Win probability publishes once the 2027 model passes its held-out validation. Until then this page describes the Cricsheet data, the planned 2027 baseline, the release gates, and how a public number will be derived.
Data sources. Ball-by-ball data: Cricsheet (licence to be confirmed).
The 2027 model will train on the checked-in Cricsheet corpus, not on live request-path scraping. The current archive covers 1,243 matches and 2,95,732 balls across IPL seasons 2008 to 2026, including 74 matches from 2026.
No model is in public use. The planned baseline is team Elo, a home indicator, a chasing indicator, venue chase bias, and Platt calibration, with separate pre-toss and post-toss variants. It ships only after the release gates pass.
A model is released only after an earned pass on fixed held-out seasons. Gates are not weakened to force a ship.
22 completed matches reviewed. Pre-match accuracy 77% with 17 correct and 5 incorrect. 1 no result excluded.
When a model is released, each published fixture will store two values at publication time: the calibrated probability used for scoring, and the public number shown to readers. The track record never retrofits either value after the result.
The public reading is a transform of the calibrated value, not a second model. It widens the calibrated reading away from a 50-50 split, then stays inside a 55 to 78 public band so the page always names a side without pretending certainty.
A separate live model will print a chase-oriented probability after the first-innings powerplay, with the published live number starting once the target is known. It is shown only from innings states that have cleared their own validation checks. Live numbers are not widened.
Unsupported numbers stay unpublished. The match page still carries schedule, squads, venue, form, toss context, and head-to-head history while the model is unreleased.
