Classifying Winning Performances in International Women’s Rugby Union

Abstract

Purpose: The efficacy of isolated and relative performance indicators (PIs) has been compared in rugby union; the latter more effective at discerning match outcomes. However, this methodology has not been applied in women’s rugby. The aim of this study was to identify PIs that maximize prediction accuracy of match outcome, from isolated and relative data sets, in women’s rugby union. Methods: Twenty-six PIs were selected from 110 women’s international rugby matches between 2017 and 2022 to form an isolated data set, with relative data sets determined by subtracting corresponding opposition PIs. Random forest classification was completed on both data sets, and feature selection and importance were used to simplify models and interpret key PIs. Models were used in prediction on the 2021 World Cup to evaluate performance on unseen data. Results: The isolated full model correctly classified 75% of outcomes (CI, 65%–82%), whereas the relative full model correctly classified 78% (CI, 69%–86%). Reduced respective models correctly classified 74% (CI, 65%–82%) and 76% (CI, 67%–84%). Reduced models correctly predicted 100% and 96% of outcomes for isolated and relative test data sets, respectively. No significant difference in accuracy was found between data sets. In the relative reduced model, meters made, clean breaks, missed tackles, lineouts lost, carries, and kicks from hand were significant. Conclusions: Increased relative meters made, clean breaks, carries, and kicks from hand and decreased relative missed tackles and lineouts lost were associated with success. This information can be utilized to inform physical and tactical preparation and direct physiological studies in women’s rugby.