Three pure models, no I/O. Elo with per-sport home advantage and margin of victory, the market's own no-vig consensus, and a weighted blend. Every forecast is settled later and scored, so a model earns its place with numbers.
Each game was predicted with only the ratings available before it, then settled. Brier is mean squared error of the probability (lower is better, 0.25 is a coin flip). The market's closing line is the bar to beat.
StandingsBuilder from events, Elo from EloRatingSystem.ratings(after:), Pythagorean expectation from points for and against. Click a heading to sort.
K-factor and home advantage differ by sport (NFL K 20 / home 55, NBA K 20 / home 100, NHL K 6 / home 50, MLB K 4 / home 24, college K 30). Margin of victory is log-scaled with the autocorrelation correction; ratings regress a third toward the mean between seasons.
PredictionEvaluator.scorecard buckets forecasts into deciles for the calibration plot and reports closing-line value, the only edge that survives contact with a sportsbook.
Logistic regression on stat differentials, pace and efficiency models for basketball, spread and total projections, and replaying stored snapshots to backtest at the opening and closing lines.