Positions are — let’s be honest here — a terrible way to think and talk about football. They’re supposed to suggest players’ job descriptions, even whole skill sets, but all they really tell you is approximately what patch of grass they run around on.
It’s like sending employers a CV that says your desk was over by the potted plant in the north east corner of the office, a few rows behind the accountants and across from marketing. You might get a few weird looks at that interview.
Advertisement
What if, instead of positions, we had a more detailed way to describe footballers’ tactical roles?
Better yet, what if these roles are based not on subjective labels for starting formations but on the type and location of a player’s actual touches?
“Players can be called one position and then do a totally different thing than another player who plays the same position. Sometimes, people don’t agree on what position a player is,” says Mike Imburgio, who started writing articles for American Soccer Analysis about using data to define player roles while earning a neuroscience PhD.
“So I got really fascinated by the idea of, ‘What kind of player is this? What does this player like to do, and what are they tasked with doing?’.”
Together with another American Soccer Analysis contributor named Sam Goldberg, Imburgio developed a set of data-based player roles that serve as the basis for a public scouting model called DAVIES. (Goldberg now works as a data scientist for the New York Red Bulls of MLS; Imburgio does some football recruitment consulting alongside his day job in data science.)
“I settled early on a two-step process that involves taking a lot of data about a lot of different players and condensing it into a smaller number of variables. That’s called dimension reduction. Rather than trying to explore 30 different variables for all of these players, you can look at them on literally a map,” Imburgio says. “From there, you use a clustering algorithm to group players into different roles.”
During pre-season, The Athletic used an approach similar to Imburgio’s to develop our own unique set of 18 player roles.
Like DAVIES, these roles are built on FBref’s trove of public StatsBomb data, which covers players from the big five European domestic leagues — England, France, Germany, Italy and Spain — over the last five seasons.
Advertisement
The resulting player classification model — a role model, if you will — has all kinds of potential uses.
It can put players’ numbers in context by determining who we should compare them to, since different roles have different opportunities to record stats. It can also help us investigate how players, squads or entire leagues evolve over time, how roles and formations interact, or what your club might need on the transfer market.
We hope the player roles will help writers and readers alike find new ways to think about data and tactics, or at least give us something more fun to talk about than positions.
But first things first. Here’s an introduction to The Athletic’s 18 player roles, what they mean and where they came from…
The 18 player roles
Central attackers
- Finisher: Focused on getting in the box and finding shots. May be good on the ball but isn’t very involved in possession. Examples from the 2021-22 season: Erling Haaland, Lautaro Martinez, Romelu Lukaku
- Target: Involved in his team’s build-up play, especially in the air or by dropping towards the ball. Typically more of a goal threat than a creative passer or dribbler. Patrik Schick, Dominic Calvert-Lewin, Gianluca Scamacca
- Roamer: Likes to drop deep or wide to create for team-mates when he’s not attacking the goal. Includes mobile strikers, false nines, and hybrid wingers. Harry Kane, Kai Havertz, Richarlison.
Wide attackers
- Wide threat: Stretches the back line and gets into the penalty area. Does some crossing but likes receiving the final ball as much as playing it. Kylian Mbappe, Vinicius Junior, Mohamed Salah
- Unlocker: Likes to play field marshal in the opponent’s half. Big on crosses, switches, and forward passes. More of a provider than an off-the-ball runner. Mason Mount, Neymar, Lorenzo Pellegrini
- Outlet: Gets on the end of dangerous passes but usually plays it safe on the ball. Takes a lot of touches in midfield or close to goal, not much in between. Draws fouls. Jadon Sancho, Jack Grealish, Joao Felix.
Advanced midfielders
- Box crasher: Doesn’t touch the ball much except when close to goal. More of an off-ball runner than a passer, but will play passes into the box. Likes to dribble. Phil Foden, Florian Wirtz, Serge Gnabry
- Creator: Looks to break lines with aggressive passes. Frequently central to his team’s play in the attacking half. Bruno Fernandes, Kevin De Bruyne, Sergej Milinkovic-Savic
- Orchestrator: Prefers midfield circulation to finding the final ball. Likes shorter, higher-percentage passes. Not afraid to do some defending. Bernardo Silva, Jude Bellingham, Nicolo Barella.
Deep midfielders
- Box to box: Defends low but also gets upfield to receive progressive passes. Not very heavily involved in possession, usually cautious on the ball. Fabinho, Wilfred Ndidi, Eduardo Camavinga
- Distributor: Favours longer, more direct passes and switches. Active in the opponent’s half more than the defensive third. Joshua Kimmich, Fabian Ruiz, Youri Tielemans
- Builder: Serves as the main circulation hub in the build-up and as a stopper at the base of midfield. Declan Rice, Rodri, Aurelien Tchouameni.
Wide defenders
- Overlapper: Gets into the final third, dribbles, and hits crosses. Would rather receive long passes and recycle them than play long himself. Includes a lot of wing-backs. Alphonso Davies, Achraf Hakimi, Reece James
- Progressor: Attempts long balls and progressive passes as well as crosses. Active in possession at both ends. Trent Alexander-Arnold, Andrew Robertson, Joao Cancelo
- Safety: Stays home, getting involved in the defensive third while not taking many risks on or off the ball. Includes hybrid centre-backs or elbow backs. Benjamin Pavard, Timothy Castagne, Aaron Wan-Bissaka.
Central defenders
- Aggressor: Likes to tackle and dribble. Not big on long passes or clearances. Includes a lot of outside centre-backs. Alessandro Bastoni, Lucas Hernandez, Ronald Araujo
- Spreader: Plays longer, more direct passes. Will sometimes step out in possession or to close down in defence. Marquinhos, Eder Militao, Jules Kounde
- Anchor: Plays safe passes and does his defending close to goal, especially with clearances and blocks. Ruben Dias, Matthijs de Ligt, Milan Skriniar.
How it works…
OK, deep breath. This is about to get nerdy.
First, we took hundreds of season-level stats for players and their teams, mixed and matched them, and came up with 14 that were useful for separating players into broad position groups, for example “wide defender” or “advanced midfielder”. This initial set of stats focused on what part of the pitch players tend to take their touches in, what kinds of passes they attempt, and how successful they are at passing and receiving.
Passing and receiving percentages may sound like a measure of skill — and to an extent, they are — but not so much as they tell you what kind of passes a player is involved in and where. That was important. The goal wasn’t to group players by how good they are at football, only how they fit into their team. (It mostly worked — by Transfermarkt values, there’s a pretty even mix of talent across each of the final 18 roles.)
We fed the first 14 stats into a dimension reduction algorithm called UMAP that arranged 7,000-plus player seasons into a two-dimensional map where players with similar stats were located near one another. Then we used a clustering algorithm called GMM to divide that map into six primary groups of players.
Up to this point, the process was more or less the same as the one used to classify team playstyles for The Five Kingdoms of Football.
As intended, the six primary clusters broke down along position lines: centre-backs and full-backs each had a group almost entirely to themselves, while the other four clusters belonged predominantly to strikers, wingers, attacking midfielders and defensive midfielders (central midfielders were divided between the last two groups).
Advertisement
But the algorithms didn’t know players’ positions. The clusters were derived entirely from stats, so unusual players such as Inter Milan’s Denzel Dumfries, who lines up as a wing-back but operates more like a winger than a defender, were grouped by how they played instead of how they were labelled positionally.
The next step was to break the six primary groups into smaller secondary clusters that would become our 18 player roles.
There was no right or wrong final number of roles, but dividing each primary group into three smaller ones of comparable size made the roles specific enough to capture different styles but large enough to stay fairly stable.
This second stage followed the same process as the first: for each primary cluster, we selected new stats that helped to map and cluster players based on broad stylistic similarity. Each of the six used different stats — dribbles per touch, for example, proved useful for separating attacking midfielders but not defensive midfielders.
The last step was to train a separate model to re-classify players into the 18 roles using all 42 stats that went into the clustering processes, ranging from the percentage of a player’s touches taken in the opponents’ penalty area to how many of his defensive actions are clearances or blocks. Every stat was normalised by team rates, which didn’t make much difference most of the time but may help adjust for more eccentric team playstyles.
After that, all that was left was to name the roles, which is both more important and harder than it sounds.
Even though the roles capture player types that you can probably imagine on the pitch, English doesn’t have as colourful a vocabulary for talking about player roles as some other languages.
“If you take two seasoned scouts and show them the same player that they have no preconceptions of and say, ‘Give me a description of how this player plays in three words’,” Imburgio says, “it’s highly likely that they’re not going to line up.”
Advertisement
The names The Athletic settled on are a little corny and ambiguous by design — they’re supposed to evoke a role’s vibe without getting all pseudo-scientific-sounding in trying to nail down its particulars, which take more than a couple of words to do justice to. (Where role names do sound familiar, like “Target” or “Box to box”, it’s because those concepts happened to map pretty well onto a certain cluster.)
How well did the whole project turn out?
There’s no sure way to say. It’s not too hard to get a feel for what sort of player belongs in which role and how the roles are different, which is the most essential thing. If the roles aren’t interpretable, they won’t be useful.
The model is confident and consistent in its ability to assign roles to players, and the roles themselves are stable enough over time.
On average, a player will stay in the same role from one season to the next 61 per cent of the time, or 46 per cent if he moves to another club. That seems reasonable, given the number of roles and all the usual vagaries of team-mates, tactics and coaching.
Players with distinctive styles, such as Roberto Firmino, pop up at almost exactly the same place on the style maps every season. As for guys whose tactical role has evolved, such as Harry Kane, who’s been dropping deeper to provide for team-mates over the last couple of seasons, the model reflects that.
These are signs that it’s doing something right.
But the real test of how well the player roles project turned out will be how useful it is for The Athletic’s writers and readers, and we’ll find that out in future articles.
“No matter what you end up with, it’s always going to be something that people can argue with,” Imburgio says. “But if you can make use of it, that’s all that matters.”
(Design: John Muller and Sam Richardson)
