Thirty rain stops in one cricket match are still one match

When rain shortens a cricket match, the Duckworth–Lewis–Stern method (DLS) revises the chasing team's target from its remaining resources: overs and wickets. A new audit asks a basic governance question. Does one resource table make comparable errors across formats and across men's and women's cricket?

Roy's answer is no. The paper finds structured state-dependent error, including an approximately 6.1-run women-minus-men gap in mean ODI over-prediction on its training period. But its best methodological lesson lies under the headline: 30 simulated rain stops from one innings share the same teams, venue, and scoring path. They are useful measurements, not 30 independent matches.

An ODI is a one-day international with up to 50 overs per side; a T20I allows 20. DLS assigns a resource percentage Z(u,w) to u overs remaining and w wickets already lost. The audit reconstructs the published resource table, samples an interruption during a completed first innings, and compares predicted remaining runs with the runs actually scored afterward.

Prediction error is predicted minus actual remaining runs. Its mean is bias: positive means over-prediction. A standard error measures how uncertain that mean is. Treating every sampled interruption as independent lets repeated views of the same trajectory shrink the standard error artificially.

def match_clustered_standard_error(rows):
    groups = {}
    for match_id, error in rows:
        groups.setdefault(match_id, []).append(error)
    means = [sum(xs) / len(xs) for xs in groups.values()]
    center = sum(means) / len(means)
    variance = sum((x - center) ** 2 for x in means) / (len(means) - 1)
    return (variance / len(means)) ** 0.5

The paper uses a stronger version of this idea: a 2,000-replicate block bootstrap that resamples whole matches. The compact code above shows the equal-cluster analogue; it is not a reimplementation of the paper's interval or a substitute for its raw match data.

Matching match-state buckets does not erase the ODI result. Nor does restricting to Full-Member teams: the paper reports +6.72 runs with a match-clustered 95% interval of [+3.38, +10.04]. It still calls the finding an observational association, not a causal effect of gender.

DLS-Cal predicts the residual—actual remaining runs minus the DLS prediction—and adds that correction back to the baseline. Its small multilayer perceptron uses match state, format, recent scoring, and a bucket-mean prior. The result stays decomposable as “DLS plus correction,” unlike a five-model stacking ensemble.

Accuracy and auditability do not pick the same winner. Stacking has the best RMSE at 21.49. DLS-Cal trails by 2.23 runs but has the lowest ODI WFR₅ and an inspectable correction. A gender-aware version reduces women's test-period residual bias from +6.19 for canonical DLS to +0.65 runs, while men's calibration changes little.

The paper's Win-Flip Rate at threshold k is simply the share of predictions whose absolute error exceeds k runs. It focuses attention on decision-relevant misses, but the paper explicitly calls it a proxy: an error over five runs does not prove the chase winner actually changed.

• Obtain official historical DLS decisions and evaluate actual revised targets at exact stoppage times.
• Weight synthetic match states by the empirical distribution of weather interruptions, venue, season, and geography.
• Report match-clustered uncertainty for every subgroup and state bucket, not only marginal averages.
• Validate whether WFR thresholds correspond to actual outcome flips and whether that relationship differs by format.
• Audit separate calibrators across era, region, team strength, and ball type before treating gender routing as a deployment answer.

The XGBoost page explains the strongest single tree baseline in the comparison, while neural networks provide the residual calibrator's building blocks. The general lesson travels far beyond cricket: when many rows come from one real-world unit, the row count measures computation; the unit count measures evidence.