Football Stats Explainer: xG

What you should, and should not, "expect"

With the World Cup about to kick off, we’re going to be watching a lot of football. That means people will be talking and writing about football endlessly, myself included. Now seems as good a time as any to brush up on some of the basics of modern analytics-based conversations. In this article we’ll look at xG, which has become the sport’s most famous metric. Then I will write a second article looking at some other concepts you may or may not be familiar with.

At this point, I think most people are familiar with “expected goals”, or xG. If you genuinely aren’t, I’ve got you covered anyway. But I still think a lot of the mainstream coverage of the metric is often misguided, and fails to understand exactly what it tells us. Expected goals have popped up absolutely everywhere without all involved necessarily understanding the utility beyond a new way to get social media engagement. When people are using xG to complain that a striker can’t finish over a sample of ten games, or that a manager is too cautious because his team keep missing chances, there’s an issue. So hopefully I can bridge some knowledge gaps while giving a better explanation of certain terms I use in this newsletter.

While I wouldn’t call myself an analyst, I did write a regular column for data analytics company StatsBomb back when they were still posting daily articles, and I broke into football writing specifically through the analytics space. I could never claim to know as much as the true data experts working inside clubs now, but I would happily argue I have a stronger and more experienced grasp on this stuff than most journalists.

Note: all the numbers in this article are pulled from FBRef, using data provided by Opta.

What is xG?

Ok, let’s start with the very basics. Football is about goals, right? Score more than the opposition and you win the game?

Where do goals come from? Shots. Again, this is obvious. So when people started getting serious about collecting data in football, shot totals were the obvious place to start. This, like many football analytics concepts, was imported from hockey. The more shots, the more chances you had to score a goal, so the best teams would take the most shots and concede the fewest. Simple, right?

Turns out, not so much. Not all shots are created equal. It’s much easier to score from a yard away from goal than it is to score from outside the box. Players are more accurate shooting with their feet than their head. All of this is pretty intuitive, but it needed to be measured. Further models have added in other elements such as the location of the defenders or whether data companies measure a situation as a “big chance”, but the very basics are about the type of shot and the location where it took place.

Expected goals, then, is an attempt to put a number on how often a chance goes in the back of the net. You can add them up and find that, for example, Arsenal created 2.1 expected goals against Chelsea, despite only scoring once. It’s taking the common notion that a player “should have scored there”, and trying to put a number on it.

What isn’t xG?

Whenever people get their hands on these numbers, they always want to flip it around and use it to tell us which players or teams are the best and worst finishers.

It can’t really do that. That’s not what the models are built to tell us.

When analysts first started applying xG to football, there were two obvious findings. The first was that long shots are less likely to go in than most people seemed to assume. This played on a pretty intuitive football logic that you’re more likely to score sitters than long range screamers, but no one had realised just how true it was. Players’ mental models were slightly off, probably due to the euphoria of that one time they scored from thirty yards. Long shots were usually bad bets, and the players would be better off working the ball into high quality situations.

The other runs totally counter to how everyone has understood football for decades: finishing skill might not be all that. The vast majority of players score goals at a rate close to what an xG model would, well, “expect” them to score at. Even the very best aren’t that far off it. Most everyone who has studied this will tell you that Lionel Messi is probably the greatest finisher football has seen in the modern era. And yet, since the start of the 2017-18 season (as far back as FBRef will provide data for), he’s only scored about 30% more than his xG would imply.

So if you see a player scoring 60% more than their xG in a given season, ask yourself a simple question: is this player twice as good at kicking the ball into the back of the net as Messi, or are they on a hot streak that will come to an end?

Messi’s most obvious comparison doesn’t even have that going for him. Over the same stretch of time, Cristiano Ronaldo has scored 97 goals from an xG of 96.9. Suggesting that one of the greatest goalscorers of all time is a bang average finisher. I don’t personally believe this, but I think it makes clear that the xG models have a hard time sniffing out the true clinical finishers from the rest.

So if xG can’t tell us which players can finish, what use is it?

xG is valuable not because of which players and teams differ from it, but because the vast, vast majority will conform to it over time.

Cristiano Ronaldo is (or, depending on who you ask, was) extremely good at scoring goals. He’s so good at it because of his movement, his athleticism, his first touch and, most of all, his ability to think quicker than anyone else on the pitch. What that means in practice is that he gets so many more shots than everyone else, often in dangerous situations, and then turns those into goals. xG tells us that he is an excellent goalscorer and, like most, they do it by getting a huge volume of good chances rather than converting at an incredible clip.

Once we know players and teams should conform to their xG over time, we can see trends that are and are not likely to continue. Take José Mourinho’s Manchester United, for example. United finished second in the 2017/18 season with 81 points. This was miles behind the runaway title winners Man City, so the vibes weren’t great. But looking purely at results, it was reasonable to expect that United could continue improving and close the gap further with the right signings.

Except that once you bring xG into the picture, things look a whole lot worse. United scored about 12 goals more than expected, while also conceding around 13 goals fewer than the model suggests they “should have” let in. None of this was rocket science, and anyone with a calculator could tell you that the club should lower expectations for the following campaign. Just as the data suggested, things started to go very wrong, very quickly. Some people will still tell you that this second place finish is a high watermark for the post-Ferguson years, and that Mourinho deserves a great deal of credit for achieving it. The xG models suggest that it was just an outlier where things broke United’s way, and not indicative of any real quality.

This gets at something important: while xG correlates closely with goals over many years, one season is not always a large enough sample size to see this happen. Conventional football wisdom says that luck evens itself out over 38 games, but the data is in and, frustratingly, one season isn’t nearly enough.

These models are imperfect. The xG value of a single shot might be pretty far from what the “true” value should be for any number of reasons. But over a larger sample, the law of large numbers will kick in and

Ok, I get all that. How can we take it further?

I think a big problem with expected goals is that when people first encounter them, it tends to be measuring single game results. That’s ok as a usage, but it’s not very exciting. You probably already had an opinion on who “deserved” to win from watching the game.

I use “xG difference” a lot in this newsletter, and in hindsight I should have explained it more thoroughly. It just takes the idea of goal difference, but uses expected goals instead. It’s just the xG a team creates minus the xG they concede. From there you can divide it down to a per game basis, giving us more of an xG-based team rating. For example, Liverpool had an xG difference per game last season of +1.52, whereas this year it’s only +0.30. Obviously you’d need to break that down a lot to understand the whys and hows, but it gives you a baseline sense that, yes, Liverpool really have got worse this campaign.

xG per shot is another revealing indicator. This is just the total xG divided by the number of shots, giving you the average quality of chance that a team is creating or conceding. This season, Arsenal have the Premier League’s highest xG per shot in attack, at 0.13. That means the model estimates the average Arsenal shot has a 13% chance of being scored. Arsenal haven’t taken as many shots as Liverpool, but the shots they do take have been of a better quality, perhaps indicative of intricate passing sequences that work the ball into dangerous areas. Or perhaps it’s a small sample size being diluted by a few outlier chances. That’s the problem with all metrics in a low-scoring sport like football.

That low-scoring problem is both why xG exists (goals are too rare to build useful stats around), and why it hits a limitation. There is no one number in football that lets you stop using your brain. Nonetheless, people will still try to argue whatever their preconceptions are because a team scored once from its last nine expected goals. When used properly, xG can tell us a lot about football. Just be careful for the ignorant and the grifters out there.