Never change a winning team, or should you? Assessing the impact of coach interventions in player lineups in sports using machine learning models.
In this thesis, we proposed a model splitting soccer matches per minute in order to examine if both predetermined match features, as well as dynamic in-game features, predict the final number of goals the home and away team will score. We used the data of the four biggest national soccer leagues (Bundesliga, La Liga, Serie A, and Premier League) from the seasons 18/19 till 21/22 to train and test the predictions of a multilayer perceptron and compared the performance with multiple linear regression and random forest. Furthermore, we examined the impact a coach can have on the final number of goals by making interventions in the player lineup during the game. The results demonstrate that overall, multilayer perceptron outperformed multiple linear regression (MSE -0.053) and random forest (MSE -0.014) with an increased accuracy of 1.42% and 0.41% respectively. Including the number of substitutions used and the change in strategy score significantly reduced the mean squared error by 0.22. Furthermore, increasing the number of substitutions used during the game significantly decreased the final number of goals for both the home and away team. Contrary to expectations, changing to a more attacking lineup by the home team decreases the final number of goals of both the home and away team, and changing to a more defensive lineup by the home team increases the final number of goals of both the home and away team.