A live win probability for every ball of a UK T20 match, from a model tested on 19,340 predictions across 2,546 matches it had never seen during training.
Most cricket “prediction” sites publish a tip: a team name, sometimes a price, and no way to check whether the last hundred tips were any good. This is a different thing. It publishes one number — the probability that a side wins from the exact position on the field right now — and it publishes the record of how that number has performed.
The number updates as the match does. A wicket in the seventeenth over of a tight chase moves it a long way; a dot ball in the fourth barely moves it at all. That is not a design flourish, it is what the historical data says those events are worth.
All eighteen first-class counties, North and South groups, through to Finals Day at Edgbaston. The Blast is the hardest UK competition to read from the scoreboard alone: grounds vary enormously, and the gap between the highest and lowest scoring of them is wider than in any other T20 competition. The model carries a scoring record for each ground rather than one national average.
Both the men's and women's competitions. The 100-ball format is not simply a shorter T20 — five-ball overs and ten-ball spells change how a chase is paced, and the site handles the ball-count arithmetic separately rather than pretending it is a 20-over game.
Home fixtures at Edgbaston, the Ageas Bowl, Old Trafford, Trent Bridge, the Kia Oval and Lord's, plus England away.
Across 10 Blast grounds and 85 matches of recorded data, the first-innings average runs from County Ground, Bristol at 187.7 down to County Ground, Hove at 125.3 — a spread of 62 runs between two grounds in the same competition. That is more than a whole powerplay, and it is why a single national par score reads a good total as a bad one at one ground and the reverse at another.
A par score is a local fact. Across 2,546 matches of test data, the gap between the highest and lowest scoring grounds in the same competition is routinely thirty runs an innings — more than a whole powerplay. A model that treats 165 as “average” everywhere will read a good total as a bad one at one ground and the reverse at another. Ground records for 335 tracked venues sit behind every prediction.
81.5% of the time, the side the model favoured went on to win — measured on 2025-26 matches held out of training entirely, not on the data it learned from. Broken down by how far into an innings the reading was taken:
| Reading taken at | Accuracy | |
|---|---|---|
| After Over 6 | 79.2% | 5,036 |
| After Over 10 | 82.1% | 4,926 |
| After Over 12 | 82.6% | 4,820 |
| After Over 15 | 82.2% | 4,558 |
Direction accuracy is the easy half. The harder question is calibration: when the model says 70%, does that happen roughly 70% of the time? That table is published in full on the accuracy page, and it is the one a sceptic should read, because a hit-rate can be inflated by only ever backing the obvious favourite and a calibration curve cannot.
It has no ball-tracking data — no speed, no swing, no spin measurement. Those are not available at this scale and we would rather say so than imply a depth of information that is not there. It works from delivery outcomes: runs, wickets, who is at the crease, who is bowling, and what the ground has historically done.
It also says very little before a ball is bowled. A confident-looking pre-match number would be decoration.
Live predictions · Vitality Blast · The Hundred · The accuracy record · How the model works