Soccer Calibration Pack
soccer_intl corpus: n=4265 rows -- far smaller than mlb/nba, wider CIs, single-fold reads not durable. Murphy: model Brier 0.3325 vs market 0.1952 (gap=+0.1374); market leads. Minute-bucket ECE (weighted): 0.3748.
How to read this figure
The Murphy decomposition splits Brier into reliability and resolution; the disagreement buckets show who is closer when model and market pull apart; the minute buckets show calibration error over the clock.
Continue with the data
Open the interactive analysis: Minute-Bucket Calibration SupportWhat it means
When model and market diverge by 0.10 or more (n=2291), the model is the closer forecast only 10.48% of the time. Late-game calibration also drifts, with expected calibration error climbing to 0.4824 in the 75-91 minute bucket. This is a published null, not a win. Segmentation cut the corpus from 9,003 rows to 4,265 and dropped the draw share from about 26% to 14% of ticks; a draw is mechanically cheap to score against a two-way probability, so both sides post a higher Brier here than they did in revision 1.
Caveats and confounds
The soccer_intl corpus is only 4,265 rows across 27 games, far smaller than MLB or NBA, so confidence intervals are wide and single-fold reads are not durable.