Other games (2022/23)

[Super Duper Branko Strupar]

51 minutes ago, The College Dropout said:

This is why I call you slow. You're desperate to disagree and be right.

 

It is my opinion that some of the money spent on the LCB position would have been more effectively spent on attacking positions. I think the money spent there has contributed to limiting our attacking options.

 

You disagree. Fine. Move on and stop quoting me.

[Kid Icarus]

Also bear in mind I'm obviously a layman like, so this is all based on me talking with maths and stats academics at work Christmas parties and them doing their version of party tricks 

[The College Dropout]

10 minutes ago, joeyt said:

 

I'm probably being monumentally thick but this doesn't seem right at all

 

If I shoot 4 times from 30 yards I'm not sure how that makes me more likely to score than if I just have one shot from like 18 yards or whatever

It's all relative. A shot from 30 yards likely has a Xg of 0.03 or so. So cumulative it's not that high.

 

Shot volume is a factor to take into consideration. For example, there's this idea that Luis Diaz is a high volume, low quality shooter. As in he takes a lot of shots from difficult to score positions (like he did last night and scored). Cumulative he may end up at 0.6 Xg per match but that's from 5 low chance of conversion shots. Someone like Mane was great at getting on the end of big chances but not a volume shooter. He may end up with a 0.6Xg match but he had a single 0.5% chance and 1 other shot for example.

[triggs]

16 minutes ago, Kid Icarus said:

 

Well no, because cumulative probability and probably are different. If you toss a coin 100 times in a row, the cumulative probability of getting heads at least once is extremely high, but each time you toss the coin the single probably of getting heads isn't then higher, it's reset to 50%

 

Likewise, if I take 100 shots at 0.10 from the halfway line with the 'keeper on the line, xG shouldn't amount to 1.00 just because I tried something that's extremely unlikely 100 times. If xG really is cumulative then that's a flaw that should be taken into account, because it can be a bit deceiving.

 

Liverpool did have half chances, but have a higher xG than Palace who had 2 huge chances.

xG is cumulative probability, that's the reason for it. If you take 50 shots with a value of 2% each, the probability is that 1 of those will go in and that will be reflected in xG. Of course every time you step up to take a shot that particular shot will still have a 2% chance of going in but over the 50 shots 1 should go in. I don't see how that's really a problem in the interpretation of xG?

[Kid Icarus]

22 minutes ago, triggs said:

xG is cumulative probability, that's the reason for it. If you take 50 shots with a value of 2% each, the probability is that 1 of those will go in and that will be reflected in xG. Of course every time you step up to take a shot that particular shot will still have a 2% chance of going in but over the 50 shots 1 should go in. I don't see how that's really a problem in the interpretation of xG?

 

Again, I'm not sure that's the case. My understanding is that it's simply that 50 shots with a 2% chance have been taken, not that 50 shots taken at 2% have a 100% chance of resulting in a goal.  

 

The problem is that xG can be stacked with high volume/low quality chances. Last night Liverpool had 24 lower probability shots and come out with a higher xG than Palace's 7 shots that included a one on one and an open goal. We can disagree about whether that was the case and Palace's chances were higher quality, but being able to stack xG with a higher volume of low quality chances is still a flaw in xG. I don't think it's really a big deal to highlight that, I'm pretty sure it's one of the known issues with it and why things need to pass the eye test etc.  

 

 

Edited August 16, 2022 by Kid Icarus

[Vinny Green Balls]

1 hour ago, KaKa said:

 

It's not a fight man!

TCD is an arrogant, bloviating nob, HTT. He may not be an idiot, but he certainly isn’t nearly as intelligent as he thinks he is.  He has been wrong on so many occasions, but appears to be too much of a pussy to hold his hands up when he is. No wonder so many people have blocked him. Kaka is right to challenge him and his constant stream of horse shit.

[triggs]

4 minutes ago, Kid Icarus said:

 

Again, I'm not sure that's the case. My understanding is that it's simply that 50 shots with a 2% chance have been taken, not that 50 shots taken at 2% have a 100% chance of resulting in a goal.  

 

The problem is that xG can be stacked with high volume/low quality chances. Last night Liverpool had 24 lower probability shots and come out with a higher xG than Palace's 7 shots that included a one on one and an open goal. We can disagree about whether that was the case and Palace's chances were higher quality, but being able to stack xG with a higher volume of low quality chances is still a flaw in xG. I don't think it's really a big deal to highlight that, I'm pretty sure it's one of the known issues with it and why things need to pass the eye test etc.  

 

 

 

If a chance has a 2% chance on xG that means that based on previous data that kind of shot has resulted in a goal 1 in every 50 times on average.

 

I just don't agree that it's much of a flaw to include the actual probabilities of low value shots in the calculation as low value shots often do actually go in. Fwiw both Nunez and Salah's chances in the first half had higher probabilities than the Crystal Palace goal

 

[madras]

12 minutes ago, Kid Icarus said:

 

Again, I'm not sure that's the case. My understanding is that it's simply that 50 shots with a 2% chance have been taken, not that 50 shots taken at 2% have a 100% chance of resulting in a goal.  

 

The problem is that xG can be stacked with high volume/low quality chances. Last night Liverpool had 24 lower probability shots and come out with a higher xG than Palace's 7 shots that included a one on one and an open goal. We can disagree about whether that was the case and Palace's chances were higher quality, but being able to stack xG with a higher volume of low quality chances is still a flaw in xG. I don't think it's really a big deal to highlight that, I'm pretty sure it's one of the known issues with it and why things need to pass the eye test etc.  

 

 

 

Does the probability of a shot change depending on who is doing the shooting ? If not how valuable is the stat ?

[Kid Icarus]

1 minute ago, triggs said:

If a chance has a 2% chance on xG that means that based on previous data that kind of shot has resulted in a goal 1 in every 50 times on average.

 

I just don't agree that it's much of a flaw to include the actual probabilities of low value shots in the calculation as low value shots often do actually go in. Fwiw both Nunez and Salah's chances in the first half had higher probabilities than the Crystal Palace goal

 

 

You don't have to wed yourself to it man, along with plenty of other factors it's a known flaw with xG

[Kid Icarus]

1 minute ago, madras said:

Does the probability of a shot change depending on who is doing the shooting ? If not how valuable is the stat ?

 

It doesn't, it's based on the average shot taker and the average goalkeeper. Leading to players like Messi, or Son in the PL regularly outperforming 'their' xG. 

[triggs]

Just now, madras said:

Does the probability of a shot change depending on who is doing the shooting ? If not how valuable is the stat ?

No. Interestingly there are very few players who actually beat their xG every year, that is to say there are very few consistently good finishers in the world. It's all about getting into the positions to take the shots.

[The College Dropout]

Just now, triggs said:

No. Interestingly there are very few players who actually beat their xG every year, that is to say there are very few consistently good finishers in the world. It's all about getting into the positions to take the shots.

The best players tend to overperform Xg. Messi, Kane, Son etc.

 

Funnily enogh Messi & Kane underperformed their Xg for the first time in a long time last season

[triggs]

Cristiano Ronaldo has finished below his xG 7 of the last 10 seasons

[madras]

Just now, triggs said:

No. Interestingly there are very few players who actually beat their xG every year, that is to say there are very few consistently good finishers in the world. It's all about getting into the positions to take the shots.

I was meaning the probability of scoring a penalty will be higher if its Shearer than if its me for example.

[triggs]

4 minutes ago, Kid Icarus said:

 

You don't have to wed yourself to it man, along with plenty of other factors it's a known flaw with xG

I don't think it overrates the low quality chances. It underrates the hypothetical of a counter attack and the better quality chance that would probably result in but that's impossible to judge

[Rich]

3 minutes ago, madras said:

I was meaning the probability of scoring a penalty will be higher if its Shearer than if its me for example.

 

The base "score" has to be the same for every player, because only then can you infer anything from the data created.

 

If you played 10 seasons in the PL and Shearer played 10 seasons in the PL, the data would clearly show the difference between the two of you in terms of penalties, headers, shots from various distances/locations, etc.

[Hanshithispantz]

I don't think xG is anywhere near accurate enough to be used for 1 game like, it has it's uses but anyone with eyes can see that Zaha should have put the winner away.

[Rich]

The current problem, as I see it, is the lack of granular data that's publicly available for free when it comes to xG.

 

Models have come a long way in the last decade, but, as punters, we don't really get to have a proper look at the really interesting stuff.

 

I was creating models in the industry professionally ~10 years ago that were far, far better than the stuff that's now in the public domain, and gave a lot more context around the data points. We also did a lot of work on " attacking/defensive efficiency", largely due to Leicester City's insane season, and got quite a long way to explaining how they were able to do what they did.

 

At the point I left that role, things had moved onto AI taking over from human analysts so I assume that's where everything is going now at that level. Fully mapping the pitch, assigning precise probabilities of goals being scored to every single square yard or metre based on thousands and thousands of games worth of data from all over the world, and then going back historically to fill in the data related to each and every player and each and every league/competition using video.

 

The problem is always going to be how small the sample size is for such an undertaking at an individual level, though. Say a professional plays ~35 games per season for 10 seasons, on average, you've only got 350 games to get data from at an individual level. Actions might be repeated a few thousand times, in the main, or up to ~50,000 times at the top end, and although that sounds like a lot of data to work with, it's really not.

 

It will get there, though, it's still incredibly early days, relatively speaking. And even now, with what we have available, xG is by far, by far, the most reliable method of judging team quality in the public domain (if not player quality just yet).

[triggs]

6 minutes ago, Hanshithispantz said:

I don't think xG is anywhere near accurate enough to be used for 1 game like, it has it's uses but anyone with eyes can see that Zaha should have put the winner away.

Just as Nunez should have put his chances away.

 

Agree one off games aren't it's ideal use though.

[triggs]

9 minutes ago, Rich said:

The current problem, as I see it, is the lack of granular data that's publicly available for free when it comes to xG.

 

Models have come a long way in the last decade, but, as punters, we don't really get to have a proper look at the really interesting stuff.

 

I was creating models in the industry professionally ~10 years ago that were far, far better than the stuff that's now in the public domain, and gave a lot more context around the data points. We also did a lot of work on " attacking/defensive efficiency", largely due to Leicester City's insane season, and got quite a long way to explaining how they were able to do what they did.

 

At the point I left that role, things had moved onto AI taking over from human analysts so I assume that's where everything is going now at that level. Fully mapping the pitch, assigning precise probabilities of goals being scored to every single square yard or metre based on thousands and thousands of games worth of data from all over the world, and then going back historically to fill in the data related to each and every player and each and every league/competition using video.

 

The problem is always going to be how small the sample size is for such an undertaking at an individual level, though. Say a professional plays ~35 games per season for 10 seasons, on average, you've only got 350 games to get data from at an individual level. Actions might be repeated a few thousand times, in the main, or up to ~50,000 times at the top end, and although that sounds like a lot of data to work with, it's really not.

 

It will get there, though, it's still incredibly early days, relatively speaking. And even now, with what we have available, xG is by far, by far, the most reliable method of judging team quality in the public domain (if not player quality just yet).

Good post.

 

Interesting your caveat on player quality at the end, would have thought it was decent in terms of rating attackers obviously in conjunction with the eye test?

 

Think Brentford used it as part of their Moneyball stuff for signing Scott Hogan and Maupay for peanuts and then making a massive profit. Their models were probably more sophisticated than just the basic xG though.

[Wullie]

Rory Smith's got a book about the use of data in football coming out soon.

 

 

[Rich]

14 minutes ago, triggs said:

I don't think it overrates the low quality chances. It underrates the hypothetical of a counter attack and the better quality chance that would probably result in but that's impossible to judge

 

This is it. And this was the issue back in 2014, and still doesn't appear to have been properly rectified in the public domain.

 

At that time the biggest step forward was adding active attacking/defensive bodies between the ball and the goal (in a sort of cone shape from ball to both posts) to adjust the base level xG for the position on the pitch. I believe we were the only people using this metric back then, and it made our model miles better.

 

From there we introduced "pressure" metrics to add another layer of context in terms of how "free" the player was when taking the shooting action (be it header or shot), which was another step forward in helping to explain how an uncontested shot from 12 yards on a counter-attack was different to a contested shot from the exact same position that might occur on the second phase of a corner kick or other set piece.

 

The thing that tickled me the most was how a completely open goal from inside the six-yard box was nowhere near as close to 1.00 as you would expect. Largely because those situations were quite rare compared to other shots, but also because people missed them with such frightening regularity at all levels.

[Rich]

6 minutes ago, triggs said:

Good post.

 

Interesting your caveat on player quality at the end, would have thought it was decent in terms of rating attackers obviously in conjunction with the eye test?

 

Think Brentford used it as part of their Moneyball stuff for signing Scott Hogan and Maupay for peanuts and then making a massive profit. Their models were probably more sophisticated than just the basic xG though.

 

They absolutely did, aye. I just mean publicly, there's not enough data out there to do so still today for us plebs. The likes of Brentford will have so, so much more data.

 

I was working with Matthew Benham's right-hand man for most of my time being involved in data analysis and models. We basically had Brentford's base model from ~2010 as our starting point, then spent the next few years making tweaks and improvements. SmartOdds (Benham's company) were no doubt doing the same with a team 10x the size, of course . When we got access to more video and when our team was big enough, we built our own model with everything we had learned.

 

We got it to a point where it was really, really good considering the limitations of the time. The amount of data I used to have to wade through was glorious. It's amazing how quick that knowledge goes when you no longer have access to it all, sadly. No regrets at all, but I'd love to know where we'd be at if I hadn't stepped away in 2017.

[Kid Icarus]

5 minutes ago, Rich said:

 

This is it. And this was the issue back in 2014, and still doesn't appear to have been properly rectified in the public domain.

 

At that time the biggest step forward was adding active attacking/defensive bodies between the ball and the goal (in a sort of cone shape from ball to both posts) to adjust the base level xG for the position on the pitch. I believe we were the only people using this metric back then, and it made our model miles better.

 

From there we introduced "pressure" metrics to add another layer of context in terms of how "free" the player was when taking the shooting action (be it header or shot), which was another step forward in helping to explain how an uncontested shot from 12 yards on a counter-attack was different to a contested shot from the exact same position that might occur on the second phase of a corner kick or other set piece.

 

The thing that tickled me the most was how a completely open goal from inside the six-yard box was nowhere near as close to 1.00 as you would expect. Largely because those situations were quite rare compared to other shots, but also because people missed them with such frightening regularity at all levels.

 

This had an xG of 0.3 using Opta's model. 

 

[Rich]

Opta's public model, aye. Not the one they sell to all of the clubs, bookmakers, and professional gamblers for hundreds of thousands of pounds each per season.

 

It was ~0.8 at the time for an average open-goal effort from inside the six-yard box, obviously meaning only 4/5 were scored.