W-LNet+ - How It Works

W-LNet+ treats the player as the fundamental unit. Instead of rating five-man lineups - a 15-team league doesn't repeat enough distinct fives to fit them directly - it estimates each player's offensive and defensive loading straight from every possession she is on the floor for, jointly with everyone else out there. A team's rating is the usage-weighted sum of its players' loadings. Net = offense + defense, in points per 100 possessions against a league-average opponent - the same number the matchup sim and the nightly game projections run on.

It is the same engine as the mens LNet+: the Pyro/SVI model, the recency weighting and the team rollup are imported from the CBB fit untouched. What differs is the data it's pointed at and a short list of deliberate omissions, below.

Fifteen franchises playing a short season produce far too few repeated five-man lineups to rate lineups directly. Estimating players instead pools every possession a player appears in, so even a combination seen twice still contributes.

Bias A player who only ever shares the floor with the same four teammates is hard to separate from them - collinearity, and in a small league it bites harder than in Division I.

Each possession conditions on the five defenders actually on the floor, so a rating is adjusted for exactly who was guarding - not a season-long team-defense number applied uniformly.

Bias Small samples against specific lineups are noisy; the prior absorbs most of it.

During the season possessions decay with age (exp(−days/τ), τ = 90 days) so the rating tracks the team a coach has now. Once the season is complete - after October 31 - the fit switches to flat weighting, because a finished season should be judged whole rather than tilted toward its last month.

Bias Early in a season the rating leans on the prior until enough recent possessions accumulate.

Returning players are seeded on last season's loading shrunk to 75%, the same shrink the mens fit validated. The mens model also seeds freshmen from recruiting rank; the WNBA has no equivalent public tiering, so a rookie simply starts at league average and earns her number from possessions.

Bias Rookies and first-year arrivals are prior-flat, so their early-season numbers move fast and mean little until the sample builds.

The mens rating carries three tuning layers the womens one does not: a zero-sum box-production prior, partial pooling of defense toward the team, and skill weighting. Those were fitted and validated against Division I data, and shipping them here would mean importing mens-tuned constants into a different league. W-LNet+ v1 is the base model without them.

Bias Offensive collinearity gets no box-prior attack, and individual defense is fully individual - so a starter who never sits can absorb credit that belongs to her unit. Read defensive numbers on thin rotations with suspicion.

The fitted loadings are not constrained to sum to zero, so the raw board drifts: on one 2026 snapshot it read mean offense 106.3 / mean defense 97.2, a +9.2 average net, with 13 of 15 teams positive. In a closed league every point scored is a point allowed, so the true average net is 0. Two uniform transforms fix it - CENTER subtracts the possession-weighted mean loading from each side, and ANCHOR shifts the offense/defense columns onto the league's real points per 100.

Bias None to the ordering - both are uniform shifts, so ranks and gaps are untouched. It only makes the numbers mean what the label says.

The rating doesn't know guard from post - a +5 is a +5. That surfaces undersized creators and stretch bigs without a positional thumb on the scale.

Bias Loses position context: +5 means different things for a lead guard and a rim-running five.

Offense positive = points added per 100. Defense positive = points prevented per 100, flipped so good always reads positive. Net = offense + defense.

Bias None - display ergonomics.

• Adjusted Off / Def / Net rating (team pages, Team Rankings) - a KenPom-style iterative opponent adjustment solved from per-game offensive and defensive efficiency. It is a separate estimate from W-LNet+, computed without the possession stream, which is exactly why the fit checks its own team net against it. The mens version additionally weights possessions by leverage; the WNBA one adjusts the stored per-game efficiencies directly.
• RAPM (player pages) - classic two-way ridge regularized plus-minus, one row per possession, scaled to per 100. Ridge shrinkage toward zero is what makes it estimable at all despite collinearity. Defense is sign-flipped on the way in so higher reads better. When RAPM and W-LNet+ disagree, W-LNet+ is the canonical number.
• The matchup sim (/wnba/sim) - the mens sim's core: stored team ratings, injury toggles priced on a replacement basis through the same team-rating math, opponent-adjusted pace, and a central-limit possession simulation. The CBB-only matchup and volatility terms are not applied here.
• Graded game projections (on the Team Ratings page) - a nightly loop projects upcoming games from the stored ratings and grades every past projection against the final score. Its sim math is lockstep with /wnba/sim. Because a game re-projected on three days keeps all three snapshots, the record reads only the latest pregame snapshot per game.

• Net = +15  → this player adds 15 points per 100 possessions, offense plus defense, versus a league-average opponent. It is the same loading the sim sums into a team rating.
• Team +6.0  → six points per 100 better than a league-average team. After re-centering (#6) the board averages zero, so the sign is meaningful on its own.
• Sample size  → on-floor possessions. A full-time starter clears a couple of thousand in a season; below a few hundred, treat a number as a rumor.

Two checks run continuously rather than being quoted once here. The fit reports the correlation between its team net and the independently-computed adjusted ratings, which share no code and no inputs beyond the games themselves. And every pregame projection is graded against the final score, so the straight-up record, Brier score and margin/total errors on the Team Ratings page are a live scoreboard for the rating, not a backtest we picked. A Brier of 0.25 is a coin flip; lower is better.