Skip to content
DunkScope

Transparency

Model performance

Every figure below is computed automatically from graded predictions, at the moment you load this page (Fri, Oct 9, 8:37 AM ET). Nothing is edited by hand. How to read these numbers

Live predictions, 2026-27 season

No graded live prediction yet.

Predictions are graded automatically once each game is final.

Backtest, 2023-24 to 2025-26

3959 games, tested season by season

Out-of-sample predictions made season by season, each one with models trained only on earlier seasons (backtest run 3). These seasons were inspected once during model development, so the live season above is the cleaner test.

  • DunkScope model

    Log loss
    0.6051
    Brier
    0.2090
    Accuracy
    66.3%
    Calib. error
    0.018
    vs Elo
    −0.0059
    95% CI −0.0096 to −0.0021

  • Elo baseline

    Log loss
    0.6110
    Brier
    0.2117
    Accuracy
    66.1%
    Calib. error
    0.027
    vs Elo
    —

  • Home-team baseline

    Log loss
    0.6892
    Brier
    0.2480
    Accuracy
    54.9%
    Calib. error
    0.021
    vs Elo
    +0.0782
    95% CI 0.0680 to 0.0888
Backtest: calibration of the DunkScope modelpredicted 9.3%, observed 0.0% (5 games); predicted 16.8%, observed 11.9% (118 games); predicted 25.2%, observed 23.3% (373 games); predicted 35.0%, observed 33.6% (456 games); predicted 45.3%, observed 49.4% (544 games); predicted 55.2%, observed 53.7% (758 games); predicted 65.0%, observed 62.8% (640 games); predicted 74.8%, observed 75.4% (625 games); predicted 84.0%, observed 84.0% (425 games); predicted 91.2%, observed 93.3% (15 games)0%0%25%25%50%50%75%75%100%100%
Predicted probability (horizontal) vs how often the home team actually won (vertical). Bins with fewer than 10 games are hidden.
Log loss by season
SeasonDunkScopeEloHome team
2023-240.61040.61410.6901
2024-250.60660.61420.6902
2025-260.59820.60470.6874

How to read it

  • Log loss and Brier score measure how good the probabilities are (lower is better). They are the model’s main yardsticks.
  • Accuracy counts how often the favorite won. It is reported, never optimized alone.
  • Calibration error compares predicted probabilities with observed frequencies: when the model says 70%, the team should win about 70% of the time.
  • vs Elo is the average difference in log loss with a classic Elo rating (negative means better than Elo), with its 95% confidence interval.

Probabilities, not guarantees. A team given 60% loses about 4 times in 10 when the model is well calibrated. DunkScope explains model outputs and gives no betting advice; betting is for adults only (18+ or your local legal age).