Methods and data
Quantities
| Quantity | Interpretation |
|---|---|
| Strength | Estimated ability at a specified date and reference context |
| Contribution | Value assigned to performances in completed matches |
| Runs above replacement | Run value relative to an explicitly defined substitute |
| WAR | Expected wins above that replacement baseline |
| Win probability added | Change in modeled winning probability around observed events |
These quantities answer different questions. Each model will keep its own scale, uncertainty method, and assumptions. An estimate of zero relative to the average player does not establish replacement level.
Data and coverage
The source is Cricsheet, whose JSON documentation defines participants, registries, replacements, and outcomes. Source data remain subject to Cricsheet’s licensing terms.
The implemented preparation stage supports official JSON and the existing YAML downloads. It writes separate Test, ODI, and T20I match and participant tables, plus player identity tables. Delivery records remain in the raw source for subsequent ball-level work. The source receipt and per-file hashes identify the exact input snapshot.
Coverage is the set of available source matches, not a certified census of international cricket. Men’s and women’s source coverage may differ. The manifest reports counts separately and distinguishes exclusions after the requested cutoff from other formats. It does not infer that a complete historical archive was available at an earlier forecast date.
Participants and eligibility
Cricsheet’s players lists include replacements as well as the starting eleven. The participant table therefore makes no starting-XI claim. A match is flagged has_fixed_xi=1 only if each side lists eleven distinct people and no match replacement or supersub is recorded. The initial binary lineup eligibility flag additionally requires a decisive result. The Test ranking model also admits fixed-XI draws and ties. These are sample flags, not deletions.
Stable IDs are required for every listed participant. Missing IDs, duplicated participants, invalid outcomes, and duplicate match IDs stop preparation with the source filename. Missing participant lists remain visible as ineligible matches. Names are preserved per appearance; the identity table uses each person’s latest observed name within the cutoff. Team labels and venue labels remain source strings, so joins to external team ratings need an audited mapping.
Ties, draws, and no-results remain distinct. A tie resolved by an eliminator is retained as a tie with the eliminator recorded separately. Toss and outcome fields are separate: outcome-dependent information must not enter pre-match predictions.
Selection and match difficulty
A player who appears mainly in easier fixtures can receive an inflated rating if the model misses what makes those fixtures easy. The current model includes all 22 player identities: weaker opposing XIs and stronger teammates therefore enter the predictor, alongside the toss and batting first. This adjusts for some schedule differences, following the logic of regularized adjusted plus-minus. It does not make team selection random.
The present ratings do not separately adjust for home advantage, neutral venues, pitch suitability, injuries, player availability, or changes in underlying team conditions over the five-year window. A specialist selected only on favorable pitches, or a reserve selected only against weaker touring sides, can remain confounded with those conditions. Shrinkage addresses instability; it does not remove this selection bias.
The next comparison should add audited home/away/neutral status and venue context, and examine team strength measured strictly before each match. A team-strength baseline changes the target: player coefficients then describe departures from that baseline. It should be reported as a separate specification, with changes in rankings and chronological predictive performance, rather than treated as an automatic correction to the existing player-only rankings.
A useful diagnostic is each player’s distribution of pre-match fixture difficulty and whether alternative players actually appear in comparable fixtures. Little overlap means estimates depend heavily on extrapolation. Selection weighting would require a defensible pool of available but unselected players and measured selection predictors; playing XIs alone do not provide that information. Such adjustment still relies on no important unmeasured confounding and adequate overlap. See Goetghebeur et al. on causal questions and adjustment.
Use pre-match conditions for these comparisons. Final run margins, wickets, and match performances are consequences of the lineup and must not be added as controls for its pre-match contribution. The published rankings remain descriptive model associations, with selection bias unresolved.
Existing WAR calculations
The original notebooks used an assumed runs-to-wins divisor and a percentile or scaled-expectation replacement baseline. Mean victory margin among matches won by runs is not a marginal runs-per-win estimate. The original calculations are preserved for inspection, with their limitations, in the exploratory collection.