Soccer is a low-scoring game where one goal can make the difference. Thus, counterattacks have been recognized by modern strategy as an effective way to create scoring opportunities from a position of stable defense. This coincidentally requires teams on offense to be mindful of taking risks, i.e. losing the ball. To assess these risks, it is crucial to understand the involved mechanisms that turn ball losses into counterattacks. However, while the soccer analytics community has made progress predicting outcomes of single actions (shots or passes) [1, 2] up to entire matches [15], individual sequences like counterattacks have not been predicted with comparable success. In this paper, we give reasons for this and create a framework that allows understanding complex sequences through comprehensible features. We apply this framework to predict counterattacks before they happen. Therefore, we find turnovers in soccer matches and create transparent counterattack labels from spatiotemporal data. Subsequently, we construct comprehensible features from sport-specific assumptions and assess their influence on counterattacks. Finally, we use these features to create a simple binary logistic regression model that predicts counterattacks. Our results show that players behind the ball are the most important predictive factors. We find that if a team loses the ball in the center and more than two players are not behind the ball, they concede a counterattack in almost 30% of cases. This stresses the importance to avoid ball losses in build-up play. In the future, we plan to extend this approach to generate more differentiated insights.
© Copyright 2022 Machine Learning and Data Mining for Sports Analytics. MLSA 2022. Communications in Computer and Information Science. Veröffentlicht von Springer. Alle Rechte vorbehalten.