The model's biggest well-sampled miss: late-inning long shots
Late in MLB games, a team the model gives 0.262 to win actually wins 0.0843 of the time -- a 0.1778 calibration error and the model's largest miss in any cell with n over 500 (n=617).
How to read this figure
Each row is a (game-phase x probability band) cell; calibration_error is the absolute distance between what the model said (mean_p) and what happened (mean_y).
Published coverage
The committed artifact is partial; only its published coverage is shown here.
| Population | Status | Games | Usable buckets | Published reason |
|---|---|---|---|---|
| MLB | unknown | - | - | - |
| SOCCER_INTL | unknown | - | - | - |
Continue with the data
Open the interactive inspector: State reliabilityWhat it means
This exhibit buckets every graded in-game prediction by confidence band and game phase to find where the model drifts furthest from reality. On the segment-clean corpus the worst well-sampled cells are low-probability bands, where the model leaves too much hope on teams that are already beaten. The late-inning heavy-favorite cell that revision 1 called the biggest miss is now close to honest: the model says 0.9101 and those teams win 0.9498, a 0.0397 error over 2,369 rows. Tinier buckets look far worse but rest on almost no data.
Caveats and confounds
The very worst rows in ranked_worst_buckets have tiny n (the top one, calibration_error 0.7892, is only n=7) and are noise-dominated; the n-weighted ECE summary (mlb model 0.0494 vs market 0.0397) is the stable read.