Breaking the Ice: ExpectedGoals in HockeyAllsvenskan

Sammanfattning

Predicting the probability that any given shot will result in a goal is known as expected goals or xG. It is an increasingly common method for analyzing performance in sports. This thesis project presents the development of an xG model tailored specifically for Swedish professional hockey, more precisely the second highest league in Sweden, HockeyAllsvenskan. The model were developed using event data from league games, with features such as shot distance, shot type and characteristics of the pass before the shot. Two different machine learning models was explored, a feed forward neural network and an XGBoost classifier. The model performance was evaluated using ROC AUC where it achieved a score of 0.78. However, more importantly, the predicted probabilities were compared with actual goal frequencies across different categories such as players, teams, features and shot types. The model aligned well with the actual goal rates confirming the models interpretability. This xG model can assist coaches, analysts and players in better understanding shot quality and what makes a scoring chance a good scoring chance. While the results are promising, the model could be further improved with the addition of tracking data.