In the world of football (soccer) analytics, we’ve moved far beyond simple box scores. However, most advanced models still focus on the team—predicting wins, losses, or expected goals at a collective level. A new paper by Gavin A. Whitaker and colleagues, "A Bayesian approach for determining player abilities in football," shifts the lens toward the individual, using a sophisticated Bayesian framework to answer the age-old question: How much is a specific player actually contributing to the scoreline?
TL;DR
The researchers developed a Bayesian Poisson model to infer individual player abilities across various event types (like scoring or tackling) using touch-by-touch data. By scaling this model with Variational Inference and integrating it into a hierarchical team-level model, they significantly improved the accuracy of match goal predictions and demonstrated a profitable betting strategy in the "over/under 2.5 goals" market.
What Problem Does This Paper Address?
Traditional player rating systems, like the EA Sports Player Performance Index, often condense a player's entire contribution into a single number. This masks specific skills and often rewards players simply for being on a winning team.
Furthermore, popular skill-rating algorithms like TrueSkill™ struggle with "teammate confounding"—if two players always play together, the model can’t easily tell who is the real driver of success. This paper seeks to disentangle individual talent from team performance by modeling the frequency of specific events (goals, passes, interceptions) while accounting for playing time, home advantage, and the quality of the opposition.
Methods
The authors employ a Poisson model to capture the occurrences of event types. Under the hood, the rate at which a player performs an action (like scoring) is determined by:
- Latent Ability ($\Delta$): The unobserved "true" skill of the player for that event.
- Team Effects: The influence of a player’s own teammates and the opposition’s ability to stop them.
- Time Factor ($\tau$): The fraction of the match the player was actually on the pitch.
- Home Advantage: A constant effect added to the rate for the home team.
To handle the massive parameter space (hundreds of players and millions of events), the authors use Variational Inference (VI) instead of traditional Markov Chain Monte Carlo (MCMC) sampling. VI turns the inference problem into an optimization problem, making it computationally feasible to update player abilities as new data arrives.
Finally, they extend the well-known Baio and Blangiardo (2010) hierarchical model. While the original model only looked at team-level attack/defense stats, this version "plugs in" the inferred individual player abilities to predict how many goals a specific lineup is likely to produce.
Datasets
The study utilizes a massive "touch-by-touch" dataset from the English Premier League (EPL):
- Seasons: 2013/2014 (used for training/learning abilities) and 2014/2015 (used for prediction/validation).
- Volume: Approximately 1.2 million individual events across 380 games per season.
- Event Types: 39 types, including goals, passes, and tackles. The authors also created custom "grouped" events like GoalStop (defensive actions that prevent goals) and ChainEvents (actions leading up to a shot).
Key Insights and Ideas
- The Suarez Outlier: In the 2013/14 season, the model identified Luis Suarez not just as the best striker, but as a statistical outlier whose inferred ability was significantly higher than any other player in the league.
- Defense is Harder to Quantify: The model found that the gaps between the top-ranked defensive players (in "GoalStop" ability) were much smaller than the gaps between top attackers. This suggests that offensive brilliance is more individualized, while defensive success might be more of a collective team effort.
- Tracking Evolution: By using a "block" structure for predictions, the model successfully tracked the rise of players like Harry Kane, whose inferred scoring ability spiked as he moved from a fringe player to a regular starter.
- Lineup Sensitivity: Because the model uses individual abilities, it can react to team news. If a star striker is benched for a youngster, the model’s predicted goal rate for that team drops immediately, whereas a team-level model would remain "blind" to the lineup change.
Key Results
The model was validated by predicting whether a match would have over or under 2.5 goals in the 2014/15 season:
- Predictive Accuracy: The model including latent player abilities consistently outperformed the baseline team-only model. For example, in one prediction block, the player-augmented model achieved an AUC (Area Under Curve) of 0.68, compared to 0.55 for the baseline.
- Financial Validation: In a simulated betting strategy on the "over/under" market, the player-ability model generated a cumulative return of £4,486.73 (assuming £100 bets), while the baseline model actually lost money (-£378.54).
Limitations and Future Work
The authors acknowledge several areas for improvement:
- Independence Assumption: The model assumes different event types are independent when grouped, which ignores the nuanced "flow" of a football match.
- The Transfer Problem: Currently, the model assumes a player’s latent ability remains constant even if they move to a different team or league, which may not account for tactical fit.
- Correlations: The current "mean-field" variational approach assumes player abilities are independent. Future work could use more complex approximations to model how one player's presence might actually boost the ability of another (the "playmaker" effect).
Reference: Whitaker, G. A., Silva, R., Edwards, D., & Kosmidis, I. (2020). A Bayesian approach for determining player abilities in football. arXiv:1710.00001v2 [stat.AP].
Summary updated