INTRODUCTION: In recent years, advanced data analytic techniques have been utilised to determine performance indicators in professional rugby union. Typically, these studies report 20+ key PI`s that collectively can be used to predict match outcome. Whilst, acknowledgement of the PI`s is strategically important, their high number can impede the deployment of actionable interventions by practitioners. No key research has investigated the use of variable selection methods to build and validate simplified models capable of predicting successful performances. Equally, there has also been no evidence of research within the United Rugby Championship (URC). Therefore, the primary aim of this study was to investigate team level PI`s associated with winning performances within the URC, through machine learning and feature selection methods. Secondary aims included to test differences between individual team PI`s (isolated data) and team PI`s relative to their opposition (relative data) and to examine whether variable selection can be used to simplify modelling processes. METHODS: A range of 27 PI`s were taken from all 96 matches within the 2020-21 URC season (formerly known as the PRO14). These PI`s were selected to characterise different areas of the game, including attack, defence, set piece and infringements. Random forest classification modelling was completed on both isolated and relative datasets, using the binary match outcome (win/lose) as the response variable. Maximum relevance and minimum redundancy was used to locate an optimal subset of PI`s that are mutually and maximally dissimilar and can represent the response variable effectively. Models were used in prediction on 53 unseen matches from the 2021-22 season. RESULTS: Within the 2020-21 datasets, the full models correctly classified 83% (CI 77%-88%) of match performances for the relative dataset and 64% (CI 56%-70%) for the isolated set. When models were optimised by reducing the number of variables, these values were 85% (CI 79%-90%) and 66% (CI 58%-72%). In prediction, the reduced relative model successfully classified 90% of previously unseen match performances (CI 82%-95%). Within the relative data model, 5 PI`s were significant in differentiating between wins and losses: kicks from hand, metres made, clean breaks, turnovers conceded and scrum penalties. CONCLUSION: Indicators of success within the URC include increased kicks from hand, metres made, and clean breaks compared to the opposition, as well as less scrum penalties and turnovers conceded. Relative PI`s are more effective in predicting match outcomes than isolated data. Simplifying to a smaller number of key variables does not degrade model accuracy, suggesting that smaller groups of PI`s may be used as replacement for the wider dataset. In practical applications this could allow practitioners to focus on a select number of PI`s, to allow for a more manageable approach to training and monitoring tactics.
© Copyright 2022 27th Annual Congress of the European College of Sport Science (ECSS), Sevilla, 30. Aug - 2. Sep 2022. Veröffentlicht von Faculty of Sport Science - Universidad Pablo de Olavide. Alle Rechte vorbehalten. / Übersetzt mit https://www.DeepL.com/Translator