WNBA: a live experiment, year one

Live paper experiment — no wagers placed or recommended. Every projection on this page is locked before tip-off — fingerprinted with a cryptographic hash so it cannot be edited afterward — and graded against the final score and the closing line. Nothing is ever edited or deleted. We’re testing whether a model built for college basketball transfers to a 13-team professional league — in public, with the misses counted.

The honest prior, stated up front: in six-season backtesting this model is market-competitive on accuracy but has no established cash edge. Its known weak spots (September, large favorites) are named below and under active work. The live season — starting July 19, 2026 — is the test.

Spreads — live record vs the close

Champion model only. Graded against the closing line (the hardest benchmark); CLV = how far the market moved from our published-at opener toward the close, the experiment’s primary metric.

SliceRecordWin %Avg CLV (pts)n
All graded picks21-2347.7%-0.2144
Divergence ≥ 2 pts (cumulative)11-1150.0%-0.2622
Divergence ≥ 4 pts (cumulative)6-275.0%-0.198
Divergence ≥ 6 pts (cumulative)0

Divergence bands are cumulative (≥) and measured against the opener the pick was published into. Small samples early — percentages stabilize as the season accumulates.

Month by month

MonthRecordWin %Avg CLV (pts)n
Jul 202612-1446.2%-0.3326
Aug 20269-950.0%-0.0418

Published, awaiting grade

DateMatchupModel lineOpenerEdge
2026-08-08Las Vegas Aces @ Minnesota Lynx-2.4-4.52.1
2026-08-08Indiana Fever @ Chicago Sky+3.5+6.53.0
2026-08-08Seattle Storm @ Portland Fire-1.2-1.50.3
2026-08-09Phoenix Mercury @ Washington Mystics-3.6-1.52.1
2026-08-09Golden State Valkyries @ Los Angeles Sparks+7.5+6.01.5

Totals EXPERIMENTAL

Backtesting showed no edge on totals — they’re tracked here for the record, not as a claim. That’s what an experiment publishes: the nulls too.

Graded: 22-18 (55.0%) · avg CLV -0.23 · n=40

Today’s picks — published, awaiting grade →
Full picks log — every projection, hashed and graded →

Methodology & the honest priors

The model: opponent-adjusted efficiency ratings computed from box scores (a from-scratch WNBA equivalent of the ratings that drive our college model), the same line-assembly anatomy as our NCAA champion, with every parameter — home-court, scale, variance, seasonal weights — re-fit on four seasons of WNBA data under walk-forward rules (the model only ever sees data available before each game it predicts).

Backtest, labeled as backtest: across 2023-25 (835 games vs true closing lines) the model missed final margins by 9.96 points on average vs the market’s 9.47 — market-competitive, not market-beating. Its disagreements showed an ordered curve (52.7% at ≥2 divergence, 54.1% at ≥4, 56.3% at ≥6 on n=167) — but tournament-thin samples and a known September weakness mean we claim no cash edge. Those backtest figures never appear in the live tables above; the live season stands on its own.

Known open problems, named before the season instead of after: September performance (late-season roster states the model can’t yet see) and very large favorites. Both are under active, pre-registered improvement work; any model change during the season will be disclosed here with a date.

Every row’s pre-tip hash is re-verified when this page’s feed is generated — a mutated snapshot would be caught mechanically. Questions about the methodology are welcome; the answers are longer and nerdier than you want.