Estimating Probability Distributions of FIFA-Defined Phases of Play Based on Inter-Analyst Diversity

Automatic recognition of play phases in soccer is essential for tactical analysis. Traditional methods have two primary challenges. First, they often fail to achieve general-purpose understanding from their recognition results, largely due to the lack of datasets based on standardized definitions. Second, play phases have typically been treated as a single, deterministic label. This approach overlooks the ambiguity experts often perceive, where a situation may have multiple possible interpretations or coexisting tactical intentions. We propose a method to estimate the probability distribution of the nine FIFA-defined “phases of play” (Build up, Progression, Final third, Counter-attack, High press, Mid block, Low block, Counter-press, and Recovery) to tackle these challenges. Our approach treats these phases probabilistically, leveraging the diverse opinions of multiple analysts. Our methodology consists of two stages. First, we construct a ground-truth probability distribution by aggregating annotations from multiple analysts, where each analyst assigns a single play phase at every moment following the FIFA definitions. Next, we build a deep learning model to estimate these probability distributions from player trajectories. As a result, we developed a dataset comprising 180 minutes of scenes extracted from 10 matches, achieving an inter-analyst reliability with a weighted Krippendorff’s α of 0.68. Furthermore, the estimation model achieved a Mean Absolute Error of 0.06 and a Root Mean Squared Error of 0.16, but transition phases remain difficult to estimate. This approach provides a new foundation for tactical analysis that quantitatively addresses the ambiguity inherent in play phases.