Foundations of Expected Points in Rugby Union: A Methodological Approach

This study explores the feasibility of an Expected Points metric for rugby union, aiming to shift performance analysis from descriptive indicators to a predictive metric of possession quality. Notational analysis was conducted on 132 Premiership Rugby matches, producing a dataset of 35,199 unique phases of play containing variables such as team in possession, pitch location, type of play, score differences, time remaining, cards and the next scoring outcome. Four

machine learning algorithms were explored to predict scoring outcomes: multinomial logistic regression, random forest, support vector machine and k-nearest neighbors. After extensive feature engineering and hyperparameter optimisation, the best-performing model (a random forest classifier) achieved 39.7% ±2.8 ppts accuracy. However, this did not meet a literature-derived baseline for practical usability (44.3%), thus the model was not suitable for applied contexts. A key

challenge was predicting minority scoring outcomes due to severe class imbalance. SMOTE was explored to address this imbalance, resulting in a lower accuracy (35.7%) but an improved F1-score of 34.4%. This study highlights the inherent limitations of modelling scoring outcomes in dynamic, open-play team sports, challenging the predominant positivist paradigm in sports performance analysis. The methodology provides critical foundational groundwork and a benchmark for future research to build upon. It recommends exploring advanced samplers for minority classes, expanded feature sets and alternative modelling techniques, such as recurrent neural networks.