How accurate is it?
The model was run on 66 men's Twenty20 internationals played in 2026 — matches it had never seen, using only data available before them. No fixtures were selected or excluded after the fact.
Where the projections landed
| Projection was | Share of matches |
|---|---|
| within 10 runs | 20% |
| within 20 runs | 42% |
| within 30 runs | 58% |
| within 40 runs | 71% |
| within 50 runs | 83% |
Close to three in five projections were within 30 runs of the actual first-innings total. About one in six was out by more than 50.
November 2026 correction: our own number was too kind
Until 1 November this page said the average error was 27 runs. Checking why our live projections had come in low, we found a leak in that measurement: each ground's average already included the match being tested, so the model had partly seen the answer. A second leak worked the same way: a few grounds have a manual baseline update, set using later results, and it was applied to matches played before it existed. With both removed the error is 30 runs, not 27. The bias stays close to zero.
The figures above are the corrected ones. One smaller leak remains: matches played at the same ground later in the year are still in its average. We say so rather than leave it out.
October 2026 update: recent batting form
In October our published T20 and ODI projections came in too low seven times in a row. Looking back over 2025–26, India and New Zealand batting first had beaten the model by 15–18 runs on average: team ratings were following form too slowly.
The model now adds a cautious correction from each side's first innings over the previous twelve months, shrunk towards zero when there are few matches. It was tuned on 2024–25 and only kept because it also improved 2026 matches it had not seen. On the same 62 matches the average error fell from 27.5 to 25.6 runs and the bias from −4 to 0 (measured before the November leak fix described below). The figures on this page include it, measured using only data from before each match. Every earlier projection stays on the results page as it was published.
Is it biased in one direction?
If a line had been set at exactly the model's projection, the innings finished above it 59% of the time and below it 41%. A perfectly centred model would sit at 50/50; this is close enough that no systematic lean is detectable at this sample size.
The average bias of -1 runs says the same thing. That matters more than raw accuracy: a model that is consistently wrong in one direction is worse than one that is imprecise but centred.
What this does not mean
An average error of 30 runs is not small. It reflects something real about the sport rather than a shortcoming that better data would fix: measured across matches where the same side batted at the same ground, totals still varied by around 34 runs. The model's error is already close to that floor, which means most of what remains is the genuine volatility of a cricket innings — an early dropped catch, a middle-order collapse, rain.
So these projections are useful for judging whether a total is likely to be passed, not for predicting a precise score. Anyone claiming to do the latter is not measuring themselves.