Method

Style and quality are kept as separate vectors: similarity runs on what a player does, quality ranks the result. The aging curve uses the delta method on season-de-trended goals added (the league's expansion moved the baseline by as much as real aging). Projections are Marcel-style: weighted recent form, regressed by sample size, aged. Cap-implied wages distribute the league's published wage pool in proportion to goals added above replacement — the NWSL publishes no salaries, so the model answers what a player should command, not what she is paid.

Across 1,014 held-out player-seasons (targets 2019 and 2021–2026) the shipped projection cuts RMSE against last season repeats by 17.3% (18.9% minutes-weighted), with skill growing from 13.9% at under 900 prior minutes to 24.5% at 2,100+ — exactly the pattern shrinkage should produce. A positional-mean baseline actually scores below the naive line, so the signal is player-specific. Style neighbours chosen a season ahead sit 12.4% closer to a player's next-season profile than random same-position peers, winning 73% of the time.

We removed a component because the data said to. The classic Marcel aging adjustment cost 2.8 points of skill. A damping sweep settled it: skill falls monotonically as the age weight rises (17.1% at a quarter weight, 16.5% at half, 15.7% at three-quarters, 14.5% at full) in every age band including 31+. No fractional weight beat zero, so the term carries no point-forecast signal and is now disabled. The aging curve remains real and is still published as a description of the league — it simply is not a forecast.