About the WNBA Model

Game predictions, a full prop board, and a public scoreboard — built from structure, not a black box.

Two models, one number per market

Win probability and spread come from a stacked model — team Elo ratings combined with opponent-adjusted offensive and defensive efficiency (points per 100 possessions) and pace, refit every season. Totals come from a separate possession-level Monte Carlo simulator: it plays out the game possession by possession using each team's pace and efficiency, so the final total is the mean of a real simulated distribution, not a linear projection.

Game pages show that distribution directly — five real percentiles (p5 through p95) from the simulation, not a smoothed curve fit after the fact.

Player props, gated

Every posted points, rebounds, assists, three-pointers, and combo prop (Pts+Reb, Pts+Ast, Reb+Ast, Pts+Reb+Ast) is priced from the same player projections that feed the game model — expected minutes and per-minute production, not book-line reverse-engineering. A prop only prices when the player has enough 2026 sample and a confirmed role; otherwise it publishes with an explicit reason instead of a guess.

The guardrails (when we refuse to publish a number)

  • No market, no claim. If no book has posted a line, we say so rather than price against nothing.
  • Roster confirmation. Games with an unconfirmed rotation carry a flag and no pick.
  • Divergence abstention. When our number and the market disagree too much to trust, we publish both and claim no edge.
  • EV cap. An edge too large to be credible usually means a stale line, not real value — we don't publish it as a pick.
  • Sample floors. Players below a minutes/sample threshold get no projection and no prop price rather than a number built on noise.
  • Injury-aware. A player listed Out is never priced as available.

A pick — on the game board or the prop board — only ever means the model actually staked it: a non-null stake and no abstention. Every abstained row stays visible with its reason; the board showing its refusals is the product's honesty surface, not a gap to explain away.

Evaluation discipline

Every published claim is walk-forward: each season is predicted using only models trained on seasons before it, refit annually, and nothing is scored in-sample. There is no clean historical WNBA closing-line archive to benchmark against, so we don't claim a market baseline in the backtest — the live closing-line-value tracker is the real market comparison: every pick is graded against the closing line once its game settles, and the scoreboard publishes once the sample is large enough to mean something.

Predictions and analysis are for informational purposes only. Nothing on this site is betting advice.
For analytical/entertainment purposes only. Not betting advice. · Model B13 · © 2026 Predictium · Follow @PredictiumAI