Why Attribution
Largest calibrated win-prob drop: mlb mid(inn4-6):.8-1 -> late(inn7+):0-.2 = -0.8985 (min support n=1462). Any in-game move decomposes into the adjacent-time bucket transition it crossed.
Data integrity
This timing measurement is revision 2, rebuilt on the segment-clean corpus (2026-09-17). The stored corpus contains 178 MLB and 27 international soccer games. Individual measurements can use fewer games after eligibility checks; read their own denominators. Revision 1 values are withdrawn and kept in the timing regeneration receipt.
- why_attribution
How to read this figure
Each row is a transition between consecutive time buckets; the delta is mean_y(B) minus mean_y(A), the realized (calibrated) win rate of each state.
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 contrastsWhat it means
Any in-game probability move can be decomposed into the adjacent time-bucket transition it crossed, with impact measured as the difference in realized win rate between the two states rather than an observed frequency. Soccer's biggest drop is a 15-30 minute .4-.6 team falling to .2-.4 by 30-45, worth -0.775.
Caveats and confounds
mean_y is treated as the calibrated win prob; transitions are restricted to consecutive buckets with no skipping; buckets thinner than n=30 are dropped. NBA is skipped because only a checkpoint-Brier file exists, with no per-state realized win rate to attribute. These are revision 2 numbers: the source state buckets were rebuilt on the segment-clean corpus, which is why the largest drop is now an MLB transition on n=1462 rather than a soccer one on n=120.