Evaluating Association Football Player Performances Using Markov Models

Historically, association football player performance evaluation has been based on subjective opinions and intuition, which is subject to cognitive bias. This thesis develops data driven player ratings, argued to tend towards objectivity. Two Markov models are developed, and used to derive player ratings. Markov modelling of football requires discretization of a dynamic sport, but contextual features are chosen to minimize the loss of information. In order to identify an accurate measure of player performance, different player ratings are proposed and validated. The validation consists of measuring the predictive power of the player ratings on match outcomes, the ratings' correlation with benchmark ratings and inter-season correlations. Based on the validation, one measure of player performance is chosen for further analysis. Lists of the top performing players in different positions in Eliteserien are produced from the ratings. Furthermore, player performance profiles are constructed, which include players' measured performance across different types of involvements. In order to identify similar players, the distances between players' performance profiles are calculated, believed useful for scouting replacement players. The impact of team quality on player performance is identified and measured. This enables a player's estimated performance in a new club to be calculated. The combination of player ratings, identification of similar players and estimated player performances can potentially be used as a framework for scouting players in professional association football clubs