Markov Chain Models for the Near Real-Time Forecasting of Australian Football League Match Outcomes

The Australian Football League (AFL) is a fast-paced invasion style ball game where two teams of 18 players jockey for a position through the strategic use of on-field tactics and skills. Each match consists of four quarters which range in length from 20 to 40 min and are heavily influenced by the irregular nature of game stoppages. The purpose of this study was to identify whether it was possible to model and subsequently forecast match outcome based on both prior match information as well as real-time transactional data. 23 professional level matches were analysed with each match containing an average of 1907 transactional epochs. To account for the irregular frequency at which each epoch occurs, a time-inhomogeneous Markov process was adopted, with model evaluation assessed on both epoch accuracy and final match result. The model performed notably well with an average epoch accuracies in excess of 71% and match outcome results in excess of 78%. The results of this study demonstrate that accurate near real-time prediction is achievable under real-world conditions using real match AFL transactional data.