Descriptive conversion of performance indicators in rugby union

Abstract

Objectives: The primary aim of this study was to examine whether accuracy of rugby union match prediction outcomes differed dependent on the method of data analysis (i.e., isolated vs. descriptively converted or relative data). A secondary aim was to then use the most appropriate method to investigate the performance indicators (PI's) most relevant to match outcome.

Methods: Data was 16 PI's from 127 matches across the 2016-17 English Premiership rugby season. Given the binary outcome (win/lose), a random forest classification model was built using these data sets. Predictive ability of the models was further assessed by predicting outcomes from data sets of 72 matches across the 2017-18 season.

Results: The relative data model attained a balanced prediction rate of 80% (95% CI - 75-85%) for 2016-17 data, whereas the isolated data model only achieved 64% (95% CI - 58-70%). In addition, the relative data model correctly predicted 76% (95% CI - 68-84%) of the 2017-18 data, compared with 70% (95% CI - 63-77%) for the isolated data model. From the relative data model, 10 PI's had significant relationships with game outcome; kicks from hand, clean breaks, average carry distance, penalties conceded when the opposition have the ball, turnovers conceded, total metres carried, defenders beaten, ratio of tackles missed to tackles made, total missed tackles, and turnovers won.

Conclusions: Outcomes of Premiership rugby matches are better predicted when relative data sets are utilised. Basic open-field abilities based around an effective kicking game, ball carrying abilities, and not conceding penalties when the opposition are in possession are the most relevant predictors of success.

Keywords: Partial dependence plots; Performance indicators; Random forest; Team sport.