Beating the Odds - A State Space Model for predicting match results in the Australian Football League

Abstract

This thesis investigates whether state space models have the potential to pre- dict the outcome of Australian Rules Football matches and can produce significant positive return over the bookmaker’s odds. The point of departure is a sample of 18 Australian football teams over the period 2012 to 2016. Modeling and predicting a football match is a challenging task, since the model should incorporate two di ↵ er- ent random processes. Firstly, the evolution of parameters, i.e. the team strengths change over time as it incorporates changes in team composition, coaching, train- ing ground, injuries etc. Secondly, the distribution of ranking data with these time-varying propensities changes stochastically over time. Given that we cannot observe all team-specific and location-specific factors, a dynamic state space model for team strengths is introduced. The team strengths are assumed to follow an order-one autoregressive process and are estimated using a recursive Kalman filter algorithm. Smoothed state estimates are applicable for ranking teams and predict- ing future outcomes of the matches. We show that beating the bookmaker’s odds is a challenging task indicating that the betting markets are e cient. Keywords : Sport Analytics; Kalman Filter Algorithm; State Space models; Predictive inference; Ranking; Australian Rules Football.