One number: the chance a side wins from where the match stands right now. Here is what goes into it, how to read it, and what it has actually scored.
Win probability is not a rating of how well a team has played. It is a forecast from the current position, and it holds no memory of how that position was reached. A side that collapsed to 40 for 5 and recovered to 130 for 5 is read exactly the same as a side that reached 130 for 5 serenely, because from here the two are the same problem.
Those feed a gradient-boosted model trained on roughly a decade of ball-by-ball T20 records, which is then updated over by over: each over is treated as evidence for or against the reading the model already held, so a single expensive over shifts the number without overturning it.
85% means that in a hundred matches from positions like this one, about fifteen are lost. T20 is a format where fifteen in a hundred happens often enough to watch for.
At over 6 the model scores about 79%; by over 15 it is about 82%. It knows less early because there is less to know, and it says so rather than presenting every reading with equal confidence.
A number that barely moves through a 20-run over would be wrong more often, not less. What the model avoids is reacting to noise — each over is weighed against what was expected at that stage.
81.5% of the time the favoured side went on to win, across 19,340 predictions on 2,546 matches from 2025-26 that were held out of training entirely.
| 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 |
This is the harder test and the one that matters. Group every prediction by what the model said, then check what actually happened in each group. A model that only ever backs the obvious favourite can post a strong hit-rate; it cannot post a straight calibration line.
| Model said | Predictions | Actually won | |
|---|---|---|---|
| 0-10% | 4,529 | 3.8% | 1.2 |
| 10-20% | 1,564 | 15.3% | 0.3 |
| 20-30% | 1,188 | 25.4% | 0.4 |
| 30-40% | 1,074 | 34.5% | 0.5 |
| 40-50% | 1,175 | 43.9% | 1.1 |
| 50-60% | 1,122 | 50.4% | 4.6 |
| 60-70% | 1,350 | 62.1% | 2.9 |
| 70-80% | 1,568 | 72.5% | 2.5 |
| 80-90% | 1,708 | 81.1% | 3.9 |
| 90-100% | 4,062 | 96.1% | 1.1 |
Full method and the reproduction command are on the accuracy page.
It is not a tip, a price, or advice. It carries no ball-tracking data — speed, swing and spin measurements are not available to us at this scale, and the model works from delivery outcomes only. And it says little before the first ball, because from nothing there is little to forecast.