What makes pressing succeed?
SPORTS ANALYTICS · FOOTBALL · MACHINE LEARNING · TRACKING DATA
Part 2 of a two-part investigation. Part 1 showed that 94% of European teams press more aggressively at home than away, a pattern consistent with the “Bourgeois” strategy from evolutionary game theory. Here we ask a different question: does pressing work better at home, and what actually makes a pressing attempt succeed?
Introduction
Part 1 measured pressing aggressiveness as the frequency of pressing attempts in the opponent’s half. But pressing aggressiveness has another dimension: the style of the press. Do home teams close in faster on the ball holder? Do they outnumber the holding team near the ball more often? And do these stylistic choices actually predict success?
To answer this we needed tracking data (recordings of every player’s exact position on the pitch many times per second) rather than discrete match events. We used the SkillCorner open dataset: broadcast tracking data for 20 Australian A-League matches from the 2022/23 season, yielding 1,410 labelled pressing chains. Each chain ends in a binary outcome: the pressing team wins the ball back (through a clean recovery or by forcing a turnover), or the holding team retains possession.
Data exploration and visualization
Before building any features, we generated frame-by-frame animations of individual pressing sequences from the raw tracking coordinates. Each animation shows the positions of all outfield players, their velocity arrows, and the ball as the pressing sequence unfolds in real time. Watching those clips, three patterns stood out, each suggesting a feature worth measuring. First, it looked as though the speed with which the pressing team closes in on the ball holder increased the success probability of a pressing attempt. Second, when the pressing team had more players near the ball than the holding team, the pressing seemed to succeed more often. Finally, pressing close to the touchline or goal line gives fewer degrees of freedom to the ball holder and should increase the recovery probability. Our intuition is partially confirmed by a simple t-test: the effect of closing speed and pitch zone is significant, though numerical superiority does not reach significance on its own.
Example 3 in the video also teaches us something important: even when the pressing effort is well coordinated, aggressive and sustained, the holding team may keep control of the ball thanks to the technical skills of its players. In other words, the success of a pressing attempt does not depend only on the pressing team but also on the holding team. This is beyond the scope of our models, but worth keeping in mind when interpreting the results.
Logistic regression
Based on the above exploration, we designed features that we thought had an effect on recovery probability at the end of a pressing chain. Two of the dimensions we hand-crafted captured pressing aggressiveness (closing speed of the nearest pressers, numerical superiority of the pressing team near the ball holder) while others were contextual (pitch zone, home/away status). A logistic regression on these features serves as an interpretable baseline.
After training, the logistic regression shows that numerical superiority and closing speed have an effect on pressing success. Having more players around the ball than the holding team does more to tip the outcome than arriving faster, though both dimensions raise the probability of a regain.
The home/away coefficient is positive and significant: home teams recover the ball more often (42.8% vs 37.4%, p = 0.046). But here is the unexpected result. When we compare home and away teams on the features that predict success (closing speed and numerical superiority), we find no significant difference. Home teams do not press in a stylistically more aggressive way. Yet their pressing is more effective.
Deep learning
The logistic regression can only see what its hand-crafted features encode. A natural question is whether there are geometric properties of pressing organisation not captured by those summaries.
We tested this with DeepSets, a permutation-invariant deep learning architecture that takes the raw spatial configuration of all outfield players as input and learns its own representation from scratch. The permutation-invariance is important: the model’s prediction is identical regardless of which player is listed first, which is the right inductive bias when there is no natural player ordering. We also tested a temporal extension using five keyframes per pressing chain.
The snapshot DeepSets model reached an AUC of approximately 0.65, comparable to the logistic regression. The temporal extension underperformed (not shown here), likely because 1,410 pressing chains are too few for it to learn that early keyframes add little predictive signal beyond the end-of-chain snapshot. The conclusion is that our hand-crafted features capture the geometrically relevant information relatively well; there is no large residual signal left in the raw spatial data.
Wrap up
Combined with Part 1, the home advantage in pressing now operates along two independently documented dimensions: home teams press more often and more effectively. The effectiveness gap is not explained by observable differences in pressing style or spatial organisation, which suggests the mechanism runs through channels not visible in the tracking data: possibly psychological effects via crowd influence on the away team’s composure, or aspects of home-team collective movement that the current features do not capture.
The Bourgeois pattern from evolutionary game theory runs deeper than Part 1 suggested. It is not just that home teams are more willing to contest the ball; their contests are more likely to succeed, and we have yet to find what explains this difference.
Part 1: Territorial behavior in football: the Bourgeois strategy in European leagues