Schemer Soft launch · daily batch

Method

Pregame game predictions

Method
m3-v0
Version
m3-v0
As of
2026-10-08
Family
Models (game and rating)
Grain
game (pregame: one home win probability and one expected margin per game)

Cohort rule

No cohort rule applies to these stats.

Stats this method computes

KeyNameDefinitionUnitTransform
pregame_wp_homePregame home win probability (Pre WP)The home team's chance of winning, stated before tip from the pregame Elo difference, recent form, rest and home court; an analytics estimate with a published calibration record by season and decile.probability (0-1)
exp_marginExpected margin (Exp margin)The expected final margin, home minus away, stated before tip, with an 80 percent interval from the model's training residuals.points
cal_brierBrier score (Brier)The mean squared error of a win probability against the result (1 for a home win, 0 otherwise). Lower is better; a constant 0.5 scores 0.25. Graded per game on a slate, per holdout season on the calibration record.squared probability error
cal_log_lossLog lossThe mean negative log likelihood of the results under the stated win probabilities. Lower is better; it punishes a confident miss harder than Brier does.nats
cal_margin_rmseMargin RMSERoot mean squared error of the expected margin against the final margin, home minus away, over a holdout season.points
cal_interval_coverage80 percent interval coverage (80% coverage)The share of final margins that fell inside the stated 80 percent interval. A calibrated interval covers about 80 percent.share of games
cal_max_decile_gapMax decile gapThe largest absolute gap between the mean predicted home win probability and the observed home win share across the ten equal-count deciles of a season.probability
cal_train_windowTraining window (Fit on)The seasons the holdout season's model was fitted on: every season from the start of the window through the one before the holdout, so no game it scores was in the fit.seasons
cal_hc_brierHome-court Brier score (HC Brier)The Brier score of the home-court baseline, which gives every home-floor game the training seasons' home win rate and a neutral-site game 0.5. The model has to beat it in every holdout season.squared probability error
cal_elo_brierElo Brier score (Elo Brier)The Brier score of the elo-winprob model's pregame win probability over the same holdout games, shown for reference.squared probability error
cal_hc_log_lossHome-court log loss (HC LL)The log loss of the home-court baseline's win probabilities: the training seasons' home win rate at every home-floor game, 0.5 at a neutral site.nats
cal_hc_margin_rmseHome-court margin RMSE (HC RMSE)Root mean squared error of the home-court baseline's margin (the training seasons' mean home margin, 0 at a neutral site) against the final margin, over a holdout season.points
cal_margin_maeMargin MAE (MAE)Mean absolute error of the expected margin against the final margin, home minus away, over a holdout season.points
cal_margin_errorMargin errorThe expected margin minus the final margin, home minus away, for one graded game. Positive means the home team did worse than expected.points

This page carries the definitions only. The full method note, its coverage and its calibration record publish with the data pages. Every value a data page shows will name this method and its version. All methods