Attacking stats are role-dependent too
A quick note before we start. A couple of weeks ago my Twenty3 colleagues presented some research we’d done on ‘How to break down a set defence’ at the StatsBomb Innovation in Football conference. The full paper is now online here.
For a good couple of years, it’s been accepted that football players’ defensive statistics are extremely dependent on the role that they’re asked to play. It’s not enough to sort a table by the numbers of tackles per 90 to order players as good defenders to bad defenders — that’ll just give you the ‘search and destroy’ DMs at the top and everyone else below them. Or, to be more specific, the names at the top will probably be ‘good’ but being further down the list doesn’t necessarily mean that they’re not good.
In the realm of attacking stats, we’ve traditionally been far readier to take the numbers at face value. A large part of this is just that it seems to work. Forwards with high numbers of expected goals are generally the best forwards. Attacking midfielders with high numbers of expected goals assisted are generally the best attacking midfielders.
But part of it, I think, is that there are fewer ways to skin a cat in attack than defence, and that’s blinded us to the effect a player’s role within a team can have on their stats with forwards. As an attacker, your job is generally to score, or directly contribute to other people scoring. You do the first by shooting and shooting alone; you do the second by passing. That’s two types of output you need to look at — in defence there can be three, four, five, as well as the very large output that isn’t measurable, that of ‘I’m defending this space so well that the opposition isn’t going to attempt going through it at all’.
It’s very rare for players in forward positions to stray from their usual roles too. Roberto Firmino is one of the few examples, a nominal central striker who’s usually outdone in expected goals by the two teammates flanking him. He’s such a notable example, though, that the entire footballing world knows about him and naturally looks to other stats. He’s involved in build-up play, so we look at his passing stats; he presses from the front, so we look at his defensive stats.
We know, though, that there’s greater variety in role as we take one step back from centre-forwards. Players who feature in attacking midfield positions may actually be something like second strikers. Mason Mount, for example, has around 0.36 expected goals for himself and 0.12 expected goals assisted for others, according to StatsBomb numbers. He presses quite a lot too, to add to the confusion about the ‘type’ of AM he is.
Particularly in the last 18 months or so, there has been a rapid development of different types of ‘Expected Possession Value’ models (see: xThreat by Karun Singh, a Possession Value framework from Nils Mackay at Opta Pro, the (slightly) older xGChain from StatsBomb, and various ‘non-shot expected goals’ models have been around for longer than that but I’m struggling to find links right now).
They’ve generally sought to measure how much value players further back in the goal-scoring process (ie, midfielders or defenders) add to the chances of goal-scoring. That makes sense, and I’ll preface all of my critiques about these models by saying that the people creating them are far smarter than me and they deserve immense credit for creating these models in the first place: progress is an iterative process, but it’s far easier to be the one coming in late in the day and saying ‘what about this?’ than it is to be the one taking the first step.
Across the board, these models have shown that passes in midfield have little (if any, really) direct value in scoring goals*. I have some quibbles about the fact that midfield is very dynamic and questions about whether the models capture this, but that concern is most relevant when applying the results of the model; however in terms of the descriptive value of the model I think that it’s valuable, and true-to-life, to say something like ‘by and large, taking an average of events in this area, actions in midfield have little direct impact on scoring goals’.
*In fact, the visualisations in Singh’s post show how quickly xT values rise as the ball gets closer to goal. It seems natural that players whose role dictates that they play close to these areas will be the ones get big ‘boosts’ from contributing to a team’s expected threat. The image on the left is a conventional heatmap of xT values in each zone; the image on the right is a 3D version.
With that in mind, we should think again about what we actually want to use these models for. Do we really want to use them to build a list of midfielders who contribute the most to goalscoring? Or, to put a different spin on that question, is that even the job of most central midfielders?
If the roles of attacking midfielders vary then the roles of central midfielders vary even more. And I’m not even convinced that ‘contribute heavily to increasing chances of the team scoring’ is part of most central midfielders’ roles.
On the offensive side, maybe this is something we need to think more fundamentally about. Is contributing to chances of scoring one continual phase, which runs from the moment that a team gets possession of the ball, or are there multiple phases to it?
I’ve thought for a while that there’s a kind of ‘endgame’ to football possessions. It’s the part where the team is solidly in the final third — or even the final quarter of the pitch — and need to find a way to get through, find a chink in the armour, find a way to manipulate the opponent to create space. (This doesn’t need to be for a shot directly, although it can be, it could be working to make space on the wing that leads to combination play that leads to a cut-back that leads to a shot. But the working of space must start somewhere). If this is a separate phase in itself, then it has implications for how we think about the game.
Is it the job of traditional central midfielders to help their team to this stage, rather than to be responsible for more directly increasing chances of scoring? (Part of their job, of course, will be defensive too, but that’s a whole different subject).
There’s two questions this post raises then. 1) What is the role of central midfielders in contributing to their team’s possession phases? 2) How much of players’ possession-based metrics is role-dependent and how much is indicative of ability.
Characteristically, I don’t have any firm answers (although I think I’ve dangled a couple of theories in this piece already), but I feel like they’re worthwhile questions to be asking.
Enjoy this? Subscribe to the newsletter:
Already subscribed? Spread the word: