Transparent enough to audit.
The product is designed around pre-match probability distributions rather than picks. Models are evaluated with walk-forward tests: a historical prediction can only use information that existed before that match.
Player Runs
Historical innings are weighted by recency, expected batting role and competition. A cohort prior stabilizes smaller player samples. The output is a full runs distribution rather than a single average.
P(Runs ≥ x) is derived from that distribution for each displayed threshold.
Expected batting position
The expected position is inferred from the player's most recent appearances with higher weight on recent matches. Role Stability measures how consistently the player bats within ±1 position of that expected role.
Player Fours
The model estimates a player-specific probability of a four per legal ball, shrunk toward players with a comparable batting role. It then integrates that event rate over the player's estimated balls-faced distribution.
Player Sixes
The sixes model uses the same opportunity framework, but learns a separate six-rate. A beta-binomial count model allows extra variation beyond a simple fixed-rate binomial model.
What we have tested so far
Data source
Historical match and ball-by-ball data in this MVP come from Cricsheet. The UI snapshot currently covers Big Bash League and T20 Blast. The production architecture should retain source attribution and the applicable data-license requirements.
Current MVP limitations
- No future fixture feed or confirmed playing XI.
- No opponent or venue context in the neutral player snapshots shown on the current site.
- No user accounts, watchlists or subscriptions.
- No bookmaker odds, comparisons or betting recommendations.