Susanne & Bjarke's masters thesis, written between April 2018 and December 2018. The defence were held in Aud 3, HCØ 20th December at 13:00. The event can be found here. Datasets used throughout the thesis can be found here.
In this thesis we present three approaches for optimizing win probability in team sports. These approaches revolve around creating dynamic applicable algorithms for computing optimal team compositions, predicting win probability, and assessing player performance. Also, we present and discuss the underlying theory in order to make the algorithms easy to interpret and implement. We see that we can apply hierarchical clustering techniques and then utilize the gap statistic in order to cluster player profiles into roles, which can be used for several team sports. When doing live predictions on win probability, we construct several algorithms for predicting, which can likely be used for other team sports as well. We show that logistic regression and a modified k-nearest neighbors algorithm can be used for computing win probabilities. Finally, we show that player assessments can be made by applying dynamic regression models, taking roles into account that were derived from the hierarchical clustering process.