Every game on the slate, graded. Win probabilities from a walk-forward XGBoost model,
fair moneylines, projected spreads, and the exact factors behind each pick — verified
against final scores across the full season archive.
Exhibition regime — SL picks come from a separate
in-tournament form model (not the NBA model), and its own backtest says it has
no measurable skill: — on Summer League 2025.
Books publish no SL lines here either. Treat every SL pick as a coin-flip lean,
never an edge. Finals are verified below regardless.
Grade
Matchup
Pick
Win prob
Fair ML
Proj margin
Result
League
Matchup
Tip (local)
Pick
Win prob
Status
Score
No games match — clear filters or pick another date.
How to read the terminal
GRADE
Deterministic pick quality. conf = |P(home) − 0.5|, agree = spread model
and win-prob model point the same way. A: conf ≥ .20 & agree · A−: ≥ .15 & agree ·
B+: ≥ .10 & agree · B: ≥ .06 · B−: ≥ .03 · C: coin flip · D: models conflict.
WIN PROB
Model probability the pick side wins. From an XGBoost classifier trained
walk-forward on 2016+ seasons — features are strictly pre-game (no leakage).
FAIR ML
The no-vig American moneyline implied by the model's probability. Compare
against a book's line: model more confident than the price ⇒ potential edge.
PROJ MARGIN
Separate regression model's projected home-team margin (spread proxy).
FACTORS
Exact SHAP attribution for this game, converted to approximate win-probability
points. Positive pushes toward the home side.
PLAYER PROPS
Per-player projections (points / rebounds / assists) from dedicated XGBoost
models over trailing form, usage, rest, archetype, and opponent context.
Six markets: points, rebounds, assists, threes, steals, blocks.
Trained through 2024-25 only; verified on 2025-26: —.
A market ships a learned model only if it beats the naive trailing-7-average
baseline; markets that can't (currently blocks) publish the baseline itself,
marked with †.
RESULT
Final score verification — every archived pick is graded against what actually
happened. Nothing is retroactively edited.
Model-fair odds are not betting advice. If you bet, bet what you can afford to lose.