Possession is not nine-tenths

There are several motivations for this post, but one is to add to the public analysis conversation and show that (hopefully) interesting work doesn’t need to be complex. Small children often ask more interesting questions than adults.

This post is being written just after Christmas, a time of the year when the fridge is full of odds and ends that are all a little imperfect. Everyone will have their own favourite ‘leftover’ recipes, ways to squeeze the most out of the ingredients available.

Football data has an equivalent: defence. The full-bodied enormity of soccer’s contests for possession are carved and trimmed into duels, interceptions, clearances. Unlike expected goals - akin to a centre-piece cut of prime meat - they require ingenuity to work with.

Ideas about adjusting defensive stats began. Ingenious. But not all ingenuity is genius.

It’s now about ten years since I first ventured into ‘possession adjusting’ as a way of tinkering with defensive stats. (Others arrived at the idea independently before and since). I’m now pretty unsure about it.

This is partly a post about some data analysis work, but if you poke it hard enough it’s about the essence of the sport. Can you tweak an ingredient enough to make it a centre-piece on its own or are there insurmountable limits? What is and isn’t separable in football?

And how do you recognise ingenuity?

Data tables are in the eye of the beholder

Code for the project can be found here: it goes without saying that I would appreciate mistakes to be ruthlessly hunted down.

The theory behind ‘possession adjusting’ is simple. A team having possession of the ball (or not) affects what their players can do. So you adjust a player’s stats based on their team’s share of possession. In one sense, it’s the same theory as averaging a player’s stats per 90 minutes - that things need to be adjusted to make them justly comparable.

However, an update to work I did a few years ago has given me the same result as previously, one that doesn’t mesh well with possession adjusting.

We’ll see the data below, but there’s a lack of a clear link between possession share and player defensive actions output. This is not what the theory behind possession adjusting would expect.

Below are three tables that you can glance at, one for each of three positions: centre-back, full-back, and defensive midfielder. They’ll show three things, which were also features of the previous work:

1. The link between possession share and defensive actions is small if existent at all

2. The link varies by position

3. The link varies by defensive action type

The tables are ordered by the absolute strength of the correlation for each position. In none of them does the strongest correlation hit +-0.2. (A little note about the data follows them, as a reward for getting through the numbers).

Centre-backs

Full-back

Defensive Midfielders

The data being used here is the Statsbomb open dataset for the 2015/16 seasons across the men’s ‘Big Five’ European leagues. (that is, the English Premier League, French Ligue 1, German Bundesliga, Italian Serie A, and Spanish La Liga). It’s accessed via the kloppy Python package. (I’m slightly wary of discrepancies this might cause compared to working directly with the event data, but I don’t think the trends would change).

The defensive action types are ones found in Statsbomb’s dataset, although the definitions are relatively common to event data. The player positions are also from the Statsbomb data, and only feature players who played 450+ minutes in those positions - the code for this project can be found here and a full table of results here.

In fairness to the concept of possession adjusting, those three tables aren’t the entire story.

There are higher correlation strengths among other positions, although none greater than +-0.35 in Attacking Midfielders, Center Forwards, or Wide Midfielders (and most not greater than +-0.25). The Winger position gives the greatest support for the possession adjusting principle (Wing Back correlations are stronger, but a tiny group):

Wingers

There is more that can (and will, soon) be said about the gap between these figures and those for the other, more defensive, positions, but one thing is clear. This marks a blow against the principle of a uniform possession adjustment of defensive statistics.

More than meets the eye

There are a lot of interesting little avenues that can be found in this data. It seems curious, for example, that positions known most for their defending - central defenders, full-backs, and defensive midfielders - appear to be affected least by their team’s share of possession.

Defending is a team game though. It’s not unusual for ‘forwards’ on weaker teams to be more concerned by defending than attacking, while their counterparts on stronger teams are sometimes given a bit of a pass from defensive duties. Although I suspect that the Winger position correlations might look different if some lower possession 4-3-3s were classed as 4-5-1s, it seems true to my ‘viewing experience’ for possession share to have a more discernable effect in ‘forwards’ than ‘defenders’.

(An aside, although it’s covered in the code repo for this project - as Statsbomb’s event data has possession sequence IDs included with it, it’s fairly straightforward to identify when players were on-pitch and use the sequence IDs as a quick reference for events which happened while they were on-field).

This is where the fun begins

If you’re only interested in the data tables, you can safely close the email now. If you’re interested in the fine line between wonderment and futility in your numerical analysis, continue.

Let’s duck back to my theory that correlation strengths for Wingers are partly affected by the formation assigned by Statsbomb. My (unverified) assumption is that in a Statsbomb 4-3-3 formation, the wide attackers are designated as Wingers, and in some other formations - like a 4-5-1 - they’re designated as Wide Midfielders (whose correlation strengths looked much more like the Center Forwards, with a mean average correlation of the six defensive actions of -0.166 and -0.186 respectively - again, full results can be found here).

Assigning a formation to a team is notoriously difficult. It’s a subject I’ve written about before, a subject that even data-reticent teams complain to data providers about, and, as a result, an area which use of tracking data has tried to solve.

Although football punditry commonly talks about a team’s ‘in-possession’ and ‘out of possession’ shape, my conclusion in the post linked above (which I still stand by) is that the ‘formation’ a team is assigned by fans, media, and data providers is the one which they ‘see’ most often. It’s a shorthand for the roles that the players within it are playing.

When you come to the division between what the shorthand ‘4-3-3’ indicates and what the shorthand ‘4-5-1’ indicates, the difference is mainly in the defensive roles of the wide attack/midfield players.

If the switch between assigning the two formations was systematically done based on the ‘defensiveness’ of a team (for which possession share might be a plausible proxy), then this would just be a quirk of data. However, if teams are sometimes but not systematically assigned to ‘4-5-1’ or ‘4-3-3’ based on their possession share, then that might be a confounding factor in this particular type of data analysis.

(In fact, in the dataset, Wide Midfielders and Wingers in this study had the lowest (44.9%) and highest (53%) median possession shares respectively).

Tying the threads of the Get Goalside cinematic universe

As I said before, the concept of possession adjustment is sort of similar to the ubiquitous concept of ‘per 90’-ing statistics. We recognise that there are ways that data isn’t ‘comparable’, and we think we can address that.

No method is perfect though (even ‘per 90’-ing gets questioned).

Possession adjusting supposed that a team having the ball affected the defensive output of players. (In reality, many defensive actions are ends of possession sequences, and so wouldn’t be affected by the duration between these sequence endings).

You could head in the direction of a different sort of possession adjusting. Instead of adjusting based on the possession share, you could adjust based on the count of possessions/possession sequences. Given that defensive actions often end sequences, you might reasonably expect that players involved in ‘pinball’ matches will have higher figures than those without. (In fact, there was a slight indication of this in my earlier work, though I haven’t tried replicating that with this 2015/16 dataset yet).

Unfortunately, football throws another spanner in the stew. There is an obvious stylistic difference between teams who play ‘high turnover’ matches and those that don’t, and there is likely to be a quality difference between these teams too. Some previous Get Goalside work with Statsbomb open data has shown a link between ‘ball-in-play time’ and breaks in play, and a suggestion that high-possession (Guardiola-type) teams have more in-play time because of this link with breaks in play. The relative quality of teams (and possibly the absolute quality) appears to have an impact on ‘style’ whichever way you slice things.

I suspect that this is inevitable. I’ve previously tried to distil football down into teams competing over space control and ball control, and everything else springing from there. If a team has the quality to do so, it will naturally want to have more control of the ball, the only method of scoring. The hugely imbalanced value of space on a football pitch - extraordinarily heavily concentrated around either goal - steers the (weaker) defensive team’s strategy. With such a dynamically tilted field, is full comparability through adjustment of statistics possible?

So what now?

According to my own memory of my views (a trustworthy record if ever there was one), I have two longstanding opinions about data.

One is that a large part of the craft of data analysis is tuning an internal ‘fuzz-meter’, the sense of when things are ‘close enough’, ‘probably legit’, etc. I am wary of false precision.

The second - perhaps paradoxically - is that measuring discrete, specific skills is very valuable. It’s just that most of event data analysis doesn’t do that.

Take a statistic that is pretty clearly useful, expected goals. Reliably getting a large amount of xG isn’t ‘a skill’ as such; it’s an indication, or a symptom, of the various skills that a top striker has.

To go back to the (far worthier) subject of defending, there are a variety of skills that you’re looking for in a defensive player. It turns out that these skills aren’t necessarily observable purely through stats like tackles and interceptions. And it turns out that they’re probably not observable through possession-adjusted versions of these stats either.

Other types of data might let you measure those specific skills (whether that’s tracking data or other types of event data).

There are some in this world who just want to create single number models.

Maybe you can create steak out of scraps.

That would be ingenious.