Clustering professional football players using Gaussian mixture models: A case study of the English Premier League 2024/25

summary

The gist

Clustering professional football players using Gaussian mixture models: A case study of the English Premier League 2024/25 analyzes player performance data using a Gaussian mixture model to identify…

In short

The study used a Gaussian mixture model to cluster 314 English Premier League players based on their performance statistics. This method revealed seven player archetypes and identified hybrid roles, meaning some players fit into multiple statistical categories. The findings show that model-based clustering offers a flexible way to see the hidden structure in complex player data.

Key concepts

Gaussian Mixture Model (GMM)
A statistical tool used here to group players based on their performance. Unlike simple grouping methods, GMM assigns each player a probability of belonging to several different groups at once. This allows the model to capture overlapping roles rather than forcing a player into just one category.
Principal Component Analysis (PCA)
A technique used to simplify the data by reducing many complex statistics into a smaller set of independent components. This helps manage 'curse of dimensionality' and removes redundancy among player stats, making the subsequent clustering process more accurate and manageable.
Hybrid Role
A player role that spans multiple statistical categories, meaning they exhibit characteristics from several identified clusters. The model quantifies this by checking if a player's probability of belonging to more than one cluster exceeds a set threshold, revealing players who are stylistically ambiguous.
Soft Clustering
A type of clustering where data points are not assigned to a single, fixed group but rather have a probabilistic membership score for every group. This is superior for real-world scenarios like sports because it acknowledges that individuals often possess mixed traits instead of fitting neatly into one defined style.

Terminology used across episodes

This episode discusses

The paper

Clustering professional football players using Gaussian mixture models: A case study of the English Premier League 2024/25 · Read on arXiv

Harry O’Donnell, Antony M. Overstall

University of Bath · University of Southampton

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Clustering professional football players using Gaussian mixture models".

Jane: Here's the gist of the paper. Clustering professional football players using Gaussian mixture models: A case study of the English Premier League 2024/25 analyzes player performance data using a Gaussian mixture model to identify probabilistic cluster memberships, revealing hybrid roles and stylistic ambiguity in modern football Model Application and Data Preparation The study applies a Gaussian mixture model (GMM) to principal component representations of player statistics for 314 outfield players from the 2024/25 English Premier League season.

Tom: Unlike hard-assignment methods such as k-means, the GMM provides probabilistic cluster memberships, allowing uncertainty and role hybridity to be quantified.

Jane: The data preparation involved using 8 tables of statistics from the FBref database and excluding goalkeepers and players with fewer than 800 minutes played.

Tom: The feature set consisted of 123 season totals, excluding rate, percentage, and average variables such as pass completion percentage.

Paper summary: Tom: So we're looking at this paper titled Clustering professional football players using Gaussian mixture models: A case study of the English Premier League two thousand twenty-four/twenty-five <ref:2611.17024#pg1,Clustering professional football players using Gaussian mixture models: A case study of>. Basically, they're taking all this performance data from three hundred fourteen outfield players in the Premier League for that season and applying a Gaussian mixture model to it.

Jane: That sounds like a big picture approach, Tom. What's the main idea they're trying to get across? I mean, what is their central claim about how we look at these players?

Tom: The core thesis is that most current clustering methods just shove every player into one single group, but in football, that’s not true because players do hybrid roles and mix styles. This paper uses a Gaussian mixture model to give us probabilistic cluster memberships instead of hard assignments like k-means.

Lu: What's interesting about the GMM here is that it produces a posterior probability distribution over cluster membership for each player, which means we get a measure of uncertainty about where they truly fit, which is different from just giving them one label > (<ref:2611.17024#pg2>).

Meng: But so what does this actually mean for deployment? If we're trying to use this to build something real, how do we handle that uncertainty in the model output? Does that probabilistic part add computational overhead or just more complexity?

Lalam: The paper states they can quantify uncertainty and role hybridity, and they even found a subset of players whose profiles span multiple clusters > (<ref:2611.17024#pg1>). This is the objective measure of tactical flexibility mentioned in the abstract.

Jane: So if we're looking at what this means for someone who just watches football, how does this probabilistic idea translate into something tangible about a player's style? What are these clusters they found?

Tom: They identified seven player archetypes in the full dataset, and one of the largest groups is Cluster three which they call Rotational and utility players with one hundred seven players > (<ref:2611.17024#pg1>). Another notable group is Cluster five which they define as Creative forwards with fifty players characterized by high scores on goals and assists above expected values > (<ref:2611.17024#pg3>).

Paper summary: Lu: I want to dig into that hybrid role quantification because that seems like the most novel part. They identified twenty-four players whose membership probabilities exceeded a twenty-five percent threshold across more than one cluster > (<ref:2611.17024#pg2>). That's a tangible way to measure stylistic ambiguity in the data.

Meng: Twenty-four players is a specific count, but what about the rest? They also noted sixteen players, which is five point one percent of the total, had their maximum posterior probability below zero point six five > (<ref:2611.17024#pg2>). That’s where things get really interesting for me at scale because it shows profiles that are most ambiguous.

Lalam: The paper also looked within-position, and they found that GMM works for sub-archetypes within groups like defenders or forwards > (<ref:2611.17024#pg2>). For defenders, Cluster three was the most tactically coherent group > (<ref:2611.17024#pg3>).

Jane: It sounds like they're not just giving us neat boxes; they're showing us a spectrum of roles where players blend different styles depending on the context > (<ref:2611.17024#pg2>). If you were a scout, what would this tell you about how to look at a player who doesn't fit neatly into any one category?

Tom: It tells you that instead of pigeonholing someone, you see them as belonging to a probability distribution across several styles > (<ref:2611.17024#pg1>). And when we look at the numbers, the median maximum posterior probability for every player was zero point nine nine five > (<ref:2611.17024#pg3>). That’s a high number, suggesting that most players are reasonably well-defined in their roles.

Lu: But what about those outliers they found? They mentioned Mohamed Salah being isolated as a singleton cluster at the point where they minimized the Bayesian information criterion > (<ref:2611.17024#pg3>). That shows the model can isolate statistical outliers instead of forcing them into an existing group.

Meng: Isolating one player is fine for research, but what about deploying this to production? If we have a player like Salah who falls outside the main structure, does that mean our system breaks or just flags him as 'exception'? We need to know what happens when the data doesn't conform > (<ref:2611.17024#pg2>).

Paper summary: Lalam: The paper points out that this approach is readily transferable to other sports and domains where individuals exhibit overlapping characteristics and cannot be adequately represented by hard cluster assignments alone > (<ref:2611.17024#pg3>). It suggests the mathematical framework itself is more valuable than just the football data.

Jane: So, when we look at the title again—Clustering professional football players using Gaussian mixture models—it really boils down to moving from a single answer to a probability of answers > (<ref:2611.17024#pg1>). It’s about acknowledging that modern football is messy and hybrid.

Tom: Exactly, and the implication for us as engineers is that we can use this framework to build similarity tools on top of those posterior probabilities to identify the statistically most similar player > (<ref:2611.17024#pg3>). That's a clear next step for model-based clustering.

Lu: And the methodology itself, using principal component analysis to reduce dimensionality down to eight components, was a smart choice because it mitigated multicollinearity and avoided the curse of dimensionality > (<ref:2611.17024#pg3>). That’s solid engineering work underpinning the statistical finding.

Meng: From an operational standpoint, reducing the feature space by using those eight principal components means our runtime cost for processing a single player's statistics is much lower than if we fed it all one hundred and twenty-three original variables > (<ref:2611.17024#pg3>). That’s a good trade-off if the dimensionality reduction works as advertised.

Lalam: The paper doesn't actually give us a specific cost metric, but the fact that they used PCA to explain seventy-five percent of the variation shows they are making an informed choice about what information is actually relevant for clustering > (<ref:2611.17024#pg3>).

Jane: So, to wrap up this discussion on Clustering professional football players using Gaussian mixture models: A case study of the English Premier League two thousand twenty-four/twenty-five it really shows how model-based clustering can uncover latent structure in high-dimensional performance data > (<ref:2611.17024#pg3>). It moves beyond just finding groups and starts quantifying the uncertainty behind those groups.

Tom: Right. It’s about making that uncertainty audible, not just mathematically present > (<ref:2611.17024#pg3>). And that probabilistic approach is something we can definitely take with us to other sports or even just other domains where individuals perform hybrid roles and hard assignments don't fit the reality > (<ref:2611.17024#pg3>). That’s the broader utility here.

Conclusion: Tom: So we're wrapping up this look at the Gaussian mixture model study on Premier League players, which is titled "Clustering professional football players using Gaussian mixture models: A case study of the English Premier League two thousand twenty-four/twenty-five."

Jane: That whole thing boils down to taking a massive amount of player performance data and using this statistical tool to figure out what different types of players actually are.

Lu: The authors Harry O’Donnell and Antony Overstall, they used principal component analysis first to shrink the huge list of stats down into eight main components before applying the Gaussian mixture model.

Meng: So, it’s like they took a hundred variables and boiled them down to eight key factors for clustering. What was the actual data set size for this study?

Tom: They worked with three hundred fourteen outfield players from that two thousand twenty-four/twenty-five English Premier League season. And they used twelve of those statistics, excluding things like rates or percentages.

Jane: The core thing they found is that most existing methods force a player into one single box, but this model shows you the probability of them fitting into several different styles at once.

Lu: Specifically, the model identifies seven distinct archetypes in the full dataset. They point out Cluster Three as having one hundred seven players because it represents rotational and utility players with mixed statistical profiles.

Meng: I mean, that sounds useful for scouting, but what about those players who don't fit into any of those seven main groups? The study did mention them.

Jane: Oh, they did. They found twenty-four players whose membership probability was above twenty-five percent across more than one cluster, which is how they measure that hybrid role or stylistic ambiguity.

Tom: That means we aren't just getting a label; we’re getting a measure of uncertainty about the player’s role, and they found that most players have a high certainty score in their assigned cluster.

Lu: They also looked at specific positions, finding that for defenders, Cluster Three was the most tactically coherent group. And for forwards, one player called Mohamed Salah stood out as an outlier.

Meng: So what does this actually change for a team manager or a scout who is trying to make decisions? Is it just more data points on paper?

Jane: It’s about moving from simple categorization to understanding the real flexibility of the squad, showing that players aren't always one-dimensional performers.

Tom: Exactly. The authors suggest this approach is useful for any domain where individuals have overlapping characteristics and hard assignments don't capture the reality of how they play.

Lu: And looking ahead, they suggest a way to build a player similarity tool directly on top of those posterior probabilities to find the statistically most similar player.

Meng: That sounds like a solid engineering path if you can actually make that run efficiently without blowing up your resources at scale.

Jane: It’s interesting because it takes something complex, high-dimensional performance data and makes it interpretable by showing you the spectrum of roles instead of just a single, rigid classification.

More episodes

← Home