United and xG (now that Ole is gone will things change?)

[Dion]

Apologies if I've missed this, but we should do a quick sense check.

Does the sum of say every premier league player's actual goals less expected goals equal close to zero? If the xg goals figure truly was expected, then over such a large sample size (200 starting outfield players per gameweek, plus say 50 subs, over 38 game weeks = 9500 data points), then the total goals actually scored in the league should be very close to the expected goals.

If it's not, it's a shit metric by its own definition.

In the Premier League this season there have been 814 goals, 810 xG.

Some seasons are downright weird though, the 2016-2017 in La Liga saw 19 teams outperform their xG (the only exception being 9th place Alaves) and 1118 goals from an xG of 952.

 

[OsloRed]

Good teams will more often than not outperform xG both offensively and defensively because of the caliber of players they have, just like bad teams will be on the other end of the scale. Nothing abnormal about it. We are probably overperforming slightly compared to the norm, but nothing completely out of wack. We have good players in good form.

Also, every stat should be taken with a whole lot of grains of salt, especially xG. It can never give us the entire picture.

 

[Revan]

Apologies if I've missed this, but we should do a quick sense check.

Does the sum of say every premier league player's actual goals less expected goals equal close to zero? If the xg goals figure truly was expected, then over such a large sample size (200 starting outfield players per gameweek, plus say 50 subs, over 38 game weeks = 9500 data points), then the total goals actually scored in the league should be very close to the expected goals.

If it's not, it's a shit metric by its own definition.

Teams in EPL are overperforming their xG for less than 0.5 points (in average) if my calculation (in head) is correct.

 

[Dion]

Teams in EPL are overperforming their xG for less than 0.5 points (in average) if my calculation (in head) is correct.

Yep, 0.49% more goals than expected. As I showed in the above example, it isn't always that clean however. La Liga in 16-17 outperformed theirs by 17.4%, that's essentially the whole league doing a Messi. I do wonder if the algorithms that measured that year weren't as accurate as they are now, but that's pure speculation. I guess given a large enough sample size an outlier like that is somewhat inevitable.

 

[Dave89]

Can you please help me then to publish the next paper? I am desperately in need for better results than other competing models.

Simple, tweak your sources and your weightings. Go back a few years more/few years less, add in/remove a few leagues. Include/exclude own goals. It'll take a lot of trial and error, but you'll see a range emerge.

 

[elmo]

If you take a hundred shots from the edge of the box and a little bit to the left you'll maybe see that 10 times out of 100 it goes into the net. That makes the next shot from that position in similar circumstances have an xG of 0.10, ie 0.10 goals expected from every shot in that situation, or one goal scored per ten shots from that position.

What that means is that if a team takes ten of those shots in a game (and creates no other chances) and score two goals, they will have amassed a total xG of 1.0 (0.10x10) while having an actual goals scored value of 2. In that situation, one interpretation is that the players are over-performing and would given enough time regress back to the mean (ie score one goal instead of two), while some others say that it's fairly natural that a team with above average players will outperform the average number.

What this essentially means for us is that we've scored more goals than we would have been expected to based on the amount and quality of chances we've created. If we had scored fewer goals than our xG value, then it would have meant that we'd missed a lot of good goalscoring opportunities (but depending on how big the xG number is, it would have been an indication of how many chances we create as well).

xG is a terribly misunderstood because it's real purpose is to tell you how much better a team is compared to the average but people are misinterpreting it as how often a team should score or concede .

 

[Ban]

It's something RAWK loves.

 

[Escape Goat]

xG is for losers, so they can somehow claim their team played "better" despite having their arses handed to them.

 

[Dave89]

Teams in EPL are overperforming their xG for less than 0.5 points (in average) if my calculation (in head) is correct.

Interesting, if it's getting that level of accuracy for a large sample over a season, without making any changes for observations in-season (i.e. retro-fitting the assumptions to the observations) then after repeating that for a couple of season, we might be on to something.

I've just become professionally jaded with the number of shitty estimates I've been handed in work, with "WE HAVE A MODEL TO SUPPORT THIS". I then spend half a day tearing the model apart, and they still reply "WE HAVE A MODEL!".

 

[Revan]

Simple, tweak your sources and your weightings. Go back a few years more/few years less, add in/remove a few leagues. Include/exclude own goals. It'll take a lot of trial and error, but you'll see a range emerge.

That sound like a clear reject, to be fair.

 

[Revan]

xG is for losers, so they can somehow claim their team played "better" despite having their arses handed to them.

Which is squawka's team?

 

[Dave89]

That sound like a clear reject, to be fair.

It is, it's completely missing the point of modelling, and is unethical, but it was an illustration of why I distrust these sorts of estimates until proven otherwise.

 

[SirAF]

oleG

 

[Revan]

It is, it's completely missing the point of modelling, and is unethical, but it was an illustration of why I distrust these sorts of estimates until proven otherwise.

To be fair it is predicting a very unbiased estimate, though the variance seem to be high (top teams outperform their xG, bottom teams underperform theirs) which might have a clear explanation that top teams have better players than bottom teams. If the same chance goes to Aguero and Kachunga there is a higher chance that Aguero will score it.

xG doesn't consider the quality of the players. Would be very interesting if someone comes with a model which adjusts xG to the player (as in, if you are outperforming your xG, then xG for you needs to be adjusted), however it would be a nightmare to do so.

Btw, I have no idea what types of models squawka is using.

Edit: Apparently 538's model does exactly that, each player has a weight which gets multiplied by his xG. In addition they add xG for non-shots (for example Smalling vs Liverpool which was a great chance but doesn't count as xG cause he didn't shoot).

 

[Adisa]

I'm probably being a bit self-indulgent but the Ole thread is so busy this specific topic might get lost. Why are xG painting such a grim picture of his time in charge? Is it unusual to have such a massive disconnect between xG and the actual points on the table? Is this is all just a complete load of bollox? Does it tell us anything useful at all?

Tweet

— Twitter API (@user) date

An enormous amount of shots we concede are from outside the area. The xG thing is doing my head in.

 

[cyberman]

xG is for losers, so they can somehow claim their team played "better" despite having their arses handed to them.

Spot on. A lot of sides also play for the keep it tight and get one chance for their top striker approach which would skew xg as an indicator.
PSG last night also played to keep the ball. They controlled the game but didn't really drive forward. Does that mean United were the bigger threat? Of course not

 

[Revan]

Btw, a very good post (in layman's terms) about xG for people who don't know much about it, but are not close minded.

https://www.reddit.com/r/soccer/comments/8a8mhp/oc_an_introductory_guide_to_the_xg_statistic_the/

 

[Aarron Swift]

The entire expected goals concede figure is probably from the Spurs game.

Not that I really Understand how the figure is calculated, Looks at the goals against Cardiff, Spurs, Southampton ....there are some long range worldies in there along that you wouldn’t expect to have a high % chance of going crazy, along with some that keeper maybe should have saved

Add to that the fact that we have played deep, & looking to counter with pace in a a fair number of Oles away games because of how tough those games have been, our expected goals is probably relatively low

 

[Valencia Shin Crosses]

I just don't know how you can do a meaningful statistical analysis for xG to be useful considering the sample sizes are always incredibly small in the grand scheme of numbers. Plus I feel like football is such a game of variables even within a single chance, that it is too difficult to account for all of them while giving some kind of meaningful number.

 

[Ramshock]

Can xG just feck off forever nobody gives a shit

This, its the sort of wankery that the scousers embrace to make up for not winning anything, just like the alternative league table

 

[syrian_scholes]

First of all if you think xG is shit then jog on, this post isn't for you.

We've seen results go against xG for a season, but from my recollection, results always regress to the mean. 2 examples come to mind:

1. Guardiola's first season, their results were worse than xG would suggest. Based on that, some people predicted they would win the league the following season, they did, by a mile.
2. The season that we finished second, we conceded less goals than xG suggested. DDG had an outstanding year of course. Was it any surprise that results dipped the following season?

If the numbers are saying we should be conceding more and scoring less, then it's worth looking into, and making the correct adjustments, before next season. Regardless, Ole has done well enough to be given a chance to look into making sure this run of form isn't a blip.

Or pep signed 5 wc players that took them to the next level while we only signed Fred.

 

[Dyslexic Untied]

So at what point of a "chance" is this assessed? Is it when the shot is taken? Or when it could be taken? Take yesterday as an example.

A penalty (3rd), a tap in from a rebound (2nd) and an open goal from a slight angle (1st). Should be three clear goals really, so somewhere around 2,5 - 3,0 I would assume. Or do you take into account Rashford's shot prior to our second goal (which probably had an expected goal ratio of close to 0). Or do you consider the expected goal ratio of our first goal when Lukaku first receives it, with Buffon between him and the goal and Silva slightly to the side of him?

 

[Revan]

This, its the sort of wankery that the scousers embrace to make up for not winning anything, just like the alternative league table

It is an extremely idiotic comment. The alternative league table converges to the real table at the end of the season, by definition. How it can be used to feel better when the final tables are in the end the same?

 

[Treble]

Statistical methods are useful but have various limitations. To dismiss them entirely is unreasonable though.

xG does not have a great predictive value over several months or even a season. In 16/17 Chelsea finished 15 pts above City but according to xG City should have comfortably won the league. Obviously, there was something statistically intangible that helped Chelsea massively outperform City pts wise.

BUT xG seem to have some predictive value long term. It's very difficult to consistently overachieve and it's natural that results tend to get closer to xG stats as was the case with Chelsea and City in 17/18 or as is the case with Spurs this season who were massively overachieving over the first 25 games but won just 1 point in the last 3 games and the difference between them and City/Liverpool is starting to reflect reality as represented by xG.

It's obvious that we are overachieving under Ole points wise. Why? Well, we are winning 2,67 pts per game which would take us to 101 pts over 38 games. Is the team really that good already? Most probably not. Results will probably get a bit worse over the next 2 months. Huge improvements in the squad and the performances will be needed to maintain the current level of 2,67 PPG.

 

[adexkola]

Or pep signed 5 wc players that took them to the next level while we only signed Fred.

Of those players you mention, only Walker and Ederson featured heavily in their league triumph that season.

And no, I don't think our dip was due to "just bringing in Fred". DDG's drop in performance from godlike levels alone left us in the red.

 

[Revan]

So at what point of a "chance" is this assessed? Is it when the shot is taken? Or when it could be taken? Take yesterday as an example.

A penalty (3rd), a tap in from a rebound (2nd) and an open goal from a slight angle (1st). Should be three clear goals really, so somewhere around 2,5 - 3,0 I would assume. Or do you take into account Rashford's shot prior to our second goal (which probably had an expected goal ratio of close to 0). Or do you consider the expected goal ratio of our first goal when Lukaku first receives it, with Buffon between him and the goal and Silva slightly to the side of him?

I guess 0.8 for pen (though some models seem to not take pens into considerations and have it an extra stat), 0.7 for Lukaku's goal (1 on 1 chance), 0.1 for Rashford's shot, 0.7 for Lukaku's second goal, 0.1 for Dalot's. So, it should be around 1.6xG + a penalty, or a 2.4xG in total.

Of course, this is slightly simplifying things.

 

[kundalini]

If I understand that correctly, they've confirmed that, working backwards, the models fit the source data used (which you would expect), but for the first full season of new predictions, did it fit?

There are numerous different xG models and they cover the big 5 European Leagues. The understat model seems to be a good fit for some European leagues some seasons, but from time to time, the total number of goals scored by all the teams in a League (say Serie A or La Liga) will be noticeably higher or lower than the predicted xG figure.

There are various graphs on the home page which show the results, https://understat.com/ simply change the league, slightly above the graph on the left.

 

[TMDaines]

So at what point of a "chance" is this assessed? Is it when the shot is taken? Or when it could be taken? Take yesterday as an example.

A penalty (3rd), a tap in from a rebound (2nd) and an open goal from a slight angle (1st). Should be three clear goals really, so somewhere around 2,5 - 3,0 I would assume. Or do you take into account Rashford's shot prior to our second goal (which probably had an expected goal ratio of close to 0). Or do you consider the expected goal ratio of our first goal when Lukaku first receives it, with Buffon between him and the goal and Silva slightly to the side of him?

Depends on the model, but most do only the best chance in a passage of play, which potentially undersells the likelihood of scoring, as it doesn't count for all the branches of the probability tree, which end in GOAL. So the 2nd Lukaku goal last night has slightly better xG than just the Lukaku tap in suggests, as the Rashford shot could have just gone straight in.

Also virtually all models can't account for non-shot xG. So a killer cross across the face of a goal, which a striker narrowly fails to make contact with, goes down as 0 xG, when actually that passage of play has some expected goals value.

xG isn't perfect, but as models evolve, it increasingly is one of the most insightful stats we have for assessing how well a team attacked or defended, based on the chances they had or conceded.

 

[Dyslexic Untied]

Depends on the model, but most do only the best chance in a passage of play, which potentially undersells the likelihood of scoring, as it doesn't count for all the branches of the probability tree, which end in GOAL. So the 2nd Lukaku goal last night has slightly better xG than just the Lukaku tap in suggests, as the Rashford shot could have just gone straight in.

Also virtually all models can't account for non-shot xG. So a killer cross across the face of a goal, which a striker narrowly fails to make contact with, goes down as 0 xG, when actually that passage of play has some expected goals value.

xG isn't perfect, but as models evolve, it increasingly is one of the most insightful stats we have for assessing how well a team attacked or defended, based on the chances they had or conceded.

Thank you and @Revan . I have nothing against xG btw, just curious as to what they feed into the model.

 

[Maticmaker]

This was my question to. I havent seen an explanation so I looked it up.



https://www.bbc.co.uk/sport/football/40699431

The factors taken into account when assessing the quality of a chance include:

• Distance from goal
• Angle of the shot
• Did the chance fall at the player's feet or was it a header?
• Was it a one on one?
• What was the assist like? (eg long ball, cross, through ball, pull-back)
• In what passage of play did it happen? (eg open play, direct free-kick, corner kick)
• Has the player just beaten an opponent?
• Is it a rebound?
• DID HE SCORE? (only stat that counts surely?)

 

[syrian_scholes]

Of those players you mention, only Walker and Ederson featured heavily in their league triumph that season.

And no, I don't think our dip was due to "just bringing in Fred". DDG's drop in performance from godlike levels alone left us in the red.

And players broken relationship with the manager and bad morale definitely affected morale, it's not as black and white "as going back to the mean", so many factors affect football results which makes xg unreliable imo.

 

[Revan]

Depends on the model, but most do only the best chance in a passage of play, which potentially undersells the likelihood of scoring, as it doesn't count for all the branches of the probability tree, which end in GOAL. So the 2nd Lukaku goal last night has slightly better xG than just the Lukaku tap in suggests, as the Rashford shot could have just gone straight in.

Also virtually all models can't account for non-shot xG. So a killer cross across the face of a goal, which a striker narrowly fails to make contact with, goes down as 0 xG, when actually that passage of play has some expected goals value.

xG isn't perfect, but as models evolve, it increasingly is one of the most insightful stats we have for assessing how well a team attacked or defended, based on the chances they had or conceded.

As I said before, 538's model takes that into consideration as it takes the quality of players who take the shoots (call it adjusted xG). At the moment, we are sixth there with only 34% chance of reaching UCL zone. However, the model doesn't take into account Mourinho and Ole era. Would be interesting to see how it is only in Ole era. For comparison, while now we have 34% chances of qualifying for UCL, when Ole took over the model was giving us only 5% chance.

538 have been very good at other types of predictions, so probably they are good here too.

 

[Bocca9978]

Can xG just feck off forever nobody gives a shit

This. It’s very Liverpool.

 

[bucky]

None of that really addresses the issue that Messi is far from the only player who consistently outperforms his xG. It's actually very common among forwards at top clubs.

They are supposed to. That's how the model tells you, who the good players are. It's an average that takes into account bad, average and good players.

 

[Ramshock]

It is an extremely idiotic comment. The alternative league table converges to the real table at the end of the season, by definition. How it can be used to feel better when the final tables are in the end the same?

Have you heard of Scousers before?

 

[Revan]

Have you heard of Scousers before?

It doesn't matter. It is a legend that Caf created which was funny, but it was never the truth. The alternative EPL table in RAWK never had Liverpool winning the league, because you know, it is impossible to win the league in AEPL table if you don't win the real EPL. But now everything that requires slightly more thinking gets labelled as bullshit despite that there is a mountain of evidence that it is far from bullshit.

I also have some news for you. Autonomous driving cars are better than you at driving, a computer can defeat you in both Go and Chess, while the same computer can defeat one on one every player in the world in DOTA. Bollocks!

 

[Ramshock]

It doesn't matter. It is a legend that Caf created which was funny, but it was never the truth. The alternative EPL table in RAWK never had Liverpool winning the league, because you know, it is impossible to win the league in AEPL table if you don't win the real EPL. But now everything that requires slightly more thinking gets labelled as bullshit despite that there is a mountain of evidence that it is far from bullshit.

I also have some news for you. Autonomous driving cars are better than you at driving, a computer can defeat you in both Go and Chess, while the same computer can defeat one on one every player in the world in DOTA. Bollocks!

Found Michael Owen

 

[Revan]

Found Michael Owen

Found someone who failed in high school.

 

[KM]

while the same computer can defeat one on one every player in the world in DOTA. Bollocks!

Off topic but did the AI really beat Sumail and Miracle? Find it hard to believe.

 

[bucky]

I'm probably being a bit self-indulgent but the Ole thread is so busy this specific topic might get lost. Why are xG painting such a grim picture of his time in charge? Is it unusual to have such a massive disconnect between xG and the actual points on the table? Is this is all just a complete load of bollox? Does it tell us anything useful at all?

Tweet

— Twitter API (@user) date

Chelsea won their title in 2016/17 while massively overperforming in that model (according to understat they had 61.80 xG and actually scored 85). It's sustainable over a certain amount of time, but it was indicating that they shouldn't be able to keep this up, which is what happened. If we properly back Ole during the summer and address our weaknesses, then all this regression oppos are hoping for doesn't necessarily need to happen.