What ball-by-ball cricket predicts beyond the scoreboard

Given the state of a T20 match — score, wickets, balls left, the target, who is at the crease and how long they have been there — how well can you price the next ball and the result, and how much more do you gain by modelling the players themselves? The second half is the expensive decision: a hierarchical model with a parameter per batter and bowler is easy to propose and costly to maintain, so the study measured its value before building it.

It is a measurement problem more than a modelling one, and measurement problems fail quietly: a leak or an optimistic interval produces a number that looks like a discovery. So most of the engineering went into making the measurement trustworthy.

A four-rung ladder, each rung scored against the one below it: a marginal baseline, then models of increasing capacity, up to gradient boosting on 27 whitelisted features of the match state. Every rung is fit on the training period only, tuned and calibrated on a later validation period — isotonic regression for win probability, temperature scaling for the next ball — and scored once on a held-out test period that no fitting decision ever saw.

On win probability the top rung reaches a test log-loss of 0.490 against 0.693 for the base rate — 29% lower, over 1,489 held-out matches. On the far harder next-ball task, eleven outcome classes, it is 4.17% better than the baseline. Fig. 1 is that model, unchanged, scoring a match from the test period.

Leakage discipline in a research repository is usually a convention, and conventions decay under deadline. Here it is a property of the code. The feature builder’s first statement is a select against a frozen whitelist, so an outcome column is unreachable by construction, and a test proves it by asserting byte-equality between a clean frame and one with label-bearing columns injected. The data loader raises on any read of the test split unless the leaderboard runner passes an explicit flag, so the single test touch is enforced rather than remembered.

The leakage canaries have to prove their own power before they are believed. The shuffled-identity test first requires that a planted identity signal on a synthetic split is detected, and only then that shuffling destroys it — a canary that cannot detect anything passes every leak. A poisoned-column canary and a temporal-split check run in continuous integration alongside it.

The corpus is an exact replay of 22,211 Cricsheet files through a pure transition function, pinned by hash and rebuilt byte-identically. A file that cannot be replayed is quarantined with a coded reason, never dropped — 5,457 of them, mostly formats outside the study’s scope — and 11 golden fixtures pin one real pathology each: a super over, a revised target, penalty runs, a concussion replacement.

Identity is worth 0.31% in log-likelihood and a per-match latent for pitch and conditions 0.024%, both below the materiality bar, so the per-player model was not built: the measurement fell short of the bar that would justify it, rather than instinct deciding.

The evaluation protocol was fixed before the test split was read; the verdict bands were entered with the evaluation itself. The register below draws that line exactly.