| dc.description.abstract | In this thesis I have developed a model for result prediction in soccer. The model is based
on chances created being modeled as a Poisson process while goals scored is seen as a result
of first creating chances and then converting them, here modeled as a Bernoulli trial. Compared to existing models, this one takes advantage of a number of data that previously have not been considered. As each team is described by four parameters, teams are able to be distinguished further allowing for better prediction of chances created and goals scored for each team in a given matchup.
Six different models are developed gradually with the goal of improving the model fit to data
and its predictive ability. In the final model the parameters can change over time so as to explain how a team can go through periods of good or bad form. Parameters are assumed to be correlated to each other - reflecting how a good offensive team also often has a good defense. Red cards are included to explain why certain surprising results took place. And lastly, the model uses shots on target to predict goals, as this is shown to have a stronger connection than between shots and goals.
A betting strategy is implemented where the size of the bet decreases with increasing odds,
while only placing bets when the expected return is above a certain value. The model struggles with consistency, but is still able to make a small profit over a run of five seasons, so it should be an excellent candidate for further development. | |
