Legacy 2025 evaluation: model identity is not embedded in this artifact. Its inspected producer, analysis/calibration_export.py, used StackedEnsemble trained on 2016–2024. This historical probability research is separate from the Games page’s EPA-OLS or SP+ spread projections. It does not validate the current card. Historical feature availability and shuffled internal stacking folds limit prospective interpretation.
Scoring rules
SCORING RULES — 2025 SEASON
BRIER
0.1850
LOG LOSS
0.5489
ACCURACY
74.2%
N
550
Brier 0.25 and log loss 0.693 are coin-flip baselines — lower is better. Accuracy is the pick-winner rate, not the betting bar.
Brier score measures squared probability error. Log loss penalizes confident wrong predictions more strongly. Both evaluate forecast probabilities, rather than betting profit.
Pooled reliability
POOLED RELIABILITY — ALL GRADED WEEKS
Dashed line: perfect agreement. Above it, wins occurred more often than predicted; below it, less often. Small samples remain uncertain.
Each probability range groups games with similar predicted win chances. Gap = actual minus predicted, in percentage points (pp). Games is the number of graded games. Orange gaps exceed 10 percentage points in either direction.