Fitting probability distributions to real-time AFL data for match prediction

In this research a Generalized Logistic Model (GLM) is used to model outcomes of Australian Rules football matches in real-time. Incorporating difference in team quality and score difference the outcome of the model is the probability of victory at each of the quarter time breaks. The parameters of the GLM are a function of opponent quality and are optimized via simulation for each quarter. Archival AFL data was obtained from seasons 2000 to 2009 which consisted of year, round, quarter, (nominal) home team, away team, home team score and away team score. Seasons 2000 to 2004 are used as a training set for the forward prediction of seasons 2005 to 2009. Comparisons are made throughout against a simple Brownian motion model. Both models are then evaluated on predicted and actual probabilities of winning.