Bennett Mark, Bezodis Neil, Shearer David A, Locke Duncan, Kilduff Liam P
The Rugby Football Union, UK; Applied Sport Technology Exercise and Medicine Research Centre (A-STEM), College of Engineering, Swansea University, UK.
Applied Sport Technology Exercise and Medicine Research Centre (A-STEM), College of Engineering, Swansea University, UK.
J Sci Med Sport. 2019 Mar;22(3):330-334. doi: 10.1016/j.jsams.2018.08.008. Epub 2018 Aug 18.
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.
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.
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.
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.
本研究的主要目的是检验英式橄榄球联盟比赛预测结果的准确性是否因数据分析方法（即孤立数据、描述性转换数据或相对数据）的不同而有所差异。次要目的是随后使用最合适的方法来研究与比赛结果最相关的性能指标（PI）。
数据为2016 - 17赛季英格兰超级橄榄球联赛127场比赛的16个PI。鉴于比赛结果为二元（胜/负），使用这些数据集构建了随机森林分类模型。通过预测2017 - 18赛季72场比赛的数据集结果，进一步评估模型的预测能力。
相对数据模型对2016 - 17赛季数据的平衡预测率达到80%（95%置信区间 - 75 - 85%），而孤立数据模型仅达到64%（95%置信区间 - 58 - 70%）。此外，相对数据模型正确预测了2017 - 18赛季数据的76%（95%置信区间 - 68 - 84%），相比之下，孤立数据模型为70%（95%置信区间 - 63 - 77%）。从相对数据模型来看，10个PI与比赛结果有显著关系；手抛球、成功突破、平均带球距离、对方控球时被判罚的点球、失误的球权转换、总带球距离、突破的防守球员、漏铲与成功铲断的比例、总漏铲次数以及赢得的球权转换次数。
使用相对数据集时，英超橄榄球比赛的结果能得到更好的预测。基于有效踢球比赛、带球能力以及在对方控球时不被判罚点球的基本开阔场地能力是成功的最相关预测因素。
