Statistical Learning of Key Performance Indicators for Swedish Football

Abstract: Football is the indisputable most popular sport globally, and the central question within this game is how to become the winning outcoming part. A possible approach to answer this question is to utilise data and its information for analysis and provide keyperformance indicators that distinguish the successful from the unsuccessful teams. This master’s thesis aims to investigate the main differences between successful and unsuccessful teams by using statistical learning approaches. Two different approaches, Binary Regression and Random Forest, were adopted. Forbinary regression, three types of models (based on the link functions) were investigated: Logit, Probit, and Hazard. Comparisons between these models were conducted for obtaining the best performing model. For analysing leagues with convergence problems, implementation of K-means clustering and permutation with restrictions of features was applied. Using the Feature Importance for the Random Forest, a comparison between each feature and its importance for the model can be visualised. Furthermore, a uniformly distributed random variable in the Feature Importance was employed to obtain a benchmark for indicators more critical than randomness. This thesis resulted in an overall, for both approaches, significance and importance of shooting/finishing for all leagues and subsets. Moreover, the results do not substantially differ between the men’s leagues, Allsvenskan and Superettan, where both shots and passes show significance and importance. However, the women’s league, Damallsvenskan, is distinctly dominated by only shots. Hence, the indicator that distinguishes and separates successful from unsuccessful teams is shots. Finally, by connecting the results with common knowledge within football, the performed analysis provides powerful tools for future work within football analysis.