Actions Speak Louder than Goals: Valuing Player Actions in Soccer

Tom Decroos, Lotte Bransen, Jan Van Haaren, Jesse Davis

TL;DR

This paper introduces SPADL, a standardized language for describing soccer actions, and VAEP, a novel framework that values individual player actions based on their impact on game outcome, accounting for context. This allows for a more nuanced player evaluation beyond traditional metrics like goals and assists.

What Problem Does This Paper Address?

Assessing the true impact of individual player actions in soccer is a significant challenge. Traditional metrics often focus on rare events like goals or shots, or assign fixed values to actions without considering the game's context. This overlooks the subtle, yet crucial, contributions players make through passes, dribbles, and defensive actions that influence the game's trajectory. The dynamic and low-scoring nature of soccer further complicates this, as many actions don't immediately lead to a goal but can create future opportunities or prevent conceding.

Methods

The authors propose two main contributions:

  1. SPADL (Soccer Player Action Description Language): A standardized, human-interpretable language to represent soccer actions. It aims to unify diverse event stream data formats by defining a consistent set of nine attributes for each action (StartTime, EndTime, StartLoc, EndLoc, Player, Team, ActionType, BodyPart, Result). This addresses the challenge of vendor-specific data formats and simplifies analysis.
  2. VAEP (Valuing Actions by Estimating Probabilities) Framework: This framework assigns a value to each player action based on its expected impact on the game's scoreline. It works by:
  3. Estimating Scoring and Conceding Probabilities: Using a machine learning model (CatBoost) trained on historical game data (event streams) to predict the probability of scoring and conceding within a short future window (k=10 actions). Features include SPADL action details, complex interactions between actions, and overall game context.
  4. Calculating Action Values: An action's value is determined by how much it changes the probability of scoring (offensive value) and conceding (defensive value). Specifically, the offensive value is the increase in the team's scoring probability, and the defensive value is the decrease in the opponent's scoring probability (or increase in the team's conceding probability, negated).
  5. Aggregating to Player Ratings: Individual action values are summed up and normalized per 90 minutes to create player ratings, reflecting their total offensive and defensive contributions.

Datasets

The study utilizes Wyscout data for English, Spanish, German, Italian, French, Dutch, and Belgian top divisions, covering 11,565 games from the 2012/2013 through 2017/2018 seasons. This data is transformed into the SPADL representation. No new benchmark datasets are introduced; rather, the paper focuses on standardizing existing event stream data.

Key Insights and Ideas

  • Context is King: The framework's ability to value actions based on their context (e.g., a pass backwards to a less favorable position being penalized) is a significant departure from traditional methods.
  • Beyond Goals and Assists: VAEP can identify and reward players whose contributions are not captured by traditional metrics, such as creative midfielders or defensive anchors who excel at ball retention and tactical positioning.
  • Playing Style Characterization: By aggregating action values per action type, the framework can provide insights into a player's playing style, helping clubs identify suitable replacements or players who fit a specific tactical system.
  • Player Value vs. Market Value: The paper suggests that VAEP ratings can sometimes offer a more accurate reflection of a player's true contribution than their market value, especially for defensive players or those in less prominent teams.

Key Results

  • The proposed SPADL language and VAEP framework are demonstrated to be effective in valuing player actions and generating player ratings.
  • Comparisons with traditional metrics (goals per 90, assists per 90, goals+assists per 90) show that VAEP identifies players who are often overlooked by these metrics but are highly valued by experts (e.g., Kevin De Bruyne, Eden Hazard).
  • The framework's ability to characterize playing styles is illustrated through comparisons of players like Neymar, Dembélé, and Coutinho, and potential replacements for Cristiano Ronaldo.
  • The CatBoost algorithm, combined with the comprehensive feature set, achieved superior performance in estimating scoring and conceding probabilities compared to baseline feature sets and other learning algorithms.

Limitations and Future Work

  • On-the-ball actions only: The current VAEP framework only values actions where a player has possession of the ball. Defensive actions that don't involve direct ball contact (e.g., positioning, anticipation) are not explicitly valued.
  • Inter-league and Inter-club comparisons: It's challenging to directly compare player ratings across different leagues or clubs due to variations in team strength and playing styles. Players in stronger teams or leagues might have their contributions masked or amplified.
  • Building trust: The novel nature of VAEP ratings may require effort to build trust and understanding among traditional scouts and stakeholders who are accustomed to more conventional metrics.
  • Future work: The authors suggest extending the framework to value off-the-ball actions, refining comparisons across leagues and clubs, and developing more intuitive ways to present the results to stakeholders.

Summary updated

Abstract

Assessing the impact of the individual actions performed by soccer players during games is a crucial aspect of the player recruitment process. Unfortunately, most traditional metrics fall short in addressing this task as they either focus on rare actions like shots and goals alone or fail to account for the context in which the actions occurred. This paper introduces (1) a new language for describing individual player actions on the pitch and (2) a framework for valuing any type of player action based on its impact on the game outcome while accounting for the context in which the action happened. By aggregating soccer players' action values, their total offensive and defensive contributions to their team can be quantified. We show how our approach considers relevant contextual information that traditional player evaluation metrics ignore and present a number of use cases related to scouting and playing style characterization in the 2016/2017 and 2017/2018 seasons in Europe's top competitions.
BibTeX
@inproceedings{decroos2019actions,
  author = {Tom Decroos and Lotte Bransen and Jan Van Haaren and Jesse Davis},
  title = {{Actions Speak Louder than Goals: Valuing Player Actions in Soccer}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2019},
  doi = {10.1145/3292500.3330758},
}
Figure