Implementing state-transition models in team sports such as football (soccer), can be complex and time-consuming if observational methods need to be used. To help alleviate this, the aims of this study are two-fold. First, based on existing phases of play models, we introduce a hand-crafted expert state system that combines in- and out-of-possession phases of play to form a set of comprehensive tactical states. Second, we introduce a machine learning approach for automatically detecting the introduced tactical states based on positional data. This is implemented by capturing positional configurations of the players and the corresponding ball location. By clustering the player configurations for each ball zone and learning a “translation table” between the identified clusters and the expert state system, using three training games, the framework can be used to automatically detect states from the positional data of new games. The performance of our framework is evaluated by comparing the automatically detected states to the state label assigned by a human annotator using twelve test games from the 2021/2022 Bundesliga season. This showed good agreement, with F1-scores greater than 0.9 for most states. Hence, our framework may provide a method for time-efficient and comprehensive match analysis in football performance analysis.
Rothe et al. (Sun,) studied this question.