A question on xG

[Silvio Dante]

So, TGHs shot on Saturday had a very low xG, I think about 0.03. My question is whether that xG is based on when he took the shot, or when he picked up the ball which was about 15 yards from release?

Reason for asking is both genuine interest, and whether anyone knows the metric starts from when the players “chance” and therefore xG phase starts at a more (unlikely) place.

To give another example , for Maradonas second against England 86 you’d expect the xG on release to be pretty high but practically zero from when he started the dribble. Where does the “count” begin and is it subjective if it’s not from initial possession, even if there isn’t a “chance” when the ball is picked up?

[Sleepy1968]

xG relates to the shot as it is taken per this

https://statsbomb.com/soccer-metrics/expected-goals-xg-explained/

I do wonder if some of the models take additional information into account.

For example, if the goalie is half way up the pitch, I'd think the xG might be higher, but maybe is it generally 'of all shots taken from here, this proportion result in a goal scored.'

[Davefevs]

From point shot is taken.

BBC counted Cornick’s “shot” from a Sykes cross in the 79th min.

which Wyscout didn’t.

5 minutes ago, Sleepy1968 said:

xG relates to the shot as it is taken per this

https://statsbomb.com/soccer-metrics/expected-goals-xg-explained/

I do wonder if some of the models take additional information into account.

For example, if the goalie is half way up the pitch, I'd think the xG might be higher, but maybe is it generally 'of all shots taken from here, this proportion result in a goal scored.'

Some are more sophisticated than others.

[IAmNick]

It's just the shot, not where they picked the ball up.

7 minutes ago, Sleepy1968 said:

I do wonder if some of the models take additional information into account.

Lots of sites do xG now, and most have their own models which is why you might see conflicting numbers.

The most basic would just be shots from the position. Nowadays they take into account loads of other things - for example Opta use the position of the keeper, how much of the goal is blocked by players, the "pressure" the player was under, the type of shot (which foot, header), the type of play the shot came from (corner, break, free kick), and so on.

I've seen some models take into account now the height of the ball as they receive it (i.e. at shin height will be less chance) and so on. It's extremely complex as you can imagine!

edit: One of the problems is as soon as you add another dimension, you greatly reduce the amount of available data which will then make your results less accurate. It's a delicate balance.

Edited November 27, 2023 by IAmNick

[cidercity1987]

Ah so these models don't take into account goalkeeper positioning?

[Silvio Dante]

Thanks all. I’m not unsold on the science but it’s definitely imperfect - taking the Maradona as the (extreme) example that wouldn’t be a chance when the player picks up but has a high xG. I’m more sold on seeing its value in the whether you’re creating chances which is the flaw here (probably not enough to make data unreliable overall).

There are definitely a few oddities from Saturday - Sykes angle was very tight and I’m not convinced 1 in 2 chances would be scored from there. Similarly, I don’t see the TGH early chance as a 1 in 3 because of (as has been said) defenders positions and how he recieved the ball. Conversely I think we’d all say Tommy should score more than one in ten times when he was played in by them and that’s reflective of where he was on the pitch, not the one on one scenario.

As an overall data collective and balance of game, I think it’s fine and broadly there. As an arbiter of individual chances though as those examples show just from one game, seems a bit flawed 

[Davefevs]

3 minutes ago, cidercity1987 said:

Ah so these models don't take into account goalkeeper positioning?

Yep, there is an extension of xG too, called PSxG (post shot XG) which takes into account what where on-target was placed, eg in the corner or straight at the keeper.

Someone like Harry Kane has a high PSxG because his shots are aimed into the corners.

https://thesporting.blog/blog/post-shot-expected-goals-what-is-it-and-why-does-it-matter

[Davefevs]

2 minutes ago, Silvio Dante said:

As an overall data collective and balance of game, I think it’s fine and broadly there. As an arbiter of individual chances though as those examples show just from one game, seems a bit flawed 

Yep, with all these things a larger sample size provides more use.  One game, one chance isn’t really what it’s meant for.

[cidercity1987]

4 minutes ago, Davefevs said:

Yep, there is an extension of xG too, called PSxG (post shot XG) which takes into account what where on-target was placed, eg in the corner or straight at the keeper.

Someone like Harry Kane has a high PSxG because his shots are aimed into the corners.

https://thesporting.blog/blog/post-shot-expected-goals-what-is-it-and-why-does-it-matter

Wow five years later and I finally understand the model

I agree with Silvio Dante that as a generality it's probably ok but in terms of specific chances it's a load of old bollocks

[1960maaan]

 

16 minutes ago, Davefevs said:

Yep, there is an extension of xG too, called PSxG (post shot XG) which takes into account what where on-target was placed, eg in the corner or straight at the keeper.

Someone like Harry Kane has a high PSxG because his shots are aimed into the corners.

https://thesporting.blog/blog/post-shot-expected-goals-what-is-it-and-why-does-it-matter

I've not read about XG for some time , I think they were just improving where defenders were etc . Looks like Statsbomb have stepped up a bit. This is a good explanation for us that struggle.

 

https://statsbomb.com/soccer-metrics/expected-goals-xg-explained/

Edited November 27, 2023 by 1960maaan

[johnbytheriver]

It was a great goal! but not an expected goal xg.More like an unexpected goal uxg. Good performance all round from the young man though!

[Mr Popodopolous]

For me, xG isn't even so useful game by game or change by chance, while not flawless I use it to extrapolate trends a bit in line with more basic data.

By way of example side scoring plenty despite unremarkable chance creation etc will either dip or have to improve underlying numbers drastically to maintain the level or not fall sharply down the line.

Likewise with defensive output vs chances allowed.

A side with good underlying numbers but poor output usually plays its way towards respectability in the end.

[Silvio Dante]

8 minutes ago, Davefevs said:

Yep, with all these things a larger sample size provides more use.  One game, one chance isn’t really what it’s meant for.

It’s not really comparing apples with apples though is it? Taking the Tommy example, I’m more than happy to say percentage wise 16% of shots taken from that position are scored. But if you overlay it being 1 on 1 then it should be 50-60%. The xG overall ends up being roughly where it should because the overquantifying Sykes goal balances it out. The xG from the actual position though is nonsense.
 

I get I’m probably asking the impossible as every chance is different based on pace on the ball, angle of receipt, defenders in the way etc. But, in effect, narrowing xG down to a chance quality just from where the shot is taken is a massive oversimplification, irrespective of the number of data points from that position. Enough that it’s not reliable and should be seen as a more “balance of play” metric as opposed to “quality of chances” which it’s frequently badged as.

Incidentally Garnachos goal had an xG of 0.08. Meaning he had as much chance of scoring as Dickie did with his header!

[Davefevs]

11 minutes ago, Silvio Dante said:

It’s not really comparing apples with apples though is it? Taking the Tommy example, I’m more than happy to say percentage wise 16% of shots taken from that position are scored. But if you overlay it being 1 on 1 then it should be 50-60%. The xG overall ends up being roughly where it should because the overquantifying Sykes goal balances it out. The xG from the actual position though is nonsense.
 

I get I’m probably asking the impossible as every chance is different based on pace on the ball, angle of receipt, defenders in the way etc. But, in effect, narrowing xG down to a chance quality just from where the shot is taken is a massive oversimplification, irrespective of the number of data points from that position. Enough that it’s not reliable and should be seen as a more “balance of play” metric as opposed to “quality of chances” which it’s frequently badged as.

Incidentally Garnachos goal had an xG of 0.08. Meaning he had as much chance of scoring as Dickie did with his header!

I think you’d be surprised at how many 1 on 1s aren’t taken!  Was saying same to Harry last night.  If you think a penalty is 0.76, is a number of yards closer, keeper on his line, etc, I don’t think 0.5-0.6 is likely.

Nahki Wells v Stoke (which was scored) was 0.09, slightly more difficult angle.

When you measure actual goals scored in the Championship vs total xG, it’s pretty close to equal.  Average 23.78 xG per team, 23.25 goals scored.

There will always be individual cases where we think it’s a bit out of kilter.  I’d probably give TC’s 0.25, one-in-four.  But that might be because we also think TC is a good finisher too.

Statsbomb is the most sophisticated but Wyscout is decent enough.

Bold bit - I agree 100%

[Silvio Dante]

13 minutes ago, Davefevs said:

I think you’d be surprised at how many 1 on 1s aren’t taken!  Was saying same to Harry last night.  If you think a penalty is 0.76, is a number of yards closer, keeper on his line, etc, I don’t think 0.5-0.6 is likely.

Nahki Wells v Stoke (which was scored) was 0.09, slightly more difficult angle.

When you measure actual goals scored in the Championship vs total xG, it’s pretty close to equal.  Average 23.78 xG per team, 23.25 goals scored.

There will always be individual cases where we think it’s a bit out of kilter.  I’d probably give TC’s 0.25, one-in-four.  But that might be because we also think TC is a good finisher too.

Statsbomb is the most sophisticated but Wyscout is decent enough.

Bold bit - I agree 100%

We might be able to use the data from when the NASL did their “shot clock” tiebreak!

Adding player ability into it is a metric that I think probably goes too far - as it’s also subjective. 
 

I think we’re broadly on the same page - it’s got value as a metric but not as the metric which people use it for. Story of every data point I’ve ever worked with…!