Source-linked AI summary
Actions Speak Louder Than Goals: Valuing Player Actions in Soccer
Tom Decroos, Lotte Bransen, Jan Van Haaren, Jesse Davis
TL;DR
Soccer player evaluation needs a way to value individual actions beyond goals and shots while accounting for context and downstream effects. The paper introduces SPADL and VAEP to represent actions and estimate their impact on scoring and conceding. The framework supports player ratings and use cases in scouting and playing-style characterization, while valuing only on-the-ball actions.
Problem
Existing soccer action-valuing approaches focus on goals and shots, use fixed values, and often ignore context and longer-term effects.
Method
The paper introduces SPADL for representing event-stream actions and VAEP for valuing action types according to context and effects on scoring and conceding probabilities.
Results
The framework produces action values and player ratings, with use cases showing its application to scouting and playing-style characterization.
Takeaways & Limitations
Aggregated action values quantify players’ offensive and defensive contributions while accounting for contextual information that traditional metrics ignore.
Takeaways & Limitations
VAEP values only on-the-ball actions, whereas defensive contribution can involve positioning and anticipation that prevent opponents from gaining possession.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
The paper addresses the underexplored problem of objectively valuing individual soccer actions, proposing a context-aware framework and action language to quantify player contributions.
- Individual action impact matters for player acquisition, evaluation, scouting, media, and fan engagement.
- Traditional approaches focus mainly on goals and shots, assign fixed action values, and overlook longer-term effects.These limitations make context important: a pass can create space for a later goal chance without immediately changing the scoreline.
- The paper proposes a data-driven framework that values all action types according to game circumstances and possible longer-term effects.An action value represents its expected influence on the scoreline, such as +0.05 expected goals for the acting team.
- The contributions include SPADL, a framework for valuing actions and rating players, and models for predicting short-term scoring and conceding probabilities.The paper also presents use cases involving scouting and playing-style characterization.
2 SPADL: A LANGUAGE FOR DESCRIBING PLAYER ACTIONS
SPADL unifies heterogeneous soccer event streams into a consistent, interpretable action vocabulary designed for automated analysis and contextual player-action evaluation.
- Soccer event-stream data varies across vendors, may omit important information, and often contains formats that hinder automated analysis.Optional information, inconsistent terminology, legacy compatibility, and variable-sized inputs increase preprocessing effort.
- SPADL unifies event-stream formats into a human-interpretable, simple, and complete vocabulary for describing actions in context.Its design supports reasoning about on-pitch events while reducing processing mistakes and representing information needed for analysis.
- The accompanying Python package automatically converts event streams to SPADL and supports Opta, Wyscout, and StatsBomb.
- 2.2 Language description: SPADL represents a game as a sequence of on-the-ball actions, with each action defined by nine attributes.These attributes include timing, start and end locations, player, team, action type, body part, and result.
- 2.2 Language description: The language uses fixed attributes for every action and defines 21 interpretable action types, with standardized body parts and results.Excluding optional snippets makes automatic analysis easier; action types include passes, dribbles, tackles, shots, clearances, and keeper saves.
3 VAEP: A FRAMEWORK FOR VALUING PLAYER ACTIONS
VAEP values each action by combining its effects on near-future scoring and conceding probabilities, then aggregates those values into player ratings across time frames and action types.
- Action values: VAEP values an action by combining its offensive effect on scoring probability with its defensive effect on conceding probability.The defensive component negates the change in conceding probability because reducing concessions is beneficial.
- Action values: An action’s value is computed for the team of the player performing it, using the transition from the previous game state to the current state.The team identity is denoted by x_i for action a_i.
- Action values: The framework converts action valuation into predicting the likelihood of future scoring or conceding events.It is independent of the representation used to describe the actions.
- Player ratings: Player ratings aggregate individual action values over a chosen time frame and normalize them per 90 minutes of game time.The resulting rating captures average net goal difference contributed to the team per 90 minutes.
- Player ratings: Ratings can also be computed by action type, enabling player profiles and characterization of different playing styles.The framework supports ratings along different dimensions depending on the use case.
4 ESTIMATING SCORING AND CONCEDING PROBABILITIES
The method estimates near-future scoring and conceding probabilities from recent action sequences and contextual features, using two calibrated probabilistic classification tasks for the team in possession.
- Prediction targets: The estimation task predicts whether the team in possession will score or concede within a user-defined sequence of future actions.The future window comprises k subsequent actions after the current game state.
- Prediction targets: The authors reduce the problem to two binary probabilistic classifiers with identical inputs but different scoring and conceding labels.Only the probabilities for the team possessing the ball are estimated; the opposing probabilities follow from team symmetry.
- Modeling: The paper uses CatBoost for both probabilistic classification problems, emphasizing that probability estimates should be well-calibrated.The authors state that other probability-predicting machine-learning models could also be used in principle.
- Prediction targets: Each game state becomes one training example, with labels determined by goals scored or conceded during the subsequent k actions.A positive scoring label indicates a goal by the possessing team, while a positive conceding label indicates that team conceded.
- Modeling: The future horizon is set to k = 10 based on domain knowledge and preliminary experiments.Here, k determines how far ahead the method looks when assigning action effects.
- Features: The model represents context using the previous three actions, including action attributes, spatial and temporal features, transitions between actions, and scoreline information.Three actions were empirically found to work well, and scoreline features reflect teams’ adaptations to the current game state.
5 EXPERIMENTS
The experiments apply VAEP to league data to examine action values, player rankings, and scouting use cases. Results show that contextual action valuation identifies impactful players beyond traditional goal- and assist-based metrics and ranks young talent across leagues.
- Experimental setup: VAEP is evaluated on 11,565 league games from seven European top divisions across the 2012/2013–2017/2018 seasons.The analysis excludes friendly, cup, and European games.
- 5.1 Intuition behind the action values: The framework visualizes six action values leading to Barcelona’s 93rd-minute goal against Real Madrid, including Messi’s +0.05 dribble and +0.09 backward pass.The backward pass raised scoring odds from 0.08 to 0.17, while the dribble raised them from 0.03 to 0.08.
- 5.2 Comparing our VAEP player ratings to traditional player performance metrics: VAEP’s top-10 Premier League list includes De Bruyne, Hazard, and Mahrez, who appear in none of the traditional goals, assists, or combined top-10 lists.This comparison tests VAEP against goals, assists, and goals + assists per 90 minutes.
- 5.2 Comparing our VAEP player ratings to traditional player performance metrics: 1,110 million euro is the combined market value of VAEP’s top-10 players, compared with 862 million euro for goals, 760 million euro for assists, and 947 million euro for goals + assists.The paper reports these observations as suggesting that VAEP better captures players’ contributions than traditional metrics.
- 5.3 Identifying promising young players and minor league talent: VAEP rankings identify young players in both the tougher English and Spanish leagues and the smaller French, Dutch, and Belgian leagues.The young-talent analysis covers players born after January 1, 1997 who played at least 900 minutes.
- 5.3 Identifying promising young players and minor league talent: The framework ranks talent across leagues when the required event-stream data is available, supporting its use as a scouting tool.The paper illustrates this use case with players such as Jonjoe Kenny, whose VAEP rating exceeds his estimated market valuation.
5.4 Characterizing playing style
The framework supports scouting by decomposing players’ contributions into action types and by examining the trade-off between action quality and quantity. These comparisons reveal stylistic similarities and differences that can inform recruitment choices.
- Scouting and recruitment: Player-style metrics help scouts assess a larger pool of candidates than judging playing styles with the naked eye.Scouts’ time is described as a limiting resource during recruitment.
- Scouting and recruitment: Dembélé and Coutinho were reasonable Neymar replacements: both passed better, while Neymar remained the superior dribbler.Dembélé also outperformed Neymar in crossing, and Coutinho outperformed him in shooting.
- Scouting and recruitment: Rashford was stylistically closer to Ronaldo, whereas Hazard was the better player and neither replicated Ronaldo’s finishing.The analysis presents Rashford as preferable for preserving Real Madrid’s style and Hazard for immediately strengthening the team.
- Trading off action quality and quantity: Players face a quality–quantity trade-off: frequent actions are harder to value highly individually, producing distinct player archetypes.The English and Spanish league examples distinguish players with fewer highly valued actions, more lower-valued actions, and intermediate profiles.
- Trading off action quality and quantity: The scatter plots compare actions per 90 minutes with average action value, and identify Messi as an outlier relative to other players.Figure 3 covers players with at least 900 minutes in the 2017/2018 Spanish or English leagues.
6 RELATED WORK
Related work has begun valuing soccer actions, but commonly limits action types or ignores their circumstances. This paper differs by modeling exact action locations and all relevant on-ball actions, while connecting action values to expected-goals state values.
- Soccer action valuation remains relatively unexplored because games are dynamic and low-scoring.
- Unlike fixed-zone approaches, this framework models the exact locations of each action.
- Unlike approaches valuing only three on-the-ball action types, this framework considers any relevant on-the-ball action during a game.
- Existing soccer approaches often value only limited action types, such as passes and shots, or omit the circumstances surrounding actions.
- Within VAEP, the expected-goals value of a goal attempt is computed by estimating the value of the game state before the attempt.
7 CONCLUSION
The paper introduces SPADL for representing event-stream data and VAEP for valuing individual player actions. VAEP values all action types using game context and possible effects on subsequent actions.
- SPADL is a language for representing event-stream data designed to facilitate data analysis.
- VAEP is a framework for assigning value to each individual player action during a soccer game.
- VAEP values all action types, including passes, crosses, dribbles, and shots.
- VAEP bases action valuation on game context and considers possible effects on subsequent actions.
A.1 SPADL action types
Table 4 presents the 21 action types used in the SPADL representation and describes each type.
- SPADL contains 21 action types, each paired with a description in Table 4.
A.3 Experimental setup and implementation
The experiments evaluated four learning algorithms on SPADL-transformed soccer actions, using large training and evaluation sets. Table 4 defines the 21 action types and their success conditions and special result values.
- Implementation: Four learning algorithms were evaluated in the experiments.
- Implementation: Logistic Regression used L2 regularization with L-BFGS, while Random Forest used 100 trees.
- Implementation: CatBoost used default parameters except for 40 parallel threads.
- SPADL action types: Table 4 covers 21 SPADL action types, with columns specifying success conditions and additional special result values.
- Datasets and runtime: The training sets contained 8,518,378 and 11,438,956 actions, while the evaluation sets contained 2,920,578 and 2,988,847 actions.