DISTRIBUTION
Live board. Empirical EWM + NegBin. The daily board runs an empirical estimator: exponentially weighted game-log average, blended with a usage-adjusted rate, scaled by an opponent-allowed factor, with a negative-binomial or normal spread fitted from the league's own game-to-game variance. NFL means are calibrated on 2021-2025 outcomes (accuracy only: the pool of quoted players is over-projected by 3-7%). It does not model injuries, lineups, minutes, weather or teammate absences, and it has not been validated against prices.
Research model. XGBoost + NegBin, PIT-recalibrated
WHAT THE LIVE BOARD LOOKS AT
- Exponentially weighted game-log average
- Usage-adjusted rate (minutes)
- Opponent-allowed factor, shrunk and clipped
LIVE BOARD ACCURACY
SAMPLE
2024 and 2025 seasons in the game-log master
MAE (ENGINE VS L5)
1.561 vs 1.632
MAE CHANGE
-4.4% (se 0.2%)
empirical_ewm mean, no opponent factor; as-of history (strictly earlier games); baseline = trailing-5 mean. Accuracy of the projection only; it says nothing about beating a price.
RESEARCH MODEL VALIDATION
SAMPLE
4 rolling-origin test seasons (2022-2025), 186,447 played player-games in the master
MAE (MODEL VS BASELINE)
1.319 vs 1.339
CRPS (MODEL VS BASELINE)
0.907 vs 0.931
Result. 1.5% lower MAE; no market edge claimable
Against prices. No edge claimable. Accuracy gains over a rolling mean are 0.5 - 2.7% of MAE, the kind of gap a book's own projection already contains. Minutes and availability are the main driver the model does not have.
Record. No graded history is published for NBA projections: they are not logged to the public ledger, so there is no win rate or ROI to show.
Source: NBA props evaluation, 2026-10-14 (CRPS: 2025-26 season; Brier / ECE: proxy lines, PIT-recalibrated). MAE and CRPS: lower is better. The live board does not run the research model.
Assists per game from recent rate and minutes.