Innovative Metrics in Soccer: A Deep Dive into xT and VAEP
Written on
Introduction to Soccer Performance Metrics
In this article, we explore a significant analysis found in the research paper titled "Valuing On-the-Ball Actions in Soccer: A Critical Comparison of xT and VAEP." The authors, Maaike Van Roy, Pieter Robberechts, Tom Decroos, and Jesse Davis from the KU Leuven Department of Computer Science, delve into two of the most important metrics in soccer analytics.
The paper emphasizes the necessity of quantifying a player's performance during a match, which can impact team selection, scouting of opponents, and player recruitment. A primary challenge in measuring player contributions is the inherently low-scoring nature of soccer and the limited frequency of on-the-ball actions.
Analyzing Player Performance
The authors discuss various performance metrics proposed by soccer analytics experts, focusing on the evaluation of specific action types in diverse game situations, including shooting chances, off-ball positioning, passing, and set pieces. Recent research aims to create a comprehensive framework that can assess a wide array of actions across different contexts in the game.
Moreover, the paper notes that similar analytical models have emerged in other sports, such as basketball and rugby. It also highlights two main data sources for valuing soccer actions: event stream data and optical tracking data. While there has been some research utilizing tracking data, the majority of studies rely on event stream data, which is more readily accessible.
The paper outlines three primary approaches for evaluating player performance: Count-based, Expected Possession Value (EPV), and Action-based methods. Count-based methods assign weights to each action type and compute a weighted sum of occurrences. In contrast, EPV approaches segment matches into possessions—sequences of consecutive on-the-ball actions by the same team—and evaluate each action based on its impact on goal-scoring probabilities. Action-based methods, like VAEP, consider a broader range of actions and their context to assess each action's effect on the likelihood of scoring or conceding.
Comparative Analysis of xT and VAEP
The paper examines two prominent models within the last two categories—Expected Threat (xT) and VAEP—and explains how their design choices lead to distinct advantages and limitations. It illustrates specific actions where the two frameworks yield varying valuations. xT is shown to correlate more with playmaking, while VAEP tends to prioritize shooting.
Both methodologies offer fresh perspectives on player performance, moving beyond traditional metrics such as goals or assists per 90 minutes. This detailed comparison will be beneficial for researchers and practitioners focused on refining player evaluation techniques and understanding the nuances of existing methods.
Action Valuation Frameworks
The discussion includes frameworks for valuing actions based on event stream data, treating soccer matches as sequences of actions. The valuation of each action ( a_i ) is determined through a specific equation that quantifies the change in game state before and after the action.
Expected Threat (xT) Explained
The xT model operates as a possession-based Markov model, segmenting games into possessions controlled by a single team. Each possession is comprised of a series of ball-progressing actions aimed at enhancing scoring probabilities. The model represents each game state ( S_i ) on a grid overlaying the pitch, with values ( xT(z) ) assigned to zones based on scoring likelihood.
The iterative nature of the xT model allows for dynamic updates of zone values as the game progresses, which aids in analyzing team performance and strategy.
Learn more about the xT metric here:
Understanding VAEP
The VAEP algorithm offers a more intricate representation of game states compared to xT. It evaluates the last three actions during a match, categorizing features into three types: action characteristics (location, type), contextual elements (game tempo), and current game conditions (time remaining, score difference).
Unlike xT, which focuses purely on location, VAEP values actions based on how they impact game state probabilities, estimating the likelihood of scoring or conceding in future actions. The VAEP model captures both offensive and defensive value, reflecting the dual motivations of players during gameplay.
Learn more about the VAEP metric here:
Comparative Insights
This section highlights the similarities and differences between xT and VAEP in evaluating individual on-the-ball actions. While both frameworks measure action value through changes in game state, xT’s location-based approach contrasts with VAEP’s feature-based representation, which allows for a more nuanced understanding of game dynamics.
xT is limited to ball-progressing actions and overlooks defensive maneuvers, while VAEP encompasses a broader spectrum of actions and contextual factors, providing a more comprehensive evaluation.
Risk-Reward Evaluation
The paper further examines the possession-based approach of xT against the window-based approach of VAEP. xT focuses on actions that progress the ball, neglecting defensive plays, while VAEP accounts for prior actions' type and context, including game conditions.
The authors discuss specific actions, like backward passes into the penalty box and first ball progression during counter-attacks, showing that VAEP can effectively assess risks associated with various actions, in contrast to xT's limitations.
Data Utilization
Utilizing data from StatsBomb for the English Premier League seasons of 2017/2018 and 2018/2019, the study employs the SPADL format to standardize event stream data representation. The XGBoost algorithm serves as the prediction method for the VAEP model, while the xT model uses a 16×12 grid, with convergence reached after several iterations.
Experimental Comparison of Action Values
In this segment, the authors compare the effectiveness of xT and VAEP in valuing specific soccer actions across various contexts. They analyze actions like risky backward passes, ball recoveries for counter-attacks, forward dribbles in the penalty box, and forward passes near the penalty area, noting how each model assigns different values based on their methodologies.
Experimental Player Rating Comparison
The authors present a comparison of player ratings derived from both xT and VAEP, normalizing ratings per 90 minutes to account for time on the pitch. The analysis highlights varying rankings for players based on each model, demonstrating the distinct qualities valued by xT and VAEP.
Comparison with Traditional Metrics
The paper contrasts the rankings generated by xT and VAEP with traditional metrics like goals and assists, revealing that VAEP correlates more strongly with goal metrics, while xT aligns better with assists. Both models, however, fail to account for the context in which actions occur, limiting their comprehensive evaluation of player performance.
Robustness Evaluation
A robustness analysis reveals that xT ratings demonstrate greater consistency across different data subsets compared to VAEP. The authors attribute this stability to the nature of the xT model, which relies on consistent player actions, while VAEP's context-sensitive ratings may introduce more variability.
Conclusion
This research offers valuable insights into the contrasting methodologies of xT and VAEP for valuing soccer actions. While both metrics present unique advantages, they also diverge from traditional performance indicators, highlighting the complexities of player evaluation. Understanding these strengths and limitations can inform future applications in player assessment and strategic development in soccer.
References
Van Roy, M., Robberechts, P., Decroos, T., & Davis, J. (2020). Valuing on-the-ball actions in soccer: a critical comparison of xT and VAEP. In Proceedings of the AAAI-20 Workshop on Artificial Intelligence in Team Sports.