Long-distance relationships: Positional classification of Australian football league players

This research demonstrates how Australian Football League (AFL) players can be accurately and efficiently classified into four recognised playing positions (Defence; Midfield; Forward; Ruck) after each match, using only a handful of collected game-related performance variables. By maximising the Mahalanobis distance between a linear combination of thirteen performance variables and their respective centroids, 7,744 individual player cases in the 2009 AFL season are assigned to one of the four positions, without any prior knowledge of that player's movement within the match. Once the discriminant functions have been developed, Bayesian probabilities are then calculated to highlight each player's level of activity across the four positions in each match. This information is crucial when developing a set of position-dependent rules with which to measure AFL player performance. The research then progresses to intra-position analysis where each player is further classified based on the Squared Euclidian distance between position-specific elements derived from that player's performance covariance matrix (PCM). A case study details how forwards can be segmented into discrete and continuous playing roles based on the distances between covariance couplets. This information is of high importance for coaching staff and pundits alike as post-match deductions can be made, not only about a player's influence on the match, but also a player's influence within each position. An appealing aspect of the research is that only a few simple game-related statistics are required to gauge a player's positional performance, without having to resort to audio-visual tools and complex mapping