Evidence & method
How the numbers were built and tested
Every figure in the app, checked on matches the model never saw. ← Back to the app
01
Team audit
Win chance each team gained from its own challenges, in wins (100 points = one win).
Teams with at least 15 challenges. Brackets show the plausible range (95% interval). "Expected" is the success rate an average team would get with the same mix of challenges.
02
Can you trust the numbers?
Trained on matches before 2024, graded on 2,813 later matches it never saw. Lower error scores are better.
When the model says 70%, does the team win 70% of the time?
Our model vs simpler alternatives (2024–26 matches)
What makes a challenge more likely to succeed?
Right of the 1× line = more likely to succeed; left = less likely. Bars show the plausible range; faded rows are not clearly different from 1×. (Logistic regression on all 5,132 challenges.)
Can a model predict which challenges will win?
Does the main finding survive tougher assumptions?
Share of real challenge decisions where a team needed less than 10% confidence — even if a saved challenge were worth three times our estimate.
03
Method & limits
The bet, in one line
Challenge if your confidence × what the call is worth is bigger than your doubt × what a lost challenge costs.
Rearranged, the minimum confidence is p* = (1 − u)L / (G + (1 − u)L), where G is what the call is worth, L what a failed challenge costs, and u the share of failures that are "umpire's call" (challenge kept).
How the pieces are measured
What a call is worth: the drop in win chance from losing that batter, from a model trained on 8,897 matches — think of a TV win-probability line. What a failed challenge costs: how much win chance teams typically gain later from the challenges they still hold, measured from real matches. Challenges left: rebuilt from the order of challenges in each innings; it matches the IPL's 2022 rule change from one to two.
What this does not claim
We cannot see calls nobody challenged, and we have no ball-tracking data, so the player's own read of the ball is the missing ingredient — the tool asks for it rather than guessing it. All numbers are estimates, and patterns in behaviour are not proof of cause and effect.
Contains information from Cricsheet, made available under the ODC Attribution License 1.0. Not affiliated with or endorsed by Cricsheet, the ICC or any league. Full methodology: docs/METHODOLOGY.md.