Umeå universitets logga

Sports and specifically Ice-Hockey analytics has emerged as a very popular area of research in the last twenty years, with studies exploring player evaluation, game prediction, and team strength and weakness detection. However, one area that few studies have explored is the extent to which the different roles in a team contribute to team performance, and more specifically, whether different roles require different skills to succeed. This study aims to address these questions.

In order to do so, several dimension reduction techniques and machine learning methods are applied to a large dataset containing player-level variables from all National HockeyLeague (NHL) games between 2008 and 2025. This study is exploratory and identifies a few interesting patterns worth further study. These include the second defenseman consistently being important for team success, especially in its offensive metrics; the goalkeeper position consistently having very low importance; and a lack of evidence for different forward roles requiring different levels of defensive skill. Consistent with prior knowledge, the first centeris found to play the most important role in the team across most models, both offensively and defensively.

Across models, the most influential variables are on-ice expected goals for and against, along with variables highly correlated with these measures, and ratios between the two. The strongest models for the purposes of this study are a combination of an initial subjective variable selection followed by forward selection, and the machine-learning representative Gradient Boosting. In terms of predictive performance, the best model achieves an Area Under The Curve (AUC) of 0.61 for predicting game outcome and a Mean Squared Error (MSE) of 1.67 for predicting the expected goal difference of a game. Both of which are not especially strong in the general case, but are in line with earlier Ice-Hockey studies, highlighting the difficulty of predicting the outcome of Ice-Hockey games.