Statistical modelling and Bayesian inference for match outcomes and team behaviour in association football
Abstract
This thesis presents advances in modelling and inference for match outcomes in
the association football English Premier League. We firstly extend earlier models
by introducing a behavioural aspect which can be used to investigate how teams
react to the state of play in a match. We show that the model, in its simplest form,
outperforms existing models and is able to select a portfolio of pro table bets against
a bookmaker. Secondly, we introduce a dynamic component to the model by allowing
team ability parameters to vary stochastically in time. We employ particle filtering
methods to cope with a mixture of static and dynamic parameters and find that the
updating of posterior distributions is particularly fast, a necessary attribute should
we wish to update parameter estimates while matches are in-play. Furthermore, it
is shown that the methods are able to recover model parameters based on simulated
league data. Finally, we propose an extension to the model so that we are able to
investigate how a team modifies its behaviour based on their league situation. We
consider league positions that are closely attainable and suggest that since teams
modify their behaviour based on their current league position, outcomes of different
matches are not necessarily independent.