How it works
No black box. Here's exactly how every number is built — and where it can be wrong.
Where the data comes from
Player game logs, advanced stats (usage, pace, minutes), season shot charts and opponent shooting-by-zone come from the official league stats feed, refreshed every night and again pre-tip. Player-prop lines (points, rebounds, assists, threes, steals, blocks) come from a multi-book odds feed (DraftKings, FanDuel, Fanatics, and more).
The projection, step by step
1 · Minutes. The biggest driver of any counting stat. We use a robust median of recent minutes so one blowout or foul-trouble night doesn't distort it.
2 · Rate. A blend of recent and season per-minute production, so a hot or cold streak moves the number without owning it.
3 · Matchup. For shooting markets we nudge the projection by how the opponent defends, from their opponent-shooting-by-zone profile (bounded, so it informs rather than dominates).
4 · Regression to the line. We pull the raw projection toward the market line by sample size — (n·model + k·line)/(n+k). A player with 2 games barely moves off the line; a player with 30 stands on the model. This is the single biggest guard against fake edges, and we regress harder in the WNBA because the season is short.
5 · Distribution. Counting stats are overdispersed (variance > mean, because minutes swing and game script correlates events), so we fit a negative-binomial distribution — not a single number — and read the probability of going over the line off it. The 50% range you see is that distribution's middle half.
From projection to edge
We strip the vig out of the book's two-sided price to get the no-vig market probability — the fair benchmark. Props carry heavy juice (often 8–15%), so this step is essential; skip it and you'll "find" edge that is really just the hold.
- Edge = our probability − the no-vig market probability.
- EV = expected return per unit at the best available price.
- Stake = quarter-Kelly, capped at 1–2% of bankroll — props are high-variance and our probability is itself uncertain, so we bet a fraction of "optimal."
- Grade & confidence turn edge and sample size into an at-a-glance read. Every player card has a "why this number" drawer showing the raw model number, the regression, the rate and the matchup factor.
Shot chart + defensive overlay
Each player's shot chart plots every attempt and colors the court zones. Toggle it: Shooting shows where and how well the player shoots vs league; Defense shows how tonight's opponent defends each zone; Matchup combines them to highlight where this player meets a soft defense.
Matchup Fit
The Fit page inverts the question: for every defense tonight, it finds the zones they're worst at, then ranks the opposing players whose shot profile best exploits those soft spots — volume in the soft zones × how well they shoot there × how soft the defense is.
The honest limitations
- Small samples. The WNBA season is ~40 games and we launch partway in. Early-season numbers are noisy; that's exactly why we regress hard to the market.
- No play-type data. Synergy play-type and player-tracking feeds aren't published free for the WNBA, so the WNBA build can't show "plays the player runs" the way the NBA build will. We surface only what we can stand behind.
- Not yet backtested in public. A walk-forward backtest with closing-line value and calibration is the real proof and is in progress; until it's published, treat these as well-reasoned model-vs-market disagreements, not certainties.
- Forward CLV is the test. Beating the closing line over hundreds of bets — not any single night — is what tells you a model is real.
For entertainment and informational purposes only. 21+. Please bet responsibly.
← Back to the Board