Where the numbers come from

The pipeline pulls the full PrizePicks NBA board nightly, then prices every prop itself from real data: full-season game logs for each player, schedule context (rest, back-to-backs), team defense and pace, injury reports, and player-prop lines from major sportsbooks.

The projection

Each player is projected as per-minute production × projected minutes, with recent games weighted more than old ones (exponential decay). Counting stats with skewed distributions (threes, steals, blocks) use a Poisson tail instead of a bell curve. The projection is then shrunk toward the sportsbook consensus line — the market is the strongest single signal in sports.

Grading & calibration

Every pick is archived and graded against box scores the next morning — wins, losses and avoid-calls all stay on the Record tab. The probability you see is calibrated on 12,000+ graded picks: an isotonic layer maps the model's raw confidence to how often picks like it actually hit, refit automatically every night.

Tiers & lineups

CORE requires calibrated probability, edge and recent form to all clear bars; FLEX and WATCH are progressively looser; PASS/FADE are avoid signals (low win rates there mean the model was right). Lineups headline the slip-level win estimate — the chance every leg hits, much lower than any single leg — and a payout-aware EV estimate (goblins pay less, demons more).

Estimates, not betting advice. Always verify the live PrizePicks line and payout before entering anything.