TL;DR
This paper introduces a Bayesian approach to quantify individual football player abilities for specific in-game events, going beyond simple aggregate statistics. By extending a hierarchical model with these player abilities, they achieve improved predictions of match outcomes, demonstrating the value of granular player assessment.
What Problem Does This Paper Address?
In football, understanding a player's true ability is crucial for team management, transfers, and even betting. Existing methods often rely on aggregate statistics (like total goals scored) or general performance ratings, which can be misleading. For instance, a player with few minutes might appear less skilled than they are, or two players who always play together might be indistinguishable in their abilities. This paper aims to address this by developing a method to infer a player's ability for specific event types (e.g., scoring a goal, making a tackle) and to use these inferred abilities to build more accurate models of team performance and match outcomes.
Methods
The core of the paper is a Bayesian hierarchical model. For individual player abilities, they use a Poisson model to capture event occurrences and infer player abilities using variational inference (VI). VI is chosen for its scalability with large datasets, transforming posterior inference into an optimization problem. They introduce a mean-field variational family, assuming player abilities are independent.
To model team performance, they extend the Bayesian hierarchical model of Baio and Blangiardo (2010). This model uses a Poisson-log normal distribution to predict goals scored by home and away teams, incorporating team-specific attack and defense abilities, and a home advantage. The key innovation is incorporating the previously inferred player abilities into the team's scoring intensity. This is done by defining a function f(q(∆)) that aggregates player abilities relevant to scoring (e.g., "Goal," "Shots," "ChainEvents") for the players on the field.
Datasets
The primary dataset used is touch-by-touch data from the English Premier League covering the 2013/2014 and 2014/2015 seasons. This dataset records every touch in a game, including time, team, player, event type, and outcome. Approximately 1.2 million events are included.
Key Insights and Ideas
- Granular Player Abilities: The model can distinguish between players even if they always play on the same team, a limitation of some previous methods like TrueSkill.
- Interpretable Abilities: The inferred player abilities are visualized through marginal posterior variational densities, allowing for direct comparison and understanding of differences between players for specific skills.
- Time-Weighted Abilities: Incorporating the fraction of time a player spends on the pitch (
Ti,k) is crucial for accurately modeling their contribution and impact. - Predictive Power: Including latent player abilities significantly improves the model's ability to predict match outcomes (specifically, over/under 2.5 goals) compared to a baseline model without these abilities.
- Defensive Midfielders vs. Goalkeepers: For "GoalStop" ability, the model suggests that investing in defensive midfielders might be more effective than goalkeepers, based on their inferred abilities.
Key Results
- The model successfully infers player abilities for various event types, with rankings for "Goal" and "GoalStop" showing sensible results that align with expert opinions. For example, top goal scorers are ranked highly, and players with less playing time but high inferred ability are identified.
- The extended hierarchical model, incorporating player abilities, showed improved predictive performance. For instance, in predicting over/under 2.5 goals, the Area Under the ROC Curve (AUC) was consistently higher for the model with player abilities across multiple prediction blocks.
- In a simulated betting scenario on the over/under 2.5 goals market for the 2014/2015 season, the model with latent player abilities yielded a cumulative return of £4486.73, while the baseline model resulted in a loss of £378.54.
Limitations and Future Work
- Event Type Grouping: The choice and grouping of event types were based on expert football analysts, and different groupings might yield different results. The independence assumption between grouped event types is also a simplification.
- Variational Inference Approximation: The use of a mean-field variational family assumes independence between latent variables, which might underestimate uncertainty. Future work could explore structured variational inference to account for correlations.
- Player Injuries/Transfers: The model assumes latent abilities remain constant even if a player is injured or transfers teams. Future work could explore random walks for variational parameters to capture these changes.
- Starting Line-up Prediction: Predictions currently rely on expert-provided starting line-ups. Future work could explore using predicted line-ups based on historical data.
- Penalty Goals: Penalties were not modeled separately, as they were deemed to add complexity without significant impact on the top goal-scoring rankings. Future work could investigate their separate modeling.
- Model Parsimony: Home effects for some event types might be negligible and could be removed to simplify the model.
- Interactions: Defining all possible interactions between event types is a complex and open research question.
- Application to Other Sports: The authors believe the approach is applicable to other sports like basketball, hockey, and rugby.
Summary updated