xG - What's it telling me?

Football, Soccer - whatever you call it. It is the beautiful game.
Post Reply
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

If you Saints can still bear to talk about football after the glorious Palace game... ;) 1-1 in h2h though which seems fair enough so I'm not gloating. Palace weren't settling for a draw even against ManCity the other day so when it went 1-1 I had a bet on Palace @ 7.4 with 18min to go as I knew they'd press on. 7.4 just seemed waaaay too big.

Anyway.....
I've been looking at the latest hot topic, xG and have questions, perhaps I've missed something. I'm trying to figure out how to use it.

Is comparing xG to Goals simply a measure of how lucky/unlucky a team was on the day?

An example:
Team A vs Team B : 0-1

Team A xG = 2.6
Team B xG = 0.7

Q1. What would you deduce from that?

But add in some further information:
the xG for Team A's last 3 games was 4.8
the xG for Team B's last 3 games was 0.2

Same question,
Q2. What would you deduce from that?

Did Team A actually underperform on the night? 2.6 vs the usual 4.8 or more? Or did they do well but were unlucky?
Perhaps Team B have the best defence in the league and Team A's 2.6 was in context terrific?

Basically, do we have to compare xG with expected xG (xxG?) to get a measure of how well they performed?

As you can see I'm finding it difficult to see how xG on a given game is helping to gauge a teams performance going forward. Have I missed the step that says it's the xG delta that's important and that strength of opposition defence should be included in the calcs somehow?

I want to be a believer but I'm just not seeing it.

(by the way, are crosses included in xG even if the defender gets to the ball first?
I suppose they are or else 'through the middle' sides taking the odd shot would appear to get more chances than a team who's style was to rain in crosses)
Simonlofc
Posts: 139
Joined: Tue Sep 30, 2014 11:07 am

Until Peter pulled up Mr Stelling about it I hadnt really taken much notice , and I have to admit Im still unsure how to use previous match data on it , Only the top leagues seem to be covered as well. But I have started following games on whoscored.com and it is ver much easier spotting there what is happening, the Wolves game last night appeared to HT to be fairly even but on WS you could clearly see the activity from brentwood was all blocked shots outside the box.Spurs on the heatmap were fairly glowing late SH,The shots West Ham were having were all in the box and close to goal.
https://www.whoscored.com/Matches/11904 ... tal-Palace Thats the match details from the saints game which is informative if you look at the positional report and the tabs there.
But your question is interesting and I havent answered it at all lol
sionascaig
Posts: 1053
Joined: Fri Nov 20, 2015 9:38 am

Hey Shaun, sorry if this is stating the obvious (although I may be missing something), but "expectations" are derived from performance over a statistically significant number of games and if good "measures" their success can only be judged over a statistically significant number of games, e.g. the season. Warning bells always go if in my head when I see discussion on a specific game in that context but you might be getting at another point...

There is a video on youtube somewhere of a Portuguese footy trader giving an insight to the techniques he uses when trading footy (think it covers a previous world cup). A few of the interesting points I took from this:

- he watches the games live on the lowest latency feed he can find
- is fairly selective of games (looks for high scoring ones, using stat expectations)
- generally only trades the 1st 60 minutes
- the actual odds don't seems to matter to him, just whether he feels he is getting value based on what he sees happening on the pitch, i.e. is one team dominating for a specific period (back them in the hope of a goal)

It sounds as if you certainly picked up on the last point which is certainly the hardest )
DoctorEvil
Posts: 16
Joined: Wed Jun 21, 2017 9:07 am

ShaunWhite wrote:
Wed Jan 03, 2018 5:53 am
Anyway.....
I've been looking at the latest hot topic, xG and have questions, perhaps I've missed something. I'm trying to figure out how to use it.

Is comparing xG to Goals simply a measure of how lucky/unlucky a team was on the day?
xG can be turned into odds of winning for home team, draw, away win and some other bets like over/under. a bit of math is required to understand what xG is about. ever heard of Poisson distribution? this is the formula. https://anomaly.io/wp-content/uploads/2 ... ormula.png
Poisson distribution is useful for situations where there are lots of trials and each trial has a relatively low probability of success. in football, one trial is one minute and "success" is if a goal is scored.
"x" in the formula is a whole number value, in football it's goals. "lambda" is a parameter that is equal to Number of trials * probability of success for individual trial. suppose that you read somewhere xG for a team is 2.6. there are 90 minutes (trials) and if 2.6 goals are expected for a team in 90 minutes, probability of a goal happening in a minute is 2.6/90 (probability of success). so, lambda parameter is 90 * 2.6/90 which is just equal to 2.6. so that xG value is actually the lambda parameter.

now, there are 2 teams in a match, each with it's own distribution. multiplying these two distributions would give you the probability for the game to end with a score x:y.

comparing xG to an actual result will have huge variability, a game can finish in many many results, some more likely, some less so.

ShaunWhite wrote:
Wed Jan 03, 2018 5:53 am
An example:
Team A vs Team B : 0-1

Team A xG = 2.6
Team B xG = 0.7

Q1. What would you deduce from that?

But add in some further information:
the xG for Team A's last 3 games was 4.8
the xG for Team B's last 3 games was 0.2

Same question,
Q2. What would you deduce from that?
there's a calculator here. http://sinceawin.com/data/tools/poisson

Q1: underdog won, odds on favourite were 1.28. nothing that was never seen before.

Q2: those numbers maybe aren't realistic because they would hugely favor team A.

ShaunWhite wrote:
Wed Jan 03, 2018 5:53 am

Basically, do we have to compare xG with expected xG (xxG?) to get a measure of how well they performed?

As you can see I'm finding it difficult to see how xG on a given game is helping to gauge a teams performance going forward. Have I missed the step that says it's the xG delta that's important and that strength of opposition defence should be included in the calcs somehow?

I want to be a believer but I'm just not seeing it.

(by the way, are crosses included in xG even if the defender gets to the ball first?
I suppose they are or else 'through the middle' sides taking the odd shot would appear to get more chances than a team who's style was to rain in crosses)
expected expected goals? lol no, one expected is enough. expected goals aren't a measure of how well they performed in your match that ended 0-1, it's a prediction how they WILL perform in that match with all the known info before the match.

xG calculation is basically odds compiling. of course it's important how both strong both teams are compared to one another. some people maybe get confused with xG since they know a game can't finish 2.6-0.7. but if you looked at average goals scored for a team, the number would almost certainly be decimal, not integer. but game scores are integers.
Wolf1877
Posts: 367
Joined: Fri Sep 08, 2017 10:59 am

Lots of seemingly differing opinions on xG so I may as well give my two-penneth.

If you look at something like understat.com it will give the retrospective xG figures for a handful of top leagues including the premier league. Sadly the championship which is the 5th biggest league in europe is not included for some bizarre reason but I won't have to worry about that next season!

So what is understat xG telling us? Basically each attempt at goal (i.e. shot, header whatever) is assigned a probability of success based on an algorithm of where the attempt is taken from and I assume other factors like proximity of defenders etc. The xG totals are then the cumulative sum of the computed probabilities of all goal attempts for each team. In Understats field view you can hover over each attempt to see the assigned probability of each goal attempt. xG is then just a metric that can be used to project forward as it is a measure of the frequency and quality of goal-scoring attempts created and conceded. So it can be used as a metric to guage of the quality of a teams ability to attack and produce quality goalscoring chances AND ALSO of a teams ability to defend and restrict the quality of chances conceded.

There are some glaring failings in xG in my view. So I'm sort of half in Stelling's camp.

For a start off there is the accuracy of the xG model's measure of the goal attempt probability. I think Peter pointed out the extreme long range Wayne Rooney goal for Everton v West Ham failed to take account of the fact that the keeper was 30 yards away from his goal. The Mahrez goal for Leicester v Spurs in November from just outside the penalty area was assigned a 0.04 probability by understat (i.e. odds of 25) which in my opinion vastly understated the probability of a player of Mahrez's quality scoring from that position with the Spurs defenders all standing a couple of yards off him. I'd estimate the probability in that case as somewhere between 0.1 and 0.2 rather than the 0.04 calculated by the understat algorithm.

Another factor that xG fails to recognise that teams tend to up the tempo and attack when they need a goal and tend to defend when they dont need a goal (unless they are completely dominant like Man City). So the timing of goals will greatly affect the xG stats for most teams IMO. So taking last nights Southampton v Palace game. Southampton went ahead on 19 minutes. After that point Southampton's xG barely rose for the rest of the game. I suspect that Southamptons xG would have finished far higher if they had not scored so early as on 19 minutes as they would have likely carried on attacking Palace and accumulating a bigger xG for longer until they actully scored.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

What a lot of long replies, brilliant.
I'll enjoy going through those later when I have time to do so properly.

Biggest issue with xG so far is not being able to search for it in the forum because it's only 2 chars :)
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

DoctorEvil wrote:
Wed Jan 03, 2018 10:14 am
ever heard of Poisson distribution? this is the formula.
Yes I know about it, thx for the info tho
DoctorEvil wrote:
Wed Jan 03, 2018 10:14 am

Q1: underdog won, odds on favourite were 1.28. nothing that was never seen before.
Now this is one of my issues, and I think you've made a mistake that I was making.
You say odds of 1.28 based on the 2.6 & 0.7 xG I put in the example. But the xG figures aren't known until the game had finished (unless I've totally misunderstood). The word 'expected' suggests a prediction, but xG is a retrospecive view. "Shoulda scored" would seem to describe it better.
DoctorEvil wrote:
Wed Jan 03, 2018 10:14 am

Q2: those numbers maybe aren't realistic because they would hugely favor team A.
Unrealistic yes but more to show that if a team has string of high/low xG games, it alters your view of the performance in the last game.
DoctorEvil wrote:
Wed Jan 03, 2018 10:14 am

expected expected goals? lol no, one expected is enough. expected goals aren't a measure of how well they performed in your match that ended 0-1, it's a prediction how they WILL perform in that match with all the known info before the match.
Again, I've never seen xG figures for a match yet to be played, it's historical analysis isn't it? A twist on 'chances created'.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

Wolf1877 wrote:
Wed Jan 03, 2018 11:55 am
Lots of seemingly differing opinions on xG so I may as well give my two-penneth.
I agree with a lot of that Wolf.
I'm always cynical about the latest statistical magic bullet especially when most of the people raving about it have a vested interest, either in selling data/services or maintaining a reputation for being bang on top of all the lastest wizzy trends. xG is a dream for these people beacuse it's complex and the clever data collection isn't possible unless you have a team of people tracking matches and who did what.

Is there even a standard was to calculate xG ?
Who's decided that when Sterling passed the ball last week he'd created a 30.2764% chance of a goal being the outcome ? To even suggest you can calcuate that from even a dozen matches is slightly sillly gven the huge number of variables, pitch condition, opponent strength, his current 'form' etc etc etc. I know these things aren't about 1 game, it's about averages, but sample size and variance have been conveniently ignored.

xG also seems to be designed to produce inherently plausible results. A scale has been invented to say a team should have scored between say 0.5 and 3.5 goals (i don't see many xGs much higher) in a sport where teams usually score between 0 and 4 goals. It's a figure that's been reverse engineered. It's always going to be quite close but give the impression that there's something to be learned from the modest variations.

In PW's xG video he mentions his secret blend of 19 herbs and spices that make up 95% of the flavour of a team. I wonder where xG has been shoehorned into that. Being an xG evangalist, it would be good if he could fill in the blanks for us, I'm obviously not the only one who's yet to be convinced it's as powerful as it's made to seem.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

DoctorEvil wrote:
Wed Jan 03, 2018 10:14 am
xG can be turned into odds of winning for home team, draw, away win and some other bets like over/under. a bit of math is required to understand what xG is about.
I think we've both seen 'Expected Goals' (historic) and mistook it for 'Goal Expectancy' (predictive)
User avatar
MemphisFlash
Posts: 2126
Joined: Fri May 16, 2014 10:12 pm
Location: Leicester

i'll tell you what it is. it's a load of bollocks.
Past performance is no Guarantee of future expectancy!!!
spreadbetting
Posts: 3140
Joined: Sun Jan 31, 2010 8:06 pm

MemphisFlash wrote:
Wed Jan 03, 2018 4:35 pm
i'll tell you what it is. it's a load of bollocks.
Past performance is no Guarantee of future expectancy!!!
Of course nothing's guaranteed but all probabilities are generally derived from past performances especially where you have related data. Saying it's a load of bollocks is like saying we should ignore all horse's form lines when setting odds.

The odds with football are so modelled it's hard to find an edge so every little helps if you can get it to work for you. I used to track actual results against the expected results to see if I could find any trace of teams that were outperforming or underperforming expectations.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

spreadbetting wrote:
Wed Jan 03, 2018 5:08 pm
MemphisFlash wrote:
Wed Jan 03, 2018 4:35 pm
i'll tell you what it is. it's a load of bollocks.
Past performance is no Guarantee of future expectancy!!!
Of course nothing's guaranteed but all probabilities are generally derived from past performances especially where you have related data. Saying it's a load of bollocks is like saying we should ignore all horse's form lines when setting odds.

The odds with football are so modelled it's hard to find an edge so every little helps if you can get it to work for you. I used to track actual results against the expected results to see if I could find any trace of teams that were outperforming or underperforming expectations.
"used to" ? Can I take it from that that you didn't find it helpful?.

I think it's reasonable to start from the assumption that everything is bollocks, until proven otherwise.
phrenetic
Posts: 45
Joined: Sun Oct 16, 2016 5:11 pm

There's an article on Pinnacle about xG at the moment - https://www.pinnacle.com/en/betting-art ... g-formula/
spreadbetting
Posts: 3140
Joined: Sun Jan 31, 2010 8:06 pm

Unfortunately it didn't give me any edge at the time possibly because i didn't put enough effort into it at the time because my coding skills were too limited for the data churning I was looking to do. The thing with football is there's so much data out there that backtesting systems based on snapshots in time aren't neccessarily too hard. I use football mainly for generating commission these days but I do see it as the holy grail because if you can get any small edge the money going through the markets will bring huge rewards.
ShaunWhite wrote:
Wed Jan 03, 2018 5:19 pm


I think it's reasonable to start from the assumption that everything is bollocks, until proven otherwise.
I actually approach all my betting ideas from the opposite angle, I start with something I think should work and the reasons why I think it'll work then test in the real world to see if my presumptions are correct or not. I think it's always best to approach systems etc with preset ideas as I find they're much easier to prove or disprove and disproving your ideas generally reveals the reason why you were wrong to make the assumption in the first place, if that makes sense? I think a lot of traders fail becuase they use a scattergun approach and it's so hard to deduce anything worthwhile if there's too many variables on the go.
User avatar
ShaunWhite
Posts: 9731
Joined: Sat Sep 03, 2016 3:42 am

phrenetic wrote:
Wed Jan 03, 2018 5:35 pm
There's an article on Pinnacle about xG at the moment - https://www.pinnacle.com/en/betting-art ... g-formula/
Didn't see anything very illuminating in that.
He'd created his own xG 'model', with no particular evidence that it was accurate.

"First, use logistic regression to find inconsistencies in the odds, then add in variables (such as expected goals) to see if you can get an edge. This edge will be small, but it could well pay off in the long term."

This is possibly the most vague conculsion I've seen for a while.
"Add in 'variables'" (including a homebrewed xG)
"see IF you can get an edge"
"COULD pay off in the long term"

..but the oddly says it "WILL be small"

All that maths and graphs didn't impress me a great deal.
I see on the other pages that they're another person who uses the word 'model' to describe a process to arrive at a betting system, when a 'model' is actually a precise description of entity relationships from which cause and effect can be established.
Post Reply

Return to “Football trading”