A fairness audit of the Duckworth–Lewis–Stern method: format-specific and gender-differential bias, with an interpretable calibration layer for cricket target revision

Abstract The Duckworth–Lewis–Stern (DLS) method has set revised targets in rain-interrupted limited-overs cricket since 1999, yet no large-scale empirical audit of its prediction bias has been published. Using ball-by-ball data for 8,150 international matches from Cricsheet, we audit DLS by sampling synthetic interruption points and comparing its resource-based projection against the runs actually scored. We find that DLS bias is far from uniform: it varies systematically with match state, over-predicting the death overs of Twenty20 and under-predicting collapse scenarios in one-day cricket, a format-specific pattern we quantify across the full space of overs remaining and wickets lost (a 137-run span of per-bucket mean bias). Our second finding concerns fairness: because a single resource table governs both men’s and women’s cricket, DLS miscalibrates women’s one-day matches relative to men’s at comparable match states, producing a gender-differential bias of several runs (a +6.13-run gap on the training split) that, to our knowledge, has not previously been documented. The gap survives match-level clustered inference, holds among matches between top (Full Member) teams, and is stable across temporal windows. We benchmark DLS against five modern learning methods and introduce DLS-Cal, a lightweight interpretable calibration layer that adds a state-conditioned correction to the published DLS prediction, reducing absolute bias by 31 % on ODI and 19 % on T20I; a gender-aware variant reduces women’s ODI residual bias from +6.19 to +0.65 runs without altering the DLS framework. We also introduce the Win-Flip Rate, a threshold-based fairness metric for target revision, and release code, trained models, and the audit dataset for reproducible research.