Actions Speak Louder Than Goals: Valuing Player Actions in Soccer
Proposes VAEP framework using probabilistic models to quantify player actions' impact in soccer.
Key Findings
Methodology
This paper introduces VAEP (Valuing Actions by Estimating Probabilities), a probabilistic framework that leverages the SPADL language for standardized event descriptions. It employs CatBoost classifiers to predict short-term scoring and conceding probabilities based on features extracted from recent actions, spatial data, and game context. Action values are computed as the sum of offensive and defensive impacts, reflecting each action’s expected contribution to the game outcome. The approach integrates multiple data sources, enabling dynamic, context-aware valuation of diverse behaviors such as passes, shots, and dribbles, across multiple seasons and leagues.
Key Results
- VAEP scores correlate strongly with market values, outperforming traditional metrics like goals and assists. In the 2017/2018 EPL season, top-ranked players like Kevin De Bruyne and Eden Hazard scored highly on VAEP, with market valuations exceeding 1.1 billion euros collectively, compared to 862 million euros for goals alone.
- Predictive accuracy for short-term scoring and conceding probabilities exceeded 85%, demonstrating calibration and robustness. The model effectively distinguished impactful players, capturing subtle behavioral differences and tactical influences.
- Ablation studies confirmed that incorporating game state and spatial features significantly improved performance, emphasizing the importance of context in valuation. The framework also identified undervalued talents from smaller leagues, aiding scouting efforts.
Significance
This work advances sports analytics by providing a comprehensive, dynamic measure of individual contributions that considers context, behavior, and game situation. It addresses the limitations of traditional metrics, offering a more nuanced understanding of player impact. The methodology supports scouting, tactical planning, and performance analysis, fostering data-driven decision-making in football. Its interpretability and adaptability make it suitable for both academic research and industry applications, promising to reshape talent evaluation and tactical insights.
Technical Contribution
The paper’s key innovations include the development of SPADL for standardized event representation, enabling cross-source data integration. The probabilistic modeling with CatBoost allows for calibrated, short-term impact prediction. The framework’s modular design facilitates multi-dimensional player ratings and real-time adaptation. The combination of context-aware features and probabilistic valuation sets a new benchmark for behavior-based sports analytics, opening avenues for future deep learning integration and real-time deployment.
Novelty
This is the first comprehensive framework to quantify all player actions’ value via probabilistic impact estimation, moving beyond static, event-count metrics. Unlike prior work focusing solely on goals or passes, VAEP captures the nuanced influence of behaviors within match context, providing a more holistic and actionable performance measure. Its integration of spatial, temporal, and situational features represents a significant methodological leap.
Limitations
- The framework relies heavily on high-quality event stream data; inaccuracies or missing data can distort valuations. Certain tactical movements and spatial behaviors are not fully captured, limiting scope.
- Model calibration may vary across leagues and styles, requiring retraining or adaptation. Computational costs for large-scale, real-time analysis remain high.
- Current validation is primarily on European top leagues; applicability to lower-tier or non-European leagues needs further testing.
Future Work
Future efforts will focus on integrating optical tracking data for finer spatial-temporal analysis, exploring deep learning models for improved impact prediction, and developing real-time evaluation tools. Extending the framework to other sports and lower-level leagues could broaden its utility. Additionally, incorporating tactical and psychological factors may enhance behavioral understanding, pushing sports analytics toward more holistic, AI-driven insights.
AI Executive Summary
In modern soccer, traditional metrics such as goals and assists offer a limited view of player contributions, often ignoring the broader tactical and contextual influences. Recognizing this gap, the authors introduce VAEP, a probabilistic framework that evaluates each player action based on its expected impact on scoring and conceding probabilities. Central to this approach is the SPADL language, which standardizes event descriptions across diverse data sources, enabling consistent analysis of behaviors like passes, dribbles, and shots.
The core innovation lies in modeling the short-term influence of actions through machine learning classifiers—specifically, CatBoost—trained on extensive European league data from 2012/2013 to 2017/2018. These models predict the likelihood of scoring or conceding in the near future, allowing each action to be assigned a quantitative value reflecting its contribution to the game outcome. By aggregating these values over time, the framework produces comprehensive player ratings that outperform traditional metrics in correlating with market value and talent potential.
Empirical results demonstrate that VAEP scores align closely with market valuations, with top players like Kevin De Bruyne and Eden Hazard ranking highly, despite not always leading in goals or assists. The model also effectively identifies undervalued talents from smaller leagues, aiding scouting decisions. Ablation studies confirm the importance of contextual features, such as game state and spatial positioning, in enhancing predictive accuracy.
This approach offers significant implications for sports analytics, enabling more nuanced, dynamic, and interpretable assessments of player performance. It supports strategic decision-making in recruitment, tactical planning, and performance optimization, fostering a shift toward data-driven football management. Future developments aim to incorporate tracking data and real-time evaluation, further transforming how the sport is analyzed and played.
Deep Analysis
Background
体育数据分析经历了从简单统计到复杂模型的演变。早期研究主要关注射门、助攻等稀疏事件（如Expected Goals），但忽略了行为的上下文和动态影响。随着事件流和追踪技术的发展，学界开始尝试结合空间、时间和行为特征，构建更全面的评价体系（如Elo、Player Impact Score）。然而，现有方法多依赖静态指标，难以反映比赛中的微妙变化。近年来，行为驱动的分析逐渐兴起，试图通过模型捕捉行为的潜在价值，为战术调整和球员选拔提供依据。
Core Problem
传统足球评价指标如进球数和助攻数，无法全面反映球员在比赛中的实际贡献，特别是忽略了行为的潜在影响和比赛环境。现有模型多忽略行为的上下文信息，导致评价偏差，难以动态调整行为价值，限制了其在实际中的应用。如何结合行为的短期影响和比赛局势，建立一种既全面又细粒度的评价体系，成为当前的核心难题。这不仅关系到球员的选拔，也影响战术制定和比赛策略。
Innovation
本研究的创新点包括：1）提出SPADL语言，标准化多源事件数据，简洁且完整，便于分析；2）引入基于概率的短期影响预测模型（如CatBoost），实现行为价值的动态估算；3）将行为价值转化为多维度的球员评分，支持不同时间尺度和行为类型的评估；4）结合比赛上下文信息，提升模型的适应性和解释性。这些创新突破了传统指标的局限，为行为驱动的体育分析提供了新思路。
Methodology
- �� 采用SPADL语言描述比赛事件，确保数据格式一致。• 利用CatBoost训练两个二分类模型，分别预测得分和失分概率。• 以行为发生的短期影响为基础，计算每个行为的价值（正向或负向）。• 提取空间、时间和比赛局势特征，增强模型表现。• 将行为价值按时间段聚合，形成球员整体贡献评分。• 通过交叉验证和消融实验验证模型性能，确保指标的稳健性。
Experiments
数据集涵盖欧洲五大联赛2012/2013至2017/2018赛季，采用训练集和验证集进行模型调优。模型性能通过分类准确率、AUC和校准度评估。与传统统计指标（如Goals/90）对比，验证VAEP在识别高影响力球员中的优越性。还进行了特征重要性分析和鲁棒性测试，确保模型的泛化能力。模型在不同联赛和风格下均表现出良好的适应性。
Results
VAEP评分在2017/2018赛季中，Kevin De Bruyne、Eden Hazard等球员排名靠前，市场估值显著高于传统指标。模型预测得分和失分概率的校准度超过85%，优于基于事件统计的指标。消融实验显示，加入比赛局势和空间特征后，模型性能提升约10%。此外，模型还能识别出被低估的潜力球员，为俱乐部提供了有价值的参考依据。
Applications
该方法可广泛应用于球探、战术分析和运动表现优化。俱乐部可以利用VAEP评分进行球员筛选和转会决策，提升投资回报。教练团队可以借助模型调整战术策略，实时监控球员表现。未来，结合实时数据和深度学习，有望实现比赛中的即时行为评价，推动足球运动的智能化管理。
Limitations & Outlook
模型高度依赖事件流数据的质量，数据偏差或缺失会影响评估准确性。某些战术移动和空间行为难以捕捉，导致价值估算不足。模型在非欧洲联赛或低级别联赛中的适应性尚未验证，未来需扩展多模态数据融合和跨区域验证。计算成本较高，实时应用仍需优化。
Plain Language Accessible to non-experts
想象你在一家工厂工作，每个工人都在做不同的任务。有些任务直接带来产品，比如装配零件；有些任务虽然不直接生产，但会影响整体效率，比如整理工具。传统评价只看工人完成了多少装配件（就像足球只看进球数），但忽略了他们在工作中的配合、策略和环境。本文提出一种方法，像是给每个工人每个动作打分，考虑他们的任务对工厂整体生产的影响。通过分析每个动作如何改善或阻碍生产流程，能更全面评价工人的贡献。这就像用智能系统分析每个工人在工厂中的表现，不仅看结果，还看过程和环境，从而找到最有效的工作方式。
ELI14 Explained like you're 14
想象你在玩一款足球游戏，你的目标不仅是踢进球，还要帮队友创造机会。以前，大家只看谁进了球或助攻了，但其实很多重要的动作都没有被重视。现在，有个聪明的机器人可以帮你分析每次传球、盘带、射门的价值。它会考虑比赛的情况，比如你们的比分、场上的位置，告诉你每个动作对赢球的帮助有多大。比如，你传球到前场，虽然没有直接得分，但让队友更容易射门，这个动作就会得到高分。这个系统用数学模型预测每个动作未来会带来多少得分机会，然后把所有动作的价值加起来，帮你知道谁其实是场上的大英雄。这样，教练可以更聪明地安排战术，球员也能知道自己哪些动作最重要。虽然还不能实时用，但未来这个技术会让足球变得更科学、更有趣！
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.