Visualizing a Team's Goal Chances in Soccer from Attacking Events: A Bayesian Inference Approach

Gavin A. Whitaker, Ricardo Silva, Daniel Edwards

TL;DR

This paper introduces a novel Bayesian inference approach to analyze soccer team attacking plays using only event data (assists and chances), rather than expensive full match tracking. The method effectively models team abilities, player contributions, and spatial patterns of chance creation, offering a computationally efficient way to visualize and understand team tactics.

What Problem Does This Paper Address?

Understanding a soccer team's attacking style and effectiveness is crucial for analysis, strategy, and even betting. Traditional methods often rely on comprehensive, expensive data like full player tracking, which isn't always accessible. This paper aims to bridge this gap by developing a method that can infer team abilities, identify key players, and pinpoint where on the pitch chances are created and taken, using only readily available attacking event data (goals, assists, and the locations of these events). This allows for a more accessible and scalable analysis of team tactics.

Methods

The authors propose a two-part Bayesian inference model:

  1. Chance Occurrence Model: A Poisson model is used to capture the number of chances a team creates within specific time blocks of a match. This model incorporates parameters for team ability, opposition ability, and a home advantage effect.
  2. Chance Composition Model: For each chance, the model infers:
  3. Assist and Chance Players: Using a "Multinoulli" (categorical) distribution, the model identifies which player made the assist and which player took the chance. This is modeled hierarchically, similar to topic modeling techniques.
  4. Spatial Locations: A Gaussian mixture model is employed to capture the spatial distribution of assist locations and the relative difference between assist and chance locations. This allows for the visualization of where players tend to operate when creating or finishing chances.

The entire model is fitted using a Bayesian inference approach, specifically a Gibbs sampling strategy, allowing for uncertainty quantification.

Datasets

The primary dataset used is Stratagem Technologies' Analyst data for the 2016/2017 English Premier League season. This dataset contains significant match events, including goals, cards, and chances created, along with the time, team, player involved, and spatial location of the event. The authors note that their method uses a subset of this data (assist and chance locations) compared to previous work that utilized complete touch data.

Key Insights and Ideas

  • Data Efficiency: The core innovation is demonstrating that valuable insights into team tactics can be extracted from relatively coarse attacking event data, significantly reducing data collection and processing costs compared to full tracking.
  • Player-Centric Spatial Analysis: Unlike some prior work that focused on team-level spatial dynamics, this method explicitly models individual player contributions to chances, both in terms of who is involved and where they operate.
  • Dynamic Tactical Analysis: By dividing matches into time blocks, the model can reveal how a team's or player's style of play evolves throughout a game (e.g., increased chance creation towards the end of halves).
  • Uncertainty Quantification: The Bayesian framework naturally provides measures of uncertainty, allowing analysts to understand the reliability of their findings and how it changes as more data becomes available.
  • Visualization of Tactics: The model's outputs, particularly the spatial density plots and radar plots, offer intuitive ways to visualize and compare team and player attacking behaviors.

Key Results

  • The model successfully captures team abilities to create chances, with top teams like Chelsea, Tottenham, Manchester City, and Liverpool showing higher abilities, particularly in later game blocks (t5, t6).
  • The model can identify individual player roles and spatial preferences. For example, Christian Eriksen's assist locations were visualized across different game blocks, showing potential shifts in his play style.
  • The home effect is consistently positive across all game blocks, indicating teams tend to create more chances at home.
  • The model's predictive distributions for chance creation were shown to fall within observed values, suggesting good model fit and predictive performance.
  • The analysis of specific matches (e.g., Liverpool vs. Crystal Palace) demonstrated the model's ability to identify potential weaknesses and predict player involvement in chances at those locations.

Limitations and Future Work

  • Computational Simplification: The authors chose to fix the means of the Gaussian mixture components for computational ease. Relaxing this and inferring means simultaneously with other parameters is a potential future direction.
  • Player Dependence: The current model assumes independence between the assist player and the chance player. Future work could incorporate this dependence, potentially using network analysis techniques.
  • Temporal Smoothing: Smoothing techniques could be applied to player and mixture component probabilities to allow for smoother variations over time, rather than discrete block-based changes.
  • Anomaly Detection: The framework could be extended to detect unusual player patterns within specific fixtures or blocks, providing insights into tactical surprises or deviations.
  • Data Volume: The authors note that the reduction in variance of posterior densities is limited by the relatively small number of observations per player. Using data over multiple seasons would help reduce this uncertainty.

Summary updated

BibTeX
@inproceedings{whitaker2018visualizing,
  author = {Gavin A. Whitaker and Ricardo Silva and Daniel Edwards},
  title = {{Visualizing a Team's Goal Chances in Soccer from Attacking Events: A Bayesian Inference Approach}},
  booktitle = {Big Data},
  year = {2018},
  url = {https://openreview.net/forum?id=WUoEcGH7if},
}
Figure