What are we doing with defensive stats?
After embarking on my journey into the football analytics world, I soon came across the problem of evaluating defensive statistics. Data on pressures, tackles and aerial duels are often reported in articles, analyses, and charts, especially when it comes to defenders. But if you have ever seen this data shown, even in a simple scatter plot with successful tackles and pressures, I am sure you were, at least initially, puzzled or uncomfortable with that chart.
That’s pretty normal because defensive stats are a bit different from the ones we are used to looking at, whether it’s expected goals, successful dribbling, or key passes.
First, it’s important to point out one fact: in my opinion, it’s not possible to accurately assess defensive skills through event data, especially for defenders. Defending well, especially for center backs, it’s mostly an off-the-ball skill that consists of many aspects: positioning, concentration, marking, space awareness, the ability to maintain physical contact with the opponent’s body to limit him without committing fouls, and many other micro-skills that are nevertheless crucial. These aspects are virtually impossible to measure through traditional event data. Let’s take, as an example, Giorgio Chiellini. In the past year, from EURO2020 onward, despite often precarious physical conditions, the Italian center-back has often given the feeling that he can still be one of the most solid defenders in the world, facing and limiting great strikers such as Lukaku, Kane, Dzeko, Ibrahimovic, and Immobile. Using the simple defensive stats produced by event data, I do not believe there is any kind of number adjustment that can effectively do justice to Chiellini’s last season.
This is not to say that in the future this will not be possible, and indeed I believe that the most advanced analytics departments of the clubs are already able, using tracking and other sophisticated types of data, to do much more than what we can do with tackles won, the success rate of aerial duels and other event data. But this doesn’t mean that this kind of data is useless and is no reason not to try to figure out how best to use it.
So, what’s the point then?
If you’re reading this you probably already know what possession-adjusted statistics are. To put it simply, it’s a kind of adjustment usually applied to defensive stats based on a clear assumption: the more you keep the ball, the fewer chances you will have to make defensive actions. That’s what I was mentioning before: if you have ever seen a chart with some non-possession adjusted defensive metrics, you may have been initially startled to see among the players with more successful actions those from non-dominant teams and conversely those from top teams with “worse” numbers.
Adjusting defensive metrics by possession, e.g. by a certain number of opponent touches or average ball possession percentage, is a mechanism that has been known for several years and is widely used in the football analytics community. Despite this, I’ve often wondered whether I was giving the right interpretation to this kind of data. In the rest of this story, I’ll try to do some reasoning, using data and some simple charts, about the use of possession-adjusted metrics and the general interpretation of defensive stats. To restrict the search space and simplify the task I will limit the analysis to center backs because I find that they are the most difficult players to evaluate with defensive metrics but are probably the ones for which it is most interesting to do it.
The dataset I’ve used is built in this way: I considered the FBref data from the Top5 leagues for the 2021/22 season, and I took the players that are considered as “Centre-back” according to the Transfermarkt main position. I kept only the players with more than 1000 minutes played in the league matches, for 313 players in total in the dataset. After that, I applied a per 90 minutes normalization and a possession adjustment (for 100 opponent touches) to the main defensive metrics available on FBref: tackles, dribble contested, pressures, and so on. I excluded aerial duel because it’s a kind of action that could be both performed when attacking and defending and using this data source it’s impossible to distinguish between the two situations.
First, is a macro example of what happens when applying possession adjustment to defensive metrics. In the two scatter plots below total tackles and pressures aggregated by teams are plotted.
The first one shows the raw values, while the second has the possession adjustment. We can see that applying the adjustment emphasizes teams that in the first chart had lower numbers because of their being dominant in ball possession such as Barcelona, Bayern, PSG, and Liverpool and penalizes teams with low possession like Osasuna, Brest, and Lorient.
This also shows us that there may be non-negligible differences in the number of defensive actions attempted by each team even when normalizing by the average possession or the opponent touches and that maybe we should be aware of that variance between teams when observing padj data at the single-player level. Another aspect to keep in mind is that not all teams make defensive actions in the same areas of the pitch or give the same tasks and directives to centre-back.
In this chart, for each team, the share of tackles and pressures made by center backs compared to the team total are plotted. That shows one of the main caveats behind some of the criticisms that are made of possession adjustment: when using adjusted metrics to evaluate players, the differences between the teams and tactical contexts (i.e. how many defensive actions are attempted and by who) are not considered.
Now, let’s move to single players' data. Here’s the chart with tackles and tackles won before and after the possession adjustment.
But still, when looking at a chart of this type or a percentile rank of a padj metric, I feel pretty uncomfortable in seeing the numbers of, let’s say, Djidji compared to those of Skriniar or Chiellini (the reference to the former is for illustrative purposes only).
As mentioned before, event data are nowhere near sufficient to cover the evaluation of a player’s defensive capabilities, and no one claims to be able to rely solely on these metrics to evaluate a centre-back, but there is still some interpretation issue. In addition, there remains a correlation between successful defensive actions adjusted for possession and the number of those attempted. And the same issues apply to other metrics, such as pressures.
This led me to wonder if I am giving the right interpretation of defensive data, as well as many others in the football analytics community, even when adjusted for possession. I have seen many analysts and content creators consider statistics such as possession-adjusted tackles and pressures as indicators of players’ ability to perform these actions and “defend well.” I, too, have often given this interpretation, because I believe that when looking at statistics and data our biases always lead us to think that “more is better.” This is roughly true in many cases: expected goals, expected assists, key passes, assists, dribbling, progressive passes, and so on, but I’m not sure the same rationale applies to defensive statistics. I asked myself this question: is a good defender simply one who does many things well or one who makes few mistakes? And in the second case, how can I tell? These are questions that are complicated to answer, but they also allowed me to give myself some answers. When we look at defensive metrics for evaluating a defender, we must always be aware of what we want to know, and that trying to figure out how a player defends is a different thing from quantifying how well he does it. And when it comes to defining a player’s defensive style, padj metrics can be very useful to distinguish, for example, a more aggressive defender from a more wait-and-see one. And in that case, attempted actions may be more meaningful that the succeded ones.
Sometimes, especially in modern football, one is led to believe that a center-back who defends aggressively is necessarily better than one who is more cautious, but it is true that one can do poorly in an aggressive defensive phase just as it is true that one can defend well in a conservative manner. But deciding to consider possession-adjusted metrics more as an indicator of style than one of quality doesn’t solve the issue of assessing the level of defensive performance of centre-backs and other players.
Adding more context to the data
Other approximations have been proposed over time to improve the interpretation of defensive metrics. Recently I have read some proposals to take into account the playing style of teams in adjusting defensive metrics. To balance the defensive numbers of players of particularly aggressive or wait-and-see teams, regardless of just the time spent in possession of the ball, these techniques further adjust individual player data based on the variance between the number of padj actions (e.g., tackles or pressures) attempted by the team compared to the average of teams in the league or multiple leagues. I have also tried to implement some of these approaches, but I still found two important critical issues.
First, this can be a further refinement to weight the influence of team playstyle when evaluating single player numbers but still in terms of style and attitude, and there’s the same problem as with padj stats: we are not saying anything more on the quality of the defensive action and their “precision”. Secondarily, adding another level of adjustment on top of the possession one further increases the level of complexity and reduces the intelligibility of the data. When communicating with less experienced people, it is sometimes already complicated to explain what 0.9 tackles won per 100 opponent touches means compared to the more simple and intuitive “per 90 minutes.” It would be really difficult, especially in mainstream debate, to convey that the 3.2 pressures carried by a player are normalized by possession and by the variance between the average number of pressures applied by the team and the league average. Definitely a situation I would not want to be in when replying to the comments below a Twitter thread of mine.
What about the good old success rates?
When analyzing event data, percentages are often not given much consideration. We know that the percentage of saves of a goalkeeper is quite meaningless because it doesn’t take into account the difficulty of the shots faced, or that the percentage of shots on target isn’t necessarily a good proxy for the finishing abilities of a striker. The problem is that those rates tell us nothing about the number of attempts, which is crucial: when comparing two wingers you will probably prefer the one that completes three dribbling per 90' with a 75% of success over another with 90% but that tries a single dribbling every two matches. With defensive metrics, there’s quite the same problem. When comparing two players, if I say that the first has 80% success on tackles and the second 70%, I am not giving any information about the number of attempts. But as mentioned before, when considering defensive metrics, as opposed to other statistics, it becomes more interesting to have fewer failed actions, and thus the success rate may gain some relevance. Ideally, it would always be better to be able to contextualize the success rate of defensive actions by also considering the number of attempts. Most people like to look at a single number to get a piece of information, but sometimes it is better to see two if that is the price to pay for a more complete understanding.
The three examples below show, respectively, the top quartile central defenders by possession-adjusted tackles won, by success rate in tackles, and the intersection of the two groups.
I’ve asked myself which of the two quartiles contained the “best” defenders, but I did not come up with an answer, and probably there isn’t a correct one for a question asked wrongly. Certainly, the group of the top players for tackles won encompasses centre-backs that I would call “most aggressive.” The intersection between the two groups is a fairly narrow club that contains a few of the centre-back considered among the best in the world and many others much less esteemed.
Of course, these arguments can be extended to all players, even those in other roles, and to other defensive metrics, such as pressures. In this case, I decided to focus on tackles, although I am not a big fan of the classification of defensive actions made on FBref, which I believe tends to underestimate the number of tackles because of the presence of contested dribbles and “recovered balls”.
There are no easy answers to complex questions
As I said at the beginning event data and current defensive metrics are not enough to assess the defensive skills of a player, and I believe that often we are not interpreting that data in the correct way. Possession-adjustment it’s a great approximation based on a solid and logical principle, and it’s fine to use that. What I will do during this season is to start to consider possession-adjusted metrics as a measure of defensive style (probably using attempted actions in place of successful ones) and try to find a more appropriate way to quantify the quality of performance. To begin with, I will try to use success rates more than in the past, always trying to put it in context by taking into account the number of attempts or successful actions.
I also hope that this story will be a starting point for me to interact with other community members and improve my understanding and interpretation of defensive metrics.
So, for this time that’s all folks! Ping me on Twitter for every feedback and above all for constructive criticism!