Under the hood
Every prediction has a paper trail.
Every number on this site was produced by a pipeline, not a pundit. Here's how it works, what goes in, and the question everyone asks: is it actually any good?
The approach
It doesn't know who “should” win. It knows who the data says will.
The model has never watched a game. It doesn't know that Kohli is clutch or that a particular captain always makes the wrong toss call. What it does know is that across 17 seasons of ball-by-ball data, certain patterns hold: teams with a fresh pace attack at certain venues win more often than their standings suggest. A side chasing under dew in Mumbai in April is a different proposition than the same side batting first at the same ground in March.
Those patterns live in the numbers. The model finds them and weighs them against each other: venue history, current form, squad strength, weather, the specific matchup between the batting lineup and the bowling attack they're about to face.
Self-auditing
The model that decides which model serves you.
Three separate model families run every prediction independently: Production (the full feature set), Lineup-only (weighted toward confirmed squads), and No-phase (ignoring powerplay/death-over splits). Before the number reaches you, a trust gate compares all three against their recent track record and picks the one that has been most accurate over the last rolling window.
If they disagree sharply, that uncertainty is surfaced. If one family consistently outperformed the others this season, it was promoted into the serving blend. The model shown in the archive is the same audited system that served the season as the data evolved.
The pipeline
Six data sources. One number.
During the season, this ran automatically before every match. No human decided what went in or out.
The question you're probably asking
OK but is it actually right?
We tracked every prediction against the actual result, re-scored after every match, and published the accuracy breakdown. You can see which model family won the season audit, how the numbers look on decisive matches versus near-50/50 calls, and what the last post-match review found.
We don't hide bad calls. When the consensus gets it wrong, and it does, the governance log says why: which signals pointed the wrong way, whether the squad data was late, where the model family disagreed. Cricket is genuinely hard to predict. The numbers reflect that honestly.
Performance data & model internals
This data is not public — request access
Accuracy breakdowns, ROC AUC, Brier scores, model comparison, and post-match learning. Submit your email to request access. Requests are reviewed manually.