Diving Deep: Possession adjusting through Convex Hulls

Today’s post will be quick and to the point due to some serious and unfortunate reasons. I hope to have the mental space and time to finish the project I’ve been working on by next week.

Recently, you’ve seen me discussing my struggle to normalize defensive actions. This was sparked after reading a post by Get Goalside, where he explored the correlation between defensive actions and possession, broadly speaking.

Inspired by his approach, I decided to replicate his research using a slightly different basis. Instead of looking at overall possession, I focused on the relationship between defensive actions and the opponent's possession (passes, receptions, carries, take ons, shots, bad touches, crosses) inside the convex hulls of players when they were on the pitch, while revisiting this work to refine and improve it.

For this, I used two types of convex hulls:

1. All actions performed by the players in a given match.

2. Actions limited to the first two-thirds of the pitch.

The results include a lot of visualizations — so feel free to skip if that’s not your thing. For those who stick around, I hope you find it insightful!

Let’s start with the overall convex hulls. The number of defensive actions performed in a match doesn’t correlate that much with the number of actions performed in his convex hull.

This type of results holds true for all the positions with different degree.

Attacking Wingers and Midfielders:

Center and defensive midfielders:

Centerbacks:

Wingbacks/Fullbacks:

Strikers:

These results are even worse when examining the relationship between actions performed agains in the convex hull and the type of defensive action performed by the player of the convex hull, with no actions type over 0.18 of correlation.

If we focus solely on convex hulls based on the first two-thirds of the pitch, the results are slightly better. However, even with this adjustment, it doesn’t fully justify using possession to adjust defensive actions under these settings.

A slight improvement overall on the 0.33 from the all pitch convex hulls:

No improvement from the 0.48 for attacking wingers and midfielders, which seems logical if we cut the main zone of their action:

And this hold true for strikers too:

There’s a slight improvement on the 0.35 of center and defensive midfielders:

And a slight improvement for wingbacks and fullbacks too (from 0.49):

And almost none for centerbacks:

If we go and watch the relationship between possession and specific type of defensive actions things don’t change a lot:

1. Recoveries are 0.19 vs 0.18

2. Tackles are 0.21 vs 0.17

3. Clearances are 0.15 vs 0.14

4. Fouls are 0.09 vs 0.11

Possession adjusting still doesn’t seem particularly effective when analyzed against event data. It’s easy to imagine that tracking data could offer better insights, helping to pinpoint the exact moments when a defender has the opportunity to perform defensive actions, thereby normalizing for those opportunities.

For now, if we want to capture defensive activity, adjusting based on the number of minutes played remains the reliable approach. However, there might be room for improvement by considering adjustments on a team basis, comparing someway defensive actions against the number of teammates. At this point, I’m not entirely sure — but it’s something worth exploring someday.