Verder kijken dan de score: Machine Learning-technieken om (vrouwen)voetbal te begrijpen

The rapid growth in the amount of data collected from soccer matches has transformed how the game is analyzed. Soccer clubs increasingly rely on data in their decision-making processes. Player recruitment and match preparation are two critical tasks in professional soccer that could benefit from the use of data. Player recruitment encompasses finding the right player to acquire for your team. However, the global player pool is too large to monitor manually and player performance in one league may not transfer directly to the other due to differences in playing style and level. Additionally, successful transfers depend not only on individual performance but also on tactical and team fit. Match preparation is equally demanding: coaching staff must analyze opponents and prepare set pieces with only a few days between matches. Although video analysis supports this process, time constraints limit how thoroughly staff can assess footage. The increasing availability of complex data from matches provides opportunities to address these challenges using Artificial Intelligence (AI) and Machine Learning (ML) techniques. Quantifying player performance is one of those challenges that has been addressed using AI techniques. Traditionally, player performance was measured using simple counting statistics such as the number of shots on target or the percentage of successful passes. The introduction of more sophisticated data enabled building metrics that better measure player performance. One of the first metrics was Expected Goals (xG), which measures the likelihood of a shot yielding a goal. Since shots only constitute a small proportion of the actions in a match, other metrics have been introduced that value all on-the-ball actions such as passes, dribbles and interceptions. However, these metrics often ignore the broader context in which they were executed. That is, they ignore context such as the player's teammates and the mental aspect. This dissertation addresses this gap by proposing novel AI-based frameworks to 1) quantify player chemistry and 2) player performance under mental pressure. Furthermore, the field of soccer analytics has largely focused on men's soccer, with limited attention to the women's game. Most existing performance models were trained on men's data. Given the machine learning principle that models should be trained on representative data, it raises the question whether such metrics could simply be transferred to the women's game. This dissertation addresses this question. Furthermore, it provides the first large-scale data-driven analysis of the technical and tactical developments in the women's game. It also introduces a novel metric to measure a player's pace of play. Finally, data-driven analysis of penalty kicks is limited and often models penalty kicks as a two-player zero-sum game, assuming that goalkeepers and kickers choose their actions independently. However, this oversimplifies reality as kickers and goalkeepers may adapt their decisions based on the opponent's movements. Moreover, not all players can adapt their actions in this way, and small sample sizes make it hard to assess individual player capacities. This dissertation introduces a simulation framework using AI techniques to optimize goalkeeper policies taking into account these action capacities. In summary, this dissertation has three main contributions to the field of soccer analytics: 1) understanding women's soccer using data, 2) capturing more context in data-driven recruitment, and 3) optimizing goalkeeper policies for penalty kicks.