Chaos vs Control

Using machine learning methods to identify team styles.

This season, there has been inklings of a shift away from the current footballing meta. That is the conservative, positional approach popularised by Pep Guardiola which puts emphasis on strangling the life out of the game through ball and territorial dominance and an obsession with preventing opposition counter attacks. This is achieved via a structured attacking shape and rest-defence that’s usually composed of 5 defensively minded players. Teams such as Iraola’s Bournemouth are the poster boys of the new “counter-culture”: direct, aggressive, and high-risk football, and they have the results to validate it - the Cherries have impressed this year and are in contention for a European spot. Even Guardiola himself has acknowledged this new trend in tactics, labelling Bournemouth’s style as ‘modern football’.

This notion isn’t exactly new, Guardiola and Klopp used to represent opposite ends of the spectrum throughout the 2010s. Klopp’s Dortmund and early Liverpool sides were fast and aggressive, characterised by their ‘gegenpress’ and Guardiola’s Bayern and Manchester City teams represented possessional control. This dynamic was illustrated perfectly in a Tifo YouTube video that is well worth your time.

Using the Klopp-Guardiola dynamic as an example, one could argue that all roads lead to control. Klopp’s Liverpool team gradually evolved into something that resembled a positional team, they placed far greater emphasis on controlling matches and picked up lots of trophies in the process. This was seen as a natural evolution, dictating the tempo and reducing variance lends itself to teams with a talent advantage and, as Guardiola has repeatedly shown, is a good blueprint to win a league title. Therefore, once Liverpool’s squad became one of the best in the world, it made more sense to embrace a more measured approach. Of course, Liverpool still kept a lot of their ‘chaotic’ roots, pressing much more aggressively than Guardiola sides and still having a higher propensity for risk and verticality, but there is no denying that the difference between Klopp’s early years at Anfield versus the latter ones was stark.

All this is to say is that a team-style focussed on control has been the meta in recent times, reducing variance in games suits the bigger teams and is a sensible approach considering league titles are a marathon and not a sprint. This begs the question as to whether “chaos-ball” has a ceiling. Teams embracing this style usually get most out of it when they are underdogs, using their style to cause disruption and close the talent gap.

What actually is chaos-ball?

At this point, I think we are all quite familiar with the “control” style of football: structured positional play, slow and intricate build-up, lots of width but a focus on using the centre of the park, and an aim to work it into the box.

The makings of the “chaos” style is less established but I would suggest it is characterised by aggressive pressing, direct and risky passing, overloads in wide areas, and a lot of physicality. Bournemouth are proponents of this style for a reason, they are incredibly direct and engage in high, man-to-man pressing with an aim to deny their opponents the opportunity to gain control of a match by inducing chaos. Another aspect of their style of play is “attacking depth quickly and frequently”, players often make direct runs in behind (read this article from spielverlagerung for a great write up on Bournemouth’s tactics) and they also have no qualms about overloading the box.

Another key difference in their style of play compared to a team like City is their risk appetite in wide situations. Not only do they encourage bursting overlapping runs from the likes of standout left-back Milos Kerkez, but they also are encouraged to take on defenders more often. In an article with the Independent, Andoni Iraola highlights this point himself, stating that he wants his players to “risk a little bit in the dribbles” and that if they “don't have a teammate ahead, forget about the pattern, just drive the ball and try to force things to happen. I want him to attack first”. It’s easy to contrast this to Guardiola’s City side, where Jack Grealish may be the perfect epitomization of the opposite. It’s easy for one to recall the Englishman encountering a one-on-one situation and deciding to pass the ball back to a supporting player, ensuring the continuation of safe possession rather than attacking the defender.

Off the ball, Bournemouth are even greater demonstrators of this style, thanks to the great data from markstats, we can see that the Cherries are league leaders in forcing the lowest opposition build-up pass completion rate (and just general pass completion rate too). They also lead the league in causing the opposition to lose the ball in dangerous situations near their own goal. This suggests that they are the best pressing side in the league, I personally think Arsenal may be superior, but Iraola’s men choose to engage in this approach more consistently and don’t really appear to have an “off button”, unlike the Gunners who are also happy to sit in a deeper block.

Suffice to say, I think that “chaos-ball” can be summarised by the word aggressiveness and the difference between chaos and control comes down to risk-appetite in your on and off ball approach.

How to categorise the two styles?

In this post, I want to categorise team styles with regards to the chaos vs control dichotomy. I aim to use machine learning methods to do this, and that involves collecting a bunch of team statistics for the season so far and then seeing how they fit into my notion of the two styles.

I would argue that there are two axes:

• Risk-appetite/aggressiveness - teams that like to be direct, vertical, and in your face versus teams that like to be cautious and risk-averse

• Possession quality - teams who shy away from the ball and when they do get it they don’t make good use of it, versus teams that are progressive and consistently make good use of the ball in dangerous areas

If we had four quadrants with these axis, here are some examples I expect to see:

• High possession quality and highly aggressive: Barcelona and Bournemouth

• High possession quality and highly conservative: Manchester City and Juventus

• Low possession quality and highly aggressive: Getafe and Brentford

• Low possession quality and highly conservative: Nottingham Forest and Ipswich Town

Now, it’s worth mentioning that these axes do not represent team quality (although, there will be some quality effects that bleed into it: i.e. teams that are good are usually better in possession and vice versa). For example, Rafa Benitez sides from the past would likely fall into the low possession quality and highly conservative camp despite being very successful. Whilst, Ten Hag’s United teams would most likely be high possession quality and highly aggressive quadrant despite being pretty awful overall.

I collected a database of team statistics for all the teams in the top-5 leagues so far this season and tried to narrow down the metrics that would best capture elements of the two axes above. I will now describe them and briefly justify their inclusion in the analysis.

• Possession & Passes completed - pretty straightforward, these stats show how much of the ball the team has in a match

• Progressive actions (carries/passes) - teams who engage in more progressive actions have better possession quality

• Pass completion % - this is a proxy for passing quality but also risk-appetite in passing

• Short passes & Long passes - these two metrics contrast each other, more shorter passes indicate a more possession-focussed style while more long passes indicate verticality

• Passes in Opp Final Third & % of passes in Opp Final Third - this shows how many passes a team is making in the opponent’s third, more suggests a team that is better on the ball and dominating territory. % of passes in final third can show a teams directness, if the total number of passes is low but this percentage is high then it can show a direct team who cares more about attacking the opponent than safe possession

• Touches in defensive third & attacking third - again this is showing where the team is using the ball most, in defensive areas or attacking ones

• Take-ons attempted - more take-ons suggest a team is attacking more

• Total aerial duels - more aerial duels suggests more direct aerial balls, something “chaos-ball” teams tend to engage in more often

• PPDA & OPPDA - passes per defensive action represents attacking pressing quality and also how aggressive opposition teams press against you

• xT & xTA - expected threat for and against, teams who dominate dangerous territory will show higher

• Pass completion rate in build-up & Opposition pass completion rate in build-up - how well a team passes when building attacks and how well their opposition does

• Defensive line height & Opposition defensive line height - how high up the pitch a team defends and how high up their opponents defend on average

• Dangerous losses & Opposition dangerous losses - how often a team loses the ball close to their own goal, and how often their opponents do. Represents a teams ability in deep build-up and managing a high-press and also their ability to cause problems in their own high-press

I may have missed some obvious ones here, so feel free to let me know in the comments any thoughts you might have! I collected this data from Fbref, Understat, and Markstats.

Method

Using these metrics, I decided to use the statistical method of Principal Component Analysis (PCA) in an attempt to simplify the the style comparison. If you are unfamiliar with this concept, I would recommend reading up on it here rather than me trying to poorly explain it in more detail. But the gist of it is that PCA takes all our different metrics and finds the best ways to combine them into n measures (in this case, two). Instead of looking at many separate metrics, PCA creates new combined metrics that capture as much information as possible. Think of it like taking a complex set of ingredients and distilling them into their two most essential flavors. The first combined measure (or principal component) captures the strongest pattern in the data, and the second captures the next strongest pattern that's different from the first. This gives us two meaningful axes to plot our data on, making it easier to see patterns and relationships. This statistical learning is unsupervised meaning I cannot dictate that I want the two components to represent aggressiveness and possession quality but the hope is that this will emerge from the data naturally.

Once the PCAs have been set up, we can analyze the Principal Component Loadings (PCL), which show how each original metric contributes to each component. For example, if possession has a loading of 0.2 for PC1 and -0.2 for PC2, this means possession strongly correlates with PC1 in a positive direction, while having a small negative correlation with PC2. Teams with high scores on PC1 would tend to have higher possession values, while teams with high scores on PC2 would tend to have slightly lower possession values.

Let's examine the metrics that show interesting relationships in our Principal Component Loadings (PCL):

Possession Quality (PC1)

Teams scoring high on PC1 generally show:

• Higher possession (0.26)

• More progressive actions (0.25)

• Better build-up passing (0.21)

• Face less aggressive pressing - their opponents have higher PPDA (0.21)

• Strong pressing - lower PPDA (-0.18)

• Fewer aerial duels (-0.13)

These loadings suggest PC1 represents a possession-oriented, methodical build-up style with technical competence.

Aggressiveness (PC2)

Teams scoring high on PC2 typically display:

• More passes in the opposition third (0.35)

• More dangerous opposition losses (0.36)

• More long passes (0.24)

• More aerial duels (0.21)

• Lower build-up pass completion (-0.24)

• Strong pressing - lower PPDA (-0.17)

• Face more aggressive pressing from opponents (-0.12)

These loadings suggest PC2 represents a more direct and aggressive style with more long, aerial balls and lots of high pressing.

As we can see from above, the PCA has reduced the dimensions of our data into two distinct axes that I think represent my hypothesis of the two competing styles well.

Results

Now that the PCs are in place, we can plot the results on a scatter chart. Below displays PC1 on the x-axis, meaning teams further on the right of the chart are better in possession whilst teams on the left are worse on the ball. PC2 is on the y-axis, meaning that teams higher up on the chart are more aggressive whilst teams lower down are more conservative.

I have also used the K-means clustering algorithm to group these teams into similar playing styles. This technique partitions teams into clusters where each team belongs to the cluster with the nearest mean. After testing different numbers of clusters, I settled on 5 clusters as the optimal balance between minimizing inertia (the sum of squared distances to cluster centers) and creating meaningfully distinct team groupings, Wikipedia does a good job of explaining this method in more detail. The clusters are represented by the colouring of the data-points on the chart.

We can now look at each cluster in more detail:

• The purple cluster - this cluster is made up of the elite teams in Europe, and Marseille. It is characterised by having very high possession quality. Bayern, PSG, and Barcelona stand out most clearly, and it’s interesting to note that these are teams who are dominating their domestic leagues. The other variation in this cluster is the difference in aggressiveness, we can see Real Madrid, Marseille, and Man City showing more towards the conservative side whilst Arsenal, Barcelona, and Bayern are more aggressive. Hansi Flick has implemented a high octane style at the Camp Nou this year, which supports their position in the chart

• The green cluster - this cluster has teams that are still pretty good on the ball, but not great. There are also more aggressive sides in this cluster, Bournemouth are the most aggressive (and the second most in all of Europe) which certainly rings true. The two Basque clubs, Monaco and Bologna join them in this more aggressive end. The rest of the teams are bunched close to the centre of the chart, I would argue that these teams don’t have a distinct style of play. Juventus are the most conservative team which again sounds about right, interestingly Tottenham are within the more conservative quadrant which is one of the biggest surprises for me.

• The orange cluster - these teams are middling to pretty bad on the ball, and all are more conservative. Strasbourg are the most conservative team in the Europe, followed by Southampton - this makes sense to me seeing as both teams have been managed by control-focussed coaches in Liam Rosenior and Russell Martin. Atletico Madrid, Aston Villa, and Manchester United are in there too but are above average in possession quality. Newcastle actually reside in the bottom-left quadrant with all of the soon to be relegated Premier League sides. This might sound surprising but when you consider Newcastle are actually quite conservative in their approach, I think it makes sense. In addition I think it’s nice that they belong in the same cluster as Atletico considering that Howe was reported to spend a lot of time at their training ground in his sabbatical before taking the Toon job.

• The blue cluster - these are the teams bad in possession and more conservative. Not much to say here other than that there is only one Premier League side in Ipswich Town, perhaps this is a byproduct of the Premier League becoming so strong that the latter ends of the other top European leagues are considerably worse than the chaff in England.

• The red cluster - this cluster represents the aggressive teams that are not very good in possession. The biggest outlier in the whole sample is Bordalas’ Getafe, known as an incredibly direct and aggressive team - in my eyes, this validates the entire approach. We also see Nottingham Forest, Everton, and Crystal Palace here.

---

Overall, I think that this has been an interesting exercise, looking at the growing trend towards fast and aggressive football over the slow and controlled approach. It does feel, to me at least, that we might be approaching the end of a tactical cycle. With Pep’s City struggling and beginning to experiment with more direct football, the soon-to-be English champions, Liverpool, enacting a far more balanced approach and embracing verticality, the rise of Bournemouth and even the dominance of Barcelona who are way more “chaos-inclined” than previous years.

It’s worth considering if there is a ceiling to “chaos-ball”, as discussed before, can a top team really maintain that high intensity all season and win a league title? Controlling matches clearly has its benefits, not only in preserving player fitness but also it’s a tried and tested method to effectively defend with the ball. I wonder if Liverpool have demonstrated a sensible balance this season and whether we might begin to see other top teams twist the dial more towards the centre.

One thing is for sure, the game would do better to embrace chaos more often.