnba_rapm(one_game)   # illustration only: one game is far too small

Limitations

  • Shrinkage is a choice. Low-possession players sit near zero because the model knows little about them, not because they are average. Filter on possessions before ranking.
  • Collinearity. Players who share nearly all their minutes cannot be separated by the data; the penalty decides how their joint effect is split.
  • Team-level validation only. The external check against Torvik validates the team aggregate. There is no public benchmark for individual women’s college RAPM.
  • Within-team RAPM is not comparable across teams, and as reported in the card it is produced by a manual step that needs the raw play-by-play bundles.
  • Not yet in the published model, as reported in the card: multi-season pooling and a box score prior, both measured in development, and a luck adjustment for three-point shooting.

Where to go next

Citation 

@misc{wehoop_ncaa_wbb_rapm,
  author = {Saiem Gilani},
  title = {Women's College Basketball RAPM Model},
  url = {https://wehoop.sportsdataverse.org/articles/womens-college-basketball-rapm-model.html},
  year = {2026}
}