DAWG — Recorded Actions covers four selected recorded box-score actions: steals, blocks, offensive rebounds and assists. It omits unrecorded off-ball work such as screens, box-outs, rotations, deflections and closeouts. It is not the full four-pillar DAWG model, and it does not measure motivation, toughness, complete defensive effectiveness or total player value.
Each component is counted per 40 played minutes, including overtime. The observed rate is count ÷ (minutes ÷ 40). The adjusted rate blends that record with the league population, using a Gamma–Poisson model fitted on nearby seasons (2019, 2020, 2022, 2023) and never on the season being scored. Fewer minutes means more pull toward the league average; with many minutes the adjusted and observed rates nearly agree. For high-minute passers the pull is slightly too strong: on the 2026 holdout their adjusted assist rates predicted about 5% worse than their observed rates.
Each player is judged against her own position. Guards, forwards and centres come from her team's roster listing that season (the first letter of a hybrid such as G-F), not from the stats being ranked, so a guard's blocks are compared with other guards' and a centre's assists with other centres'. The DAWG index averages these position-relative percentiles; the league index beside it uses the whole league.
The league-wide reference is frozen and era-local: REG player-seasons in 2021 with at least 160 minutes and all four components observed (123 player-seasons, reference 9496b23f8e85). League rates have changed a lot since 1997 (assists per 40 are roughly 25–35% higher today), so a player is compared with her own era, and the index is not an absolute comparison across eras. The index is the plain average of the four component percentiles, with equal weights chosen as a first-version policy rather than learned basketball values. It is an average of percentiles, not the percentile of the index itself, a probability or points added. If any component is missing, the index is not shown, and the remaining components are never re-weighted.
Turnovers, personal fouls and defensive rebounds are shown separately and are not subtracted. Their meaning depends on role and game context, so the index is a positive-activity profile, not net value.
- Covers four selected recorded box-score actions only; unrecorded off-ball work (screens, box-outs, rotations, deflections, closeouts) is omitted.
- Percentiles compare each player with her own position group (guards, forwards or centres, from her team's roster listing) in the same era; they are not absolute comparisons across eras or positions. The league-wide index is shown alongside.
- Position groups are coarse listings: a hybrid such as G-F counts by its first letter. Players with no listing are compared with the whole league.
- An average of component percentiles, not a percentile of the composite, a probability, points added or total player value.
- Equal component weights are a fixed policy, not learned basketball values.
- Adjusted rates are pulled toward the league-wide average, not the position average, so low-minute guards' blocks and offensive rebounds (and centres' assists) read slightly high within their group.
- Turnovers and fouls are shown separately and are not subtracted; the index is a positive-activity profile, not net value.
- Does not measure motivation, toughness or complete defensive effectiveness.
- Component intervals are conditional empirical-Bayes estimates with fitted hyperparameters; they ignore hyperparameter uncertainty.
- Conditional empirical-Bayes interval given fitted hyperparameters; hyperparameter uncertainty is not included.
- 10th-90th percentile of the index over 300 whole-game resamples with the reference fixed. A stability range, not a credible interval or ranking probability.
- Percentiles of players with few minutes sit closer to 50 because their adjusted rates are pulled harder; early-season numbers understate standouts.
- Edition
wnba_dawg_recorded_actions_v1 v1.2.0- Box source
- wnba_stats.player_game_logs · partition updated 2026-09-28
- Play-by-play
- wnba_stats.play_by_play · 192 of 192 games · diagnostics only
- Shrinkage gate
- Passed: on the registered 2026 holdout, adjusted rates predicted later counts better than observed rates in all four components.
- Status
- Method 1.2.0. A descriptive profile of recorded actions; see the limitations above.
