Author: Mike Goodman

The next in our continuing series where Kirsten asks me all (or at least one or two of) the questions you could ever want answered about using stats. This time we're talking about analyzing teams.

KS: So, even after all this education, I’m not quite clear on which stats apply to individuals, and which to teams. Is expected goals the number one valuation used for both players and teams?

MG: There are a bunch of stats that might be applied to players and teams. Pretty much everything you can count a player doing you can then total up and look at as a team number. So numbers, of course, are more useful than others. And for starting a conversation about team analysis, xG is definitely the way to go. As a metric xG is actually probably more useful at the team level than the player level.

KS: In order for me not to panic about not knowing this teams v individuals numbers, Mike sent me a piece entitled (in hindsight, ironically), “Who's gonna fix Wolves?” It’s actually a great place to start thinking about team numbers, given that when the FA hit the pause button on the season, Wolves were in sixth—but when the article was written, they were second from bottom and had yet to record a win. Mike asks if it’s time for Wolves to panic, but unequivocally states that it’s not. Turns out he was right, so it’s appropriate to ask what numbers he used to predict the future . . . and possibly to ask what other factors influenced what he saw in the numbers. Obviously the first number he looks at is xG, but for a defensive team like Wolves, it might be more appropriate to ask about what they’ve conceded. What’s the stat for that, Mike?

MG: The stat for that is xG conceded, it’s just like xG except on the other side of the ball. Look at all the shots a team has conceded, total up the expected amount of goals those shots might lead to, and then look at the actual amount of goals they’ve given up. The great thing about xG as a metric at the team level is that it operates just like goals does. You look at a team’s xG and the xG they’ve conceded and then the difference between the two, so just like you’d look at a team’s goal difference you can also look at their xG difference. And the basic rule is that we should all expect teams to have their goals and xG converse to the same values. Or, to say it slightly more nerdily, xG and xG conceded predict future goals scored and conceded better than current goals scored and goals conceded does. In Wolves' case the side had conceded a bunch more goals, 11, than the xG of the shots they’d conceded, 6.49, predicted, so it was easy to predict that their defense would improve going forward.

KS: And you were right! However, when you go on to explain the bad news, you state that Wolves’ xG isn’t nearly what it was last season—enough to land them just outside the top of the table. Instead, they’re in the European places. Does this indicate that looking at just xG and xG conceded isn’t enough?

MG: Well, I think there are three components to that answer. First is that position in the table is always going to be contingent on not only how well one team plays, but how well everybody else plays. So, part of what was going on in this once and maybe future season is Arsenal struggling, Spurs collapsing, and just in general a season where the league’s big six are being underwhelming. So, even if xG is fully capturing the contours of Wolves performance, the side’s performance in relation to everybody else can certainly change.

Second, we need to separate out two different ideas surrounding xG. There’s the idea of how a team’s actual goals stack up against their expected goals. In that arena we can say with confidence what we expect to happen (the goals they score and concede will eventually come in line with xG), but the story of how that is likely to happen is where science meets art. The question of what exactly is causing the divergence is an interesting one and highly relevant for players and managers and fans, even if we can say that whatever it is is likely temporary.

And finally, xG levels can themselves change. While xG is a pretty good proxy for how good a team is, teams get better and worse all the time. So we might observe a team’s xG improve on either side of the ball from season to season, or even within a season, and then we’d want to look for reasons why that was happening. Usually xG is pretty predictive of itself, which is to say that usually teams don’t improve a ton or get dramatically worse over the course of a season, but there are always exceptions.

So, to sum it up. Yeah.looking at xG itself isn’t enough. You need to look at a team’s metrics in relation to the rest of the league, look at how a team’s actual results are in relation to that metric and why they might differ, and then look at the movement of the metric itself and what might be causing it to change.

KS: Now, despite being barricaded inside my apartment, I don’t have time to go back and look for an article in which a team’s expected xG differed wildly from the actual number of goals scored. Here, Wolves were pretty much on track: They’d scored 6 with an xG of 5.61. However, they had conceded 11 goals from 6.49 xG conceded—and this isn’t viewed as a problem. So given that defense is simply attack in reverse (shocking!) I can look at those defense numbers and wonder, what on earth causes this difference between what is predicted and what happens in reality?

MG: Right, early this season the biggest thing that was wrong with Wolves is that the team’s goals conceded was much higher than their xG conceded. And while we’d expect that the numbers would come back in line, fixing the problem, that doesn’t tell us much about how that’s going to happen. The biggest element involved in these kinds of divergence is generally just variance in finishing. There’s really not much you can do if your opponents keep launching into the top corner against you, except rely on the numbers to reassure you it won’t keep happening.

But, there are other things that can contribute that are at least worth looking into. We can isolate keeper performance, for example. StatsBomb has a separate model, a post-shot xG model, that looks at the performance of keepers given the shots they’ve faced (this is different from normal xG models because it takes factors like the trajectory of the ball into account, which you’ll just have to trust me is the best way to go about things because working through those differences is a whole article unto itself). Using those tools we can determine if some of the divergence is down to keeper performance. Or we can separate out set pieces from open play. If the problem is that a team is conceding a lot more on set pieces than expected, it would be worthwhile to examine if there is in fact something going wrong in that phase of the game.

All of which is to say that the fact that a team is diverging from their xG or xG conceded is the start of the story, the fact that they’ll eventually come back to expectations is the end, but there are chapters and chapters to investigate in the middle about why that divergence occurs and what the likely path back is.

KS: Ok, one last question because it’s been bugging me: why are penalty goals removed from the equation? 

MG: Penalties are just not particularly predictive of anything. Just because one team got a bunch of penalties doesn’t mean they’re any more likely to get them in the future. So, including them doesn’t help us get a true picture of how good or bad a team is. It’s mostly just a function of the kind of variance that likely won’t continue. And, since the entire point of xG is to strip out the noise and look at what’s likely to continue, away go the penalties.

Do you find yourself with time on your hands these days? Suddenly staying in on a Saturday night for the good of humanity? And just to top it all off, you have to seclude yourself with no sports to watch. Separately, have you noticed an explosion of numbers in football? A sudden rash of xGs springing up all over the place? Suddenly everybody seems to be spouting off about stats and you’ve got only the vaguest notion of what they’re on about?

Well, you’re in luck. StatsBomb copy-editor and general woman about town Kirsten Schlewitz is just like you! While she’s an expert at correcting the incredibly sloppy copy you’ve all come to know and love from me, she also came in as a relative novice at this whole stat thing. So, I roped her into asking me every question she could think of that she might have otherwise been afraid to ask. So, let’s get started.

K: First of all, let’s get the elephant out of the room. What is xG?

M: xG is short for expected goals. It’s a statistic that attempts to measure how likely any given shot is to become a goal. It’s really good at predicting the future. That is, xG is better at telling you which teams will score and concede goals going forward than any other statistic we have. That’s the most basic barebones definition I can think of.

K: If you already have xG to predict who will score and who won’t, why are so many other numbers needed? I see a great deal of figures and maps when I edit pieces, and sometimes I don’t understand what their purpose is. For example, when comparing two players, you can’t rely on xG for a team. So what numbers would be used there?

M: The answer to your first question is that knowing who is more likely to score goals going forward isn’t a particularly interesting thing to know (unless all you care about is betting, which, fair enough). The interesting questions are the hows and the whys. A single number like xG doesn’t help you very much with that. I like to think of xG as being a statistic that makes sure the conversation starts in the right place, as opposed to one that tells us anything remotely close to what we need to know.

So after the conversation gets started that’s why we need all the other stuff, to examine how teams play, what individual players are doing, basically what’s going on on the pitch that leads to the xG number at the end. And when those numbers come to particular players things can get very complicated very quickly. That’s because while xG works fine for players (specifically it can tell us when a player is on a hot or cold streak that’s unlikely to continue), shots are only a relatively small part of what’s happening. And, quite frankly, the further we get away from the actual shot, the less definitive our numbers become about what’s good and what’s bad, and the more we rely on them to try and accurately describe the game, as opposed to predict outcomes.

K: So you’re saying there are more or less two sets of numbers that a StatsBomb article could use: ones to predict which team will play better going forward, and ones that tell us what happened in a previous game (games?) in a way that dives deeper than simple match reports. If someone wants to write an article about, say, how they think the Champions League would have panned out this year, would they only use the prediction numbers, or would they also examine numbers that show what happened in previous matches?

Or am I way off base here and all the StatsBomb stats are used in conjunction with one another, rather than existing as two separate sets that focus on past and future?

M: So this is exactly right conceptually. The problem is that the numbers often overlap in ways which make the divide not particularly clear cut. For example, xG is an excellent stat for predicting the future, but it’s also a pretty ok one for explaining what happened. We know more about a match if we say that Arsenal had 1.5 xG than if we said Arsenal had 15 shots. Using the xG from a single game is kind of a quick and dirty way to describe what happened, albeit one with plenty of faults.

The best use of numbers though will always combine prediction and explanation. If I wanted to look at upcoming, now cancelled, Champions League matches, I would use general xG numbers as a starting place and say, “Here’s what I think will happen based on these numbers” and then use everything else to say, “And here’s why.” Now that also doesn’t mean xG is perfect. Doing good work in stats means trying to understand the limitations of the numbers as well so that we can understand when they might be missing something. So, in theory, it might be possible to analyze all the whys and hows and decide beforehand that even though a team like Liverpool might seem much better based on xG, they would struggle against Atléti (that’s not a conclusion I would have come to, but it’s not like completely beyond the pale to suggest).

K: We keep talking about “the rest of these numbers.” For someone who’s completely intimidated by stats, to the point they’re afraid to even click on a StatsBomb link, much less pitch you an idea, what other types of numbers would you anticipate they’d need to understand?

M: From a writing perspective, understanding the numbers is somewhat less important than understanding the game. If a writer is making accurate assertions about the game then those claims are going to be reflected in the numbers and in the editing process we can work together so that your friendly neighborhood StatsBomb editor (me) can help give you the appropriate statistical support you need.

So, if a writer wanted to write about how a team relied on a midfielder for a lot of their buildup play, they wouldn’t need to know the ins and outs of StatsBombs numbers. But I’d be able to call upon stats of ours like “deep progressions” to look at how frequently they move the ball up the field, or at passing percentages when they’re pressured and not pressured to explain how they’re cool in the face of a defense, or information and graphics on pass length, etc. etc. etc.

Now, if the numbers don’t match a writer’s argument that makes for an interesting challenge. The question of why a writer perceives the game a certain way while the numbers don’t capture it is generally a really exciting place to do analysis. Figuring out why there’s a disconnect between what the numbers capture and what the eye might see is usually an interesting endeavor for everybody involved.

K: I’m here editing and writing articles, and I fully admit I don’t comprehend exactly what half of these numbers mean. But if I wanted to submit an article that showed I do understand a few of the statistics, which do you think would be most important to understand?

M: You do need to understand the basic mechanics of xG and why it works so well. It’s important to understand that a player having more goals than xG expects he “should” is likely to start scoring less. Beyond that I’m looking less for knowledge of a specific stat than for a way of thinking about questions. Questions like, “Do you have a statistic that measures XYZ” are good, questions like, “How do you go about measuring ABC” are even better.

K: From xG and its variations (non-penalty xG, open play xG etc), it’s relatively easy to assess the offensive strength, or lack thereof, of a side, even if you’re new to stats — and I can attest to this, believing I had no ability to comprehend sports statistics before I took this job. But what still tends to confuse me is the defensive measurements . . . I see the maps and figures, but even those don’t help me quite get it.

M: Yeah. Defense is hard. We can look at xG conceded, or shots conceded, or any number of other things, but those are still fundamentally measurements about what the other team’s attack is doing. And that makes sense, because on some level all defense is is preventing the other side from attacking. But it’s also unsatisfying because defenders are obviously doing SOMETHING and it would be nice to describe what those things are.

The traditional measures are things like tackles, interceptions and blocks, and while those are useful numbers, they have some major problems. The biggest is that you can’t commit those defensive actions while you have the ball, so players on bad teams tend to have more defensive actions than players on good ones that keep the ball all the time. One thing we do is adjust all of those numbers for possession, to try and give a better picture of what’s going on.

On top of that we track pressures. That is, we track every time a defender is close to an attacker with the ball and impacting him in some way. This gives us a lot of information — adding pressures into the mix demonstrates where on the field a team is making defensive actions.That gives us the ability to look at a heatmap of a team’s activity and really get a picture of where on the pitch they like to defend (the redder the square the further above average the number of defensive actions are in the zone, the bluer the square, the further below). Manchester City defend basically in their opponents penalty area, for example.

All of that’s a long winded way of saying that it’s really really hard to evaluate defenses!

K: So we know how offense is evaluated, and we know how defense is judged — somewhat, anyway. With these two necessary halves of the game described, I have one final question: What would you like to see a writer be able to demonstrate with the numbers, keeping in mind that the StatsBomb blog is there to both educate readers and show potential purchasers what they can do with the data?

M: The major thing I want to see isn’t a specific proficiency with data, but rather a framework for thinking about issues. Think about a question you want to answer, and how can you use data to answer that question. That’s what we’re all trying to do, whether it’s determining if a potential signing will be worth it, or why a player is having a career year, or if a keeper's yips will pass, everybody is fundamentally doing the same thing. Whether it’s analysts with teams, or fans in the stands, or writers for StatsBomb, they’re looking at the game, developing a question and then trying to answer it.

Offside was supposed to be the easy part of VAR, they said. It’s a black and white decision, they said. Either he’s on or he’s off, they said. It’s perfect for technology, they said. They sure said a lot of stuff. Over the last week, Premier League watchers have been subjected to watching goals get ruled off after a long video replay delay when an armpit, or a toe, or a wisp of hair has been judged to have been ruled on the wrong side of a defender by mere centimeters. It’s worth prying apart exactly what is, and isn’t, going on, as the Premier League continues to try and implement the great VAR experiment.

First, let’s ground the discussion in what has (and hasn’t) changed. Surprisingly, there’s been remarkably little change in the actual things happening on the field. By almost any measure you can find, the game is basically being played pretty close to exactly the same way it was last season. Last year teams scored on average 1.37 goals per match (to be clear that’s per team, per match, the same way I’m going to list all the stats here), this season it’s 1.36 so far. Last season, on average, a team took 12.52 shots per match. This season it’s gone up a gigantic 0.03 shots per match to 12.55. If we’re desperate to pry things apart to find a change, the best we can do is look at clear shots (shots without a defender between the shooter and keeper) and find that while last season a team would on average take 2.20 per match, this season it’s down to 1.98.

Is that something? Is that evidence that with new VAR-driven offside rules players are no longer getting in behind as much, leading to almost half a shot (remember we’re looking at per game, per team, numbers here so that’s a 0.44 total per match, 0.22 for each team) less of good old defense-splitting, goal-scoring opportunities? Well, the shift in shooting numbers might be a thing, but it’s certainly not reflected in overall expected goal totals. So, probably not. Teams are averaging an almost identical non-penalty xG per match total of 1.25 this year, compared to 1.21 last season. Their xG per shot averages to 0.10 both seasons. Penalties, interestingly, are being awarded at ever so slightly lower rates with teams getting 0.11 per match this season, as compared to 0.14 per match last season (but really that’s a tiny tiny change to read anything into).

For all the noise surrounding VAR, the good news is it hasn’t, as yet, changed the overall nature of football in the Premier League at all. Many are concerned that VAR consistently awards tight offside decisions and negates goals, thus making it harder for strikers to get in behind, harder for teams to score, and in general decreasing attacking potency across the board. And there’s ample evidence that hasn’t happened.

For all that hasn’t changed, there is one fairly large difference in gameplay that has occurred from last season to this one. It’s definitely not one I expected to see though. There have actually been notably fewer offside decisions given this season than last. Weird, huh?

We can speculate about a number of reasons why this might be. I’d be being an irresponsible stats blogger if I didn’t start off saying natural variance could be the cause. Maybe the specter of VAR has changed how teams play, with strikers already reining in their runs lest they get penalized on replay. Or maybe this is a result of assistant referees being instructed to let the game run on close offside decisions in order to better utilize VAR and a number of those situations simply don’t result in goals, and thus are never examined. Whatever the reason, the actual stats are clear. In the era of VAR, there are actually significantly fewer whistles blown for offside than before.

Yet none of this should take away from the obvious impact that VAR has had on how fans experience the offside rule. In fact, it just makes answering the question harder. If fundamentally all that’s changed is that we have fewer offside whistled, then why does everything feel so much worse?

There are, I think, two interlocking issues here. There’s a technological one about the current system’s ability to actually generate the correct decision on close calls, and there’s the experience of the VAR decision itself.

Let’s take the second one first. Before VAR, there were plenty of close offside decisions given. Many were even wrong! But they were largely forgotten because, and I think this is crucial to the experience, they didn’t actually result in goals happening. The play would be blown dead, possession awarded the other way, and everybody would rightly go on with their lives. We didn't need to listen to 30 minutes of post-match talk shows host debating whether Jamie Vardy was a toenail offside on a random pass that might have set him free just before halftime because if the flag went up, that ended the move. There was literally nothing to see here.

Now, though, the experience is entirely different. Now, that decision is close, so the flag stays down until the end of the move, and if that pass results in a goal it’s reviewed either way. Even if the linesman would have flagged it last year, and he ultimately gives it this year, now it becomes a goal that’s chalked off due to VAR, while last season it was just one more close but unremarkable call. The entire top half of this article I talked about how the stats are unchanged. Well, that’s true, but part of that is because a goal like the hypothetical Vardy one I’ve painted, which gets ruled out for offside, doesn’t count in the stats. For stats purposes it didn’t happen. In terms of the game, it didn’t count. And yet, we all saw it.

Again, the important thing to emphasize about this hypothetical is that it’s a situation in which the call was the same before and after VAR. It’s a close offside call which, by the letter of the law, is correctly given in favor of the defense. Without VAR, it’s an entirely forgettable moment, with VAR it’s a goal that’s ruled out and a contentious talking point. Is that a good thing? It certainly isn’t on its own, but you could at least make the argument that it’s a worthwhile price for getting the increased accuracy that VAR provides.

Which brings us to the technological limitations. Exactly how accurate is the VAR camera system at detecting an offside infraction anyway? As Jonathan Wilson has covered, not as incredibly precise as we’d hope. The issue has to do with frame rates and pinpointing exactly when the ball leaves the foot of the passer. It’s important not to overstate the impact of uncertainty, and in reality, though the issue of camera frame rates not being fast enough to be entirely precise on the closest of calls is real, it also has an extremely marginal effect. But, regardless, it would be a fairly straightforward fix to introduce a margin of error into the system and only overrule calls that were outside the margin of error.

That would fix the particulars of the accuracy problem, but would only serve to knock the argument down the line. Instead of arguing about whether or not the reviewed call was correctly upheld or overturned, the argument would shift to whether or not the call fell within the bounds of being reviewable. You can’t actually technologize or rule write your way out of the reality that close calls exist. Wherever you draw the offside line, or the video reviewability cut-off point, there will always be calls that narrowly fall on one side or the other of that line.

The other option, if you really don’t like close decisions being decided either by humans or cameras based on incredibly fine margins, would be to dramatically change the offside rule and insist that the spirit of the rule, “gaining an unfair advantage,” be how an infraction is judged. If you think moving from the relatively objective standard of a line to a subjective judgment will make things less controversial . . . well, let me introduce you to the handball rule.

Ultimately, the technology is the easy part. It’s not only possible, but quite likely that the game can successfully calibrate its systems, it can account for the fact that at the closest of margins some things are fundamentally unknowable. Whether that accuracy is worth the price is another question. In order to get it right, the system has to elevate a small percentage of extremely marginal calls from obscurity to center stage. These are the kinds of calls that are the hardest to get right and the most controversial even when they’re correct. Without VAR they go largely unnoticed, even when they’re wrong. Now, thanks to the system, they’re in the spotlight. The process of being more accurate by definition shows us all what might have been, and then takes it away thanks to decisions that are juuuuuust this side of arbitrary. So, when judging VAR and offside, the only question that matters is, is it worth it?

As we role into Champions League match day two, the biggest teams from across Europe will inevitably get the attention. But, as the group stage progresses, a number of smaller teams will look to stake their claim to be this year’s Cinderella story. Here’s a quick look at some of the less heralded teams in this year’s Champions League and what they do, or don’t do well.

Slavia Praha

Things could not have gone better on match day one for the team from Prague. Not only did they travel to Italy and come back with a point from Inter Milan but Barcelona and Borussia Dortmund also drew, leaving the group wide open. Getting stuck in such a difficult group meant that their odds of even making the Europa League were always going to be long, but every Cinderella Story starts with a single invitation to the ball.

If their match against Inter taught us anything, it’s that this team plans to remain true to itself, no matter how challenging that may be. Manager Jindřich Trpišovský has his team press the opposition early, often, and unrelentingly. Through nine games of their domestic season, this is what their defensive heat map looks like.

In attack, their approach is generally to shoot early, shoot often, and then use the defense to get the ball and do it again. With such an active defense, shot quality comes second to simply burying opponents under an unending deluge of attempts on goal, even if many of those individual attempts are somewhere between speculative and downright deluded.

Of course, executing that approach when you have the most talent in your domestic league is a lot easier than it is when you go up against some of the biggest teams in Europe, but as we can see from their first match against Inter, they're sure going to try.

Advancing from this group remains an extremely unlikely proposition, but if it were to ever happen, this would be the way. First the draw against Inter, and now they get to host Borussia Dortmund at a time when the team seems to be ever so slightly off the top of their game. The German side is coming off back to back disappointing 2-2 draws against Eintracht Frankfurt and Werder Bremen dropping them to eight in the Bundesliga. Slavia Praha probably won’t shock the world, but if they are going to, it helps to catch their group stage opponents at their lowest moments.

Crvena Zvezda

It seems silly to suggest that a team that has started their season with eight wins and a loss is struggling, but sometimes it fits the bill. Sure Red Star sit tied for points at the top of the table and even have a game in hand on Bačka Topola the unlikely co-owners of the league’s best record, and yes the side sits one point ahead of bitter crosstown rival Partizan, but the stats suggest, that at least early in the season all was not quite right with this perennial powerhouse.

It’s hard for Europe’s smaller teams to compete on the biggest stage if they aren’t absolutely dominant at home, and Red Star simply haven’t been. On both sides of the ball it’s been Partizan that is the stand out performer (the following stats don’t include the last two match days, which is important to note were both comfortable 3-1 wins for Red Star). On both sides of the ball Partizan have the stronger xG numbers and thus a significantly higher xG differential so far this season.

This difference clearly showed itself when the two sides squared off earlier this season. It’s one of the most hotly contested rivalries in the world, and while Partizan left it until late to score their two goals, they were clearly the better team over the course of 90 minutes.

Even though their domestic form suggests Red Star might not have enough fire power to threaten any of the favorites in their group, Tottenham Hotspur’s opening week draw away to Olympiakos has left the door ever so slightly ajar. A good result this week at home against the Greek side combined with a Spurs loss to Bayern would actually vault the Serbian team into second place in the group and at least give them a fighting chance in the weeks to come.

Galatasaray

The Turkish giants have been a mainstay of the Champions League for years. They haven’t exactly threatened to do anything historic in the knockout stages, but they’ve always been a hard out for Europe’s biggest clubs, a side good for an upset or two under the right circumstances. This year, with groupmate Real Madrid getting trounced in the opener getting trounced in the opener by PSG’s backups, it at least seems to open the door for Galatasaray to pull off a shocking upset.

This year’s version of the Turkish side is going to have a real hard time driving through that door. Despite being filled with players you know and are at least mildly fond of, the team simply hasn’t been very good this season. Despite the likes of  Jean Michaël Seri, Sofiane Feghouli, Younès Belhanda, Steven Nzonzi and Radamel Falcao leaving viewers in a constant state of “oh yeah I remember that guy” the team hasn’t gelled together at all.

Right now, Galatasaray sit in just seventh place domestically. Despite all those names, they simply struggle to score the ball.

They neither create a lot of shots, 12.87 is the ninth most in the league, nor create particularly good shots, only six teams have a worse average xG per shot than their 0.07.

And the sad reality for fans of Turkish football is that if you’re not good enough to dominate at home, you’re probably not good enough to shock the world and outlast Real Madrid or PSG, even in a year when Madrid are struggling. Not all underdogs are created equal, and this year, Galatasaray have simply shown no signs of being able to do the things that they’d need to in order to upend the expected Champions League order.