How the model works

No tips, no subscriptions, no pre-match selections. One number, updated every ball, with the test behind it published in full.

What it reads

After each ball the model is handed the state of the match and asked one question: from here, what is the probability the batting side wins? The inputs are the things a watching analyst would use:

• The chase arithmetic — runs required, balls left, current and required rate, and how far apart those two rates are.
• Wickets in hand — not just how many have fallen but who is left, since a side five down with recognised batting is in a different position from one five down with the tail exposed.
• The two batters at the crease — how they are scoring in this innings, weighted against their record, and their history against the bowler on now.
• The bowler's spell — economy and wickets in this innings rather than a career average, because a bowler having a bad night is information.
• How the pitch is actually playing — derived from the scoring pattern in this match, validated against 12,951 T20 matches, rather than a fixed venue label.
• Ground records — first-innings and chasing averages across 335 tracked venues, so a par score at one ground is not treated as par at another.

How it was trained and tested

The model is trained on ball-by-ball data from 2017 to 2024. It is then scored on 2,546 matches from 2025-26 that were not available to it during training — 19,340 predictions, 81.5% correct.

The split is by date rather than at random for a specific reason. A random split lets matches from the same season sit on both sides of the line, and the model quietly learns that season's scoring conditions. Splitting by date removes that, and the number drops to something honest. The full breakdown, including the calibration table, is here.

What it deliberately does not do

It has no pre-match tip for you

Before the first ball there is little beyond team strength and the toss, and the edge there is thin enough that publishing a confident number would be dishonest. The model is a live model. It has something to say once a game is underway.

It does not pretend to see the ball

There is no Hawk-Eye, no ball speed, no swing or spin measurement in the feeds available at this scale — only what happened to each delivery. That limit is real. Some questions, such as whether a wicket falls in one particular over, turn out to be close to unpredictable from outcome data: a properly trained model on 33,818 examples beat simply quoting the average by about one percentage point. Where that is true we say so rather than dressing up a number.

It does not quote odds

A probability is not a price. A bookmaker's price contains a margin and the weight of money staked; ours contains neither. The two are not interchangeable and we do not present them as though they are.

How to read the number

A well-calibrated 68% is supposed to be wrong about a third of the time. If every 68% call came in, the model would be badly under-confident rather than good. That is why the calibration table matters more than any single match, and why one match is never enough to judge a probability.

The number earns its keep at the moments the scoreboard is ambiguous — early in a chase, or when the run rate and the wickets in hand point in opposite directions. When a side needs 20 off 30 with eight wickets standing, everyone watching already knows; a percentage adds nothing there, and we would rather it stayed out of the way.