How the Oracle predicts a cricket match

Short version: it rates every team from every completed match it can see, accounts for who is at home, blends in recent form and head-to-head, and turns that into a probability. It was judged on matches it never saw during training — and every live call is locked before the start so the record can't be edited after the fact.

1 · The data

Every international and franchise match in the Cricsheet ball-by-ball archive — over 15,000 men's and women's Tests, ODIs, T20Is and T20 league games from 2002 to today — plus results we record ourselves between Cricsheet updates. Fixtures come from ESPN's live scoreboard.

2 · Team strength — an Elo rating

Each side has a rating that moves after every match by how surprising the result was: an upset moves it a lot, an expected win barely at all. Ratings are kept separately for each format and for men and women — a men's side never borrows its women's results. Bigger winning margins count for more. Home advantage is measured, not assumed: in our data hosts win 64% of decisive men's Tests, 59% of ODIs and 53% of T20Is, and the model learns how much that is worth in each format.

Associate nations start below the twelve ICC Full Members. Without that, a run of wins in a regional qualifier would lift a side above established nations it has never beaten — the backtest confirmed the prior makes predictions measurably better.

3 · The prediction

The rating gap (including home advantage) is combined with recent form and head-to-head by a model fitted separately for Tests, ODIs, international T20s and franchise T20s — these behave very differently (franchise squads reshuffle every auction). Where a segment has too little history, the model uses the calibrated rating alone rather than risk learning noise. Tests also get a draw probability: evenly matched sides draw more often.

4 · How it was tested

The model was built and tuned only on matches from 2019–2023, then scored once on every match from 2024 onwards — 4,159 games it had never seen. Each prediction used only information available before that match. Only decisive matches between sides with at least 10 prior games were scored.

Accuracy is the share of matches where the favourite won. The Brier score measures whether the probabilities themselves are honest — a model that says 70% should be right about 70% of the time. It is the number we optimise, because a confident wrong call should cost more than a cautious one.

5 · The locked-call guarantee

Every prediction is locked 48 hours before the scheduled start. The rule is enforced by the database itself: it rejects any prediction created after a match starts and any edit to a prediction once made — only the result can be recorded afterwards. Before the lock you'll see a provisional view that updates as new results arrive; provisional views are never counted. The live record includes every locked call, right or wrong.

6 · What it can't do

• It doesn't know the playing XI, injuries, pitch or weather — only results. A rested star or a raging turner can beat it.
• Franchise T20 is hard for any model (~55% in our tests): squads change every season and matches are short.
• No call when the data is thin — sides need 10+ matches in the format. Afghanistan isn't in the open Cricsheet archive, so we don't predict their matches yet.
• Domestic first-class and List A cricket isn't covered.

For entertainment and analysis only — CricMind predictions must not be used for betting or wagering. Model file and validation are regenerated on every retrain; last trained on data through 2026-10-07.