One stat to rule them all (xG/xA)

[samyadav]

Just in case anyone wants to look up who's who:

If you had to choose a player using those rankings, what will you consider: xG, xA or both? For example Silva is ranked 1 in xA and Salah is ranked 4 in xG. Does that make Silva a better pick if i had to choose only one of the two?

[dino1980]

Both, but goals are worth more FPL points than assists so goals first. However, of more importance is to look at the numbers to see if someone is over/underperforming their xG or xA. (you can also add a custom column on understat to combine xG+xA fwiw).

In the above scenario then Salah by a long way. Strikers should, in general, be top for xG so for Salah - a mid in FPL - to be fourth out of everyone is a testament to his goal threat this season.

[Stemania]

Here's another graph on a the xG topic. (Sorry, on holiday with too much time on my hands).

Tall Paul supports some rubbish team from the north west and postulated certain teams could consistently outperform their xGA (expected goals against) due to their individual style of play.

It sounds very plausible that a team could slightly overperform regularly, but since xGA already takes into account the number of and quality of chances conceded by a defence (a widely used metric for measuring defensive quality) it seems tough to really say anything concrete about how specific defensive organisation might actually affect the relative performance vs xGA - the likelihood of conceding a chance and the quality of it has already been taken into account. Are some teams just plain more likely to stretch that extra sinew and consistently get an extra block in from the same defensive position (are they somehow more committed)? It could be that the current xG models aren't quite sophisticated enough to take blocking likelihood of each shot fully into account - it is for sure a factor in the quality of the chance, but maybe the models are quite far from perfection on that front - though eventually I'm sure they will be very strong on that too.

Individual players, especially GKs, will definitely make a small difference since xGA assumes (the shooter and) the goalkeeper are of 'average quality'. So we should expect Man United with DDG to slightly but consistently overperform their xGA as DDG will genuinely save more chances than your average GK. Whereas <insert colloquial team with rubbish GKs here> Liverpool might suffer the opposite fate on that factor.

Either way, I don't think this answers the question (because the data is so little and there are so many other factors), but here's a chart of all the teams GA vs xGA scores for the last 4 years (which exhausts understat's data), which might be interesting. I took out the teams that have only played 1 or 2 of those seasons. Bear in mind the current year is only half done so the darkest blue bars will be slightly more erratic. The values given are GA-xGA, so a negative score indicates overperformance vs the metric and a positive score indicates underperformance. Maybe I'll call it the TPT (Tall Paul Tabulation):



Some teams have consistently overperformed, some under, some a mix. Obviously big swings are going to happen because the game is heavily based on probability, and teams may have adopted different styles of play/signed key new players in this period. Whether the difference between general trends with certain teams is down to probability/natural variance/individual GKs/styles of play is unclear. Presumably it's a mixture of many such factors.

[Stemania]

OK, so I realised an issue with choosing raw underperformance/overperformance in goal numbers as the scale; it over-exaggerates the numbers for those who have conceded many goals (since they will naturally have a larger disparity). So I re-scaled the GA and xGA disparity as a percentage of GA.



But, unfortunately it is significantly more difficult to read since its so dominated by the (frankly ridiculous) Burnley/United GA vs xGA overperformance so far this year, but hopefully still legible.

[Tall Paul]

I appreciate the name check, do I get any royalties?

For me, all those graphs are saying is that there is too much variance (for whatever reasons) in xGA for it to be much use as a predictor.

[Stemania]

A lot of variance doesn't mean it's a poor model necessarily. You're going to get variance in any model (so long as that model isn't just goals conceded), but what distribution/amount of variance we should expect from a 'perfect' model I don't know - there'd have to be some kind of probabilistic model of football in place. So whether the above shows too much for it to be useful I'd refute (or at least, not have the ability to answer properly). At least if I remembered enough statistics I'd be able to normalize this seasons data to take into account the fact we've only had half the games (using sample standard deviations etc?), and that would make the dark blue look slightly more reasonable.

Besides, I don't think anyone considers xG as a predictive tool as such - just as a way of analyzing what has happened, to try to analyse what of what's happened is genuine and what of it is variance.

[Tall Paul]

Besides, I don't think anyone considers xG as a predictive tool as such - just as a way of analyzing what has happened, to try to analyse what of what's happened is genuine and what of it is variance.

I think people that say "team x has been lucky and they're going to regress, look at their xG" are using it as a predictive tool, no?

And if it's not used as a predictive tool, what use does it have?

[Stemania]

To make predictions there's so many factors in the melting pot, upcoming fixtures, changes of manager/strategy/signings/form, injuries. No-one's going to base predictions predominantly on xG values, it would be incredibly lazy/poor analysis just to extrapolate to the rest of a season for example. But to make any predictions you certainly need to study what's happened so far. The point of xG is simply an attempt to give a better description of what has happened so far than what has actually happened so far! It's just a way of trying to remove some noise from the data (and potentially recent-points-bias from decisions based on what has happened).

''They're going to regress'' is just a broad statement saying a team has overperformed what they 'should' have got from the data so far - it's not a prediction really, it's a critique of the idea of making any prediction that is based heavily on what has happened so far, data which the xG model suggests is skewed. I.e., it's basically saying any expectations of what is to come (that have been strongly influenced by what has thus far happened) should maybe not be quite so optimistic, for example.

[Ruth_NZ]

Apropos of nothing, do the xG models factor in the degree of pressure a player is under when shooting or are all shots from the same location considered to have the same chance of resulting in a goal?

Burnley defend deep, in numbers and aggressively. They don't stand off and let someone shoot from 15, 18, 20 yards. There will be Burnley players sprinting to get to the ball, block the shot, make it difficult... it doesn't matter whether the shot is actually blocked if the player shooting is under pressure, has no time and is maybe having to rush. That's one reason why Burnley GKs have traditionally had a high save %.

There are all kinds of reasons why a team like Burnley may out-perform xGA. That is probably one of them.

[sagorjanski]

Apropos of nothing, do the xG models factor in the degree of pressure a player is under when shooting or are all shots from the same location considered to have the same chance of resulting in a goal?

Burnley defend deep, in numbers and aggressively. They don't stand off and let someone shoot from 15, 18, 20 yards. There will be Burnley players sprinting to get to the ball, block the shot, make it difficult... it doesn't matter whether the shot is actually blocked if the player shooting is under pressure, has no time and is maybe having to rush. That's one reason why Burnley GKs have traditionally had a high save %.

There are all kinds of reasons why a team like Burnley may out-perform xGA. That is probably one of them.

I'd recommend this article by Rory Smith. Echoes a lot of what you're pointing out.

[blahblah]

If you want the above in less words: elsewhere someone was espousing the goalless Ince Jnr due to his stats.

I was going to counter this with the idea that he is actually 5hite (I have pointed this out elsewhere on FISO in this season) and he has more chance of scoring in a Brothel . Posters can throw all the Championship Stats (something else that was mentioned) around that they like, but nearly all of those mean diddly squat: as the Prem is totally different. (The only relevant one, imho, is if one player scores a igh percentage of a Clubs goals\assists.)

Basically using past performances against a notion of Expected Goals shows how good a finisher a team\player is and how good they are stopping shots. For this stat to be worthwhile it has to be relative to the player taking the shots from the specific part of the pitch etc, similarly defenders and which sort of shots from where and BY WHICH OPPONENT.

Out of interest: are there any examples of when this stat has helped over and above common sense and the eye test? By Common Sense I would mean things like: Burnley (WBA of a few seasons ago etc) are tight at the back and their defenders are cheap etc. By the Eye Test I mean things like: how did Ince miss that and that and that, I'm not buying him

[Ruth_NZ]

I'd recommend this article by Rory Smith. Echoes a lot of what you're pointing out.

Excellent! Paul will love that. Believe it or not I had the aspect of cohesion (team attitude and resilience, in which I would include the supporters) in mind as well but thought mentioning that was maybe a bridge too far. Thanks for posting it.

Every article I read about Dyche seems to send him up in my estimation.

[Ruth_NZ]

Deleted/double post.

[Finisher1]

Deleted/double post.

You can delete your post if it's the latest post in a thread.

But now you can't anymore because I posted this

[Stemania]

As far as I understand it, taking into account the position of defensive players is essentially the main difficulty in perfecting xG models - all the stuff with shot position, which foot/header, game-state situation, angle to goal etc are in some sense just data collection/aggregation based on the logged match data. As mentioned, I think we'd all can expect this to get better with time.

Each model currently approaches the problem slightly differently I think, for example OPTA first split all the opportunities situationally (whether it was a through ball, so less chance of defenders being in the way, whether the shooter has dribbled past a player first, so less chance of being blocked by that player etc). Plus, famously they use their 'Big Chances' metric for judging whether a chance is actually good or not (so, no defender standing right in the way etc) as was discussed at the start of the season in a related thread - this article Blue Fire linked there is the best I know of discussing the issues with Big Chances (not that it should necessarily be thrown out).

On the other hand, Understat use neural networking for their modelling, so it's even possible they don't know themselves precisely how some final probabilities are calculated if it's predominantly based on machine learning. (Not to say it isn't a better method - it probably(?) is - just that not knowing everything about the solution you come up with is in the nature of neural networking).

[On Burnley, I'm not a huge fan of the premise of the article linked tbh - I like some of the stats mentioned, but if Burnley's style/quality of defending was so distinct to break the model we'd have seen the same thing in previous years. Burnley have been well within reasonable GA-xGA levels previously, it's just this year that's been mad and much of that is quite clearly down to variance - high team cohesion or not, having 5-10% or so higher proportion of chances blocked than many other teams obviously doesn't translate into stopping roughly 50% of the chances that 'would usually' be scored (which is roughly the level Burnley, and United, have managed so far this season)! For example, Burnley managed their best ever block per game tally last season (beating this season's current rate) and still ended up within a goal or two of their xGA at the end of the season. I'm sure there is truth in the article in that Burnley might be the type of team the model struggles with most for the reasons given, but the huge disparity in the data can't be explained away with just some nice (well written) analogy imo. In truth Burnley's block percentage isn't actually that far ahead of a few other teams for example; Burnley's PL average is 34% (this year 36%), but West Brom's 4-year average is 32% (this season 34%) and West Brom are one of the teams that have best matched their xGA tally over the years.

So, in terms of this year's numbers, as much as it's nice and romantic and contrarian to think otherwise, I'd say Burnley have for sure been fortunate this season from a goals conceded perspective (and some favorable fixtures). But they've also defended extremely well. Put those together and I guess it equals reaching top 4.]

I do think there's inevitably going to be some truth in that making a higher percentage of blocks will likely lead to a slight overperformance vs xGA. Here's a graph of percentage of chances blocked vs overperformance in GA (against xGA) - OK so the correlation doesn't look that strong (and it's simply an excel line of best fit and I didn't calculate any coefficients, edit: in fact any correlation looks quite weak), but is very likely in the melting pot of a team producing a small but regular overperformance:



Edit: btw, each blue dot is one season's PL data for some team (from the last 4 years), so 3 of the dots are Burnley, 4 are Chelsea etc.
Edit2: probably should have put the axes the other way round...

[Valeron]

Nice summary that, thanks stemania.

I think something also worth considering is this. Burnley-Stoke match recently was a 50-50 affair with Burnley at HOME when the I checked the bookies odds. The betting odds for the EPL are known to be extremely efficient for a variety of reasons. That matched happened with Burnley flying high and Stoke struggling for results. I think those odds tell you all you need to know about whether Burnley have been lucky so far or not.

[eastcentral1]

Here's another graph on a the xG topic. (Sorry, on holiday with too much time on my hands).

This is great, thanks. What is the typical margin of error on the xGA? It seems to me that we need to know that before we can talk about teams under or over performing relative to xGA. For example, there is an outstanding question (in my mind at least) as to whether Burnley actually overperformed relative to xGA last season, or whether the xGA accurately reflected their goals conceded.

[blahblah]

re Burnley: what were their stats at home last season? How do they compare to Home and Away this season?

[Stemania]

Basically using past performances against a notion of Expected Goals shows how good a finisher a team\player is and how good they are stopping shots. For this stat to be worthwhile it has to be relative to the player taking the shots from the specific part of the pitch etc, similarly defenders and which sort of shots from where and BY WHICH OPPONENT.

Out of interest: are there any examples of when this stat has helped over and above common sense and the eye test? By Common Sense I would mean things like: Burnley (WBA of a few seasons ago etc) are tight at the back and their defenders are cheap etc. By the Eye Test I mean things like: how did Ince miss that and that and that, I'm not buying him

You've actually already suggested how xG generally is worthwhile. Without stats, how would you usually judge who is a good finisher and who is a good shot stopper? You watch stuff and think, "ooh this guy scores/saves more from this position/situation than a normal player wouldn't'', right? That 'normal player' idea is exactly xG, so everyone already uses xG anyway, just not so systematically or explicitly.

So I see the use of xG is two-fold. Say for an individual, first it tells you whether a team has been creating enough chances for a player (a high xG value), then secondly it tells you how proficient/fortunate that player has been with those chances (G vs xG).

The first one gives you an idea of whether a team has actually been playing well or just had the run of the ball on the attacking front - maybe their opponents just gifted them a goal or two against the run of play and the striker essentially got a 'free' brace etc. So, what I would say one of the most useful things with xG is at judging what is more likely to be sustainable and what is more likely to be happenstance. For example, all season the xG values of City players have suggested City are the real deal not a flash in the pan.

The second one tells you how good each individual has been as a finisher. Most players vary as to whether they under- or over- perform their xG between different seasons, but you could/should view consistent overperformance of xG as evidence of how good a player's finishing ability is. If you have a player transferred to a new team who's just scored shed loads of goals, should we expect him to do as well? Was his goalscoring tally because his team simply created loads of chances for him, or was it because he's actually a brilliant finisher?

Take, I dunno, Kane and Lukaku for example. They both had amazing seasons last year overperforming their xG by a big margin. Now, every year for the past four seasons Kane has overperformed his xG slightly (this season so far is actually his worst, only just ahead) so that's good evidence he's a fantastic finisher. Lukaku on the other hand has only overperformed his xG one in the last 4 years (last year), otherwise he's underperformed by a couple of goals a season pro-rata. So that's tentative evidence (which may or may not back up eye-test impressions) that Kane is a better finisher than Lukaku, and might be something to take into account in a decision when judging their value for FPL.

For defenses, the beauty of xGA is that over a season it does take into account all opponents (by averaging) - like all data, the longer the time period the better. The difference to individual player xG is that it also takes into account much about the relative quality of each defence already, in that the chances conceded have already been conceded (much of the quality of a defence comes from avoiding giving up chances). Viewing the previous conversation from the opposite perspective, you could indeed say that a defence consistently overperforming xGA is partial evidence the individuals (especially GK) are better than average at the last second stuff. You can put in the reasons why - DDG's quality at United, maybe defenders putting strikers under pressure slightly more often, blocks, cohesion(?) etc - but all that is much harder to quantify than just 'how good someone is at finishing' and tough to really be concrete about I think. Even with that said, to judge whether a defence is good overall with the stats you still fundamentally want to be focusing at who is conceding the fewest xGA, any slight under/over performance is maybe just a little bit window dressing. You can be as great as last ditch saves as you like but it isn't going to help so much if you concede chances willy nilly. Same with a striker, it doesn't matter so much how good a finisher a striker is if he's not getting the chances.

[Stemania]

What is the typical margin of error on the xGA? It seems to me that we need to know that before we can talk about teams under or over performing relative to xGA.

Not sure there's enough data for that, but have a box and whisker anyway:


(no prizes for guessing the two outliers)

For example, there is an outstanding question (in my mind at least) as to whether Burnley actually overperformed relative to xGA last season, or whether the xGA accurately reflected their goals conceded.

They essentially conceded the amount the 'should' have last season compared to the understat model; last year Burnley's xGA was an 0.27 goal overperformance, which is obviously ridiculously far within the realms of being in line - in fact, you wouldn't expect it to be that close, it's probably luck(!) that it is.

In fact, it looks like the dials have been tweaked slightly in the last day or so (meaning my numbers previous are a very small percentage out) - I think it was around 1 under before, not that 1 and 0.27 are really any different over a season). Last year it represented roughly a 9% overperformance, on the new numbers less.

The difficult bit will always be deciding for sure whether an underperformance or overperformance is due to variance or something real. Obviously all of this (and all FPL decisions) are based on such small sample sizes that it's impossible to make actual proper statistical judgement, so all stats based stuff is still guesswork - just (hopefully) informed, pretty, pretentious guesswork.

[blahblah]

Basically using past performances against a notion of Expected Goals shows how good a finisher a team\player is and how good they are stopping shots. For this stat to be worthwhile it has to be relative to the player taking the shots from the specific part of the pitch etc, similarly defenders and which sort of shots from where and BY WHICH OPPONENT.

Out of interest: are there any examples of when this stat has helped over and above common sense and the eye test? By Common Sense I would mean things like: Burnley (WBA of a few seasons ago etc) are tight at the back and their defenders are cheap etc. By the Eye Test I mean things like: how did Ince miss that and that and that, I'm not buying him

You've actually already suggested how xG generally is worthwhile. Without stats, how would you usually judge who is a good finisher and who is a good shot stopper? You watch stuff and think, "ooh this guy scores/saves more from this position/situation than a normal player wouldn't'', right? That 'normal player' idea is exactly xG, so everyone already uses xG anyway, just not so systematically or explicitly.

So I see the use of xG is two-fold. Say for an individual, first it tells you whether a team has been creating enough chances for a player (a high xG value), then secondly it tells you how proficient/fortunate that player has been with those chances (G vs xG).

The first one gives you an idea of whether a team has actually been playing well or just had the run of the ball on the attacking front - maybe their opponents just gifted them a goal or two against the run of play and the striker essentially got a 'free' brace etc. So, what I would say one of the most useful things with xG is at judging what is more likely to be sustainable and what is more likely to be happenstance. For example, all season the xG values of City players have suggested City are the real deal not a flash in the pan.

The second one tells you how good each individual has been as a finisher. Most players vary as to whether they under- or over- perform their xG between different seasons, but you could/should view consistent overperformance of xG as evidence of how good a player's finishing ability is. If you have a player transferred to a new team who's just scored shed loads of goals, should we expect him to do as well? Was his goalscoring tally because his team simply created loads of chances for him, or was it because he's actually a brilliant finisher?

Take, I dunno, Kane and Lukaku for example. They both had amazing seasons last year overperforming their xG by a big margin. Now, every year for the past four seasons Kane has overperformed his xG slightly (this season so far is actually his worst, only just ahead) so that's good evidence he's a fantastic finisher. Lukaku on the other hand has only overperformed his xG one in the last 4 years (last year), otherwise he's underperformed by a couple of goals a season pro-rata. So that's tentative evidence (which may or may not back up eye-test impressions) that Kane is a better finisher than Lukaku, and might be something to take into account in a decision when judging their value for FPL.

For defenses, the beauty of xGA is that over a season it does take into account all opponents (by averaging) - like all data, the longer the time period the better. The difference to individual player xG is that it also takes into account much about the relative quality of each defence already, in that the chances conceded have already been conceded (much of the quality of a defence comes from avoiding giving up chances). Viewing the previous conversation from the opposite perspective, you could indeed say that a defence consistently overperforming xGA is partial evidence the individuals (especially GK) are better than average at the last second stuff. You can put in the reasons why - DDG's quality at United, maybe defenders putting strikers under pressure slightly more often, blocks, cohesion(?) etc - but all that is much harder to quantify than just 'how good someone is at finishing' and tough to really be concrete about I think. Even with that said, to judge whether a defence is good overall with the stats you still fundamentally want to be focusing at who is conceding the fewest xGA, any slight under/over performance is maybe just a little bit window dressing. You can be as great as last ditch saves as you like but it isn't going to help so much if you concede chances willy nilly. Same with a striker, it doesn't matter so much how good a finisher a striker is if he's not getting the chances.

Yep to agreeing. My point\case is the Ince one really. I read posts quoting these stats and saying that he will score loads.... Phillips was the better one during Blackpool's season, imho. Ince has shown jack 5hit since, or he wouldn't be at Huddersfield and hasn't scored this season. In less words he is 5hite (relativley speaking) and his high Xg stat just shows how bad he is in front of goal.

This is not the same as an Aguerro who is below par for a season and we can expect to score more to get back to "average". I consider Kane as very differently as he got\gets a lot from outside the box and as I was vilified for a season or two back, is a confidence goal scorer.

My question about these stats is simply this: do I need them to tell me that? Similarly for defences: there are CS tables.....

(Yeah I know I am pants at FF compared to years ago, but that is the joy\point of FISO that it considers opinions etc rather than "your 5hite (not as higly ranked as me) so fook off" )

[Stemania]

My question about these stats is simply this: do I need them to tell me that? Similarly for defences: there are CS tables.....

I would say that these stats are basically the same information you would usually use, just better/more detailed information (based on far more data).

So, a crass example for xGA: Let's say you have two teams, Team A has drawn three games 1-1, conceding one penalty per game, Team B has won three games 1-0, but an opposing striker skied a penalty in each game. Each penalty has roughly a 70-80% chance of going in regardless of the team, so Team B have really just been quite lucky and Team A a little unlucky. We might look the CS table and conclude Team B looks like it might have a far superior defence, they're 3-0 up in CSs (!), but xGA would tell you that basically the defenses are the same and that the CS table is skewed. Obviously, 'a penalty' is replaced with 'all shots', in practice.

[blahblah]

But if the GK had saved the 3 pens.....

[Stemania]

But if the GK had saved the 3 pens.....

...then he would inevitably shoot up in price and be brandished the new McAuley.

[blahblah]

But if the GK had saved the 3 pens.....

...then he would inevitably shoot up in price and be brandished the new McAuley.

Nah, thsts Otamendi, but with CS's

[Bobby Fetta]

Another article here (https://statsbomb.com/2017/12/on-burnle ... ted-goals/) about Burnley outperforming xGA, which references the Smith NYT piece. Interesting that last season their Goals against minus xGoals against was about minus 8 after 10 games but then stayed steady for the rest of the season. So a similar start to this season, although they've carried on the overperformance for longer this year.

[Stemania]

That's a great article, good find.

[Bobby Fetta]

Cheers Stemania, been lurking for a while but this seemed an easy way to get over the first post hurdle. Found your analysis above really helpful, so wanted to join the discussion.

Personally I love the xG information and generally I think it has helped my decision making this season. Other than getting the Kane captaincy decision wrong almost every week, not owning a Burnley defender has been my biggest rank killer. I guess I have a bias hoping that Burnley significantly regress so I can blame it on luck rather than completely on poor decision making. Will be interesting to continue to monitor Burnley's performance for the rest of the season.

[Stemania]

first post hurdle.

Ha, I was on Tapatalk and didn't notice it was post #1. Welcome (to a new vocal guise). Hopefully we'll see you posting regularly now, the more voices/views the better.

[konstro]

Interesting article about the accuracy of xG. As the author mentions at the beginning we are probably going through chaos stage at the moment, with multiple models available and lack of standardized approach.

http://www.aquestionoftalent.com/data_s ... stigation/

As expected there is a growing market within the game (coaches, managers) that are looking for insights from xG. This makes me think that there will be (or is happening already) a lot of investment available and soon enough, with current technological advances the most superior model will emerge. Already it seems that over the long period of time and even a moderate number of events (i.e.: shots per player) the data is pretty accurate.

Lastly, I recently read about some analytical insights in NBA that changed the way some team play with a significant increase of 3 point shots. I wonder if this means that in few years we will see the decrease of long-range shots in football as those shots success ratio is extremely low.