I quite like that xG ignores the shot taker. That means it can be used to assess the real quality of the chances.
If your style of play regularly creates seemingly gilt edged chances but they fall to a CB, you'll want a CB who can finish, or you'll realise you need to tweak some patterns to get a more natural finisher into those positions.
If your chances are falling to your striker but they're underperforming on xG, it tells you a better striker is needed.
If xG took the player or the position of the player into account, you'd only be measuring them against their own metric, meaning Rasmus's average would look just as good as Ronaldo's. Lowering the expectation based on the players lowers the accepted standard too.
I’ve been reading quite a few stats related articles lately, and you highlighted the most commonly misunderstood thing about xG - it’s a measure of chance, not quality of finish.
And for the latter, that’s why we use xGOT - expected goal on target.
But I’m sure you know that already. So I’m just gonna look forward to your next article!
I was playing around with xGOT (or post-shot xG) calculations for part III looking into measuring how good players turn their chances into - or how much they squander their chances, but ultimately I cut it out because it didn't flow right and didn't really fit
I want to start by saying that I did enjoy reading this piece. That being said, the one thing that bugged me about it is the "xG doesn't account for who's taking the shot" being a problem, - to me, that's part of the point. Stripping away any extraneous information about the player, team, time, and so on, and providing a base value. I also think that using the idea of the same player taking a shot 30 times for the cumulative number of goals is a bit misleading - in practice, it's more like 30 random players taking that shot, or 30 players taking that shot 30 times.
However, I still do like the overall point being raised here. I think xG is designed to be contextualised, and looking at it by itself doesn't really tell you much. That's where the analyst comes in. Effectively, you feed the computer raw data, it tells you what should've happened, and then you have to work out why it did or didn't. Maybe one particularly egregious miss skewed the values. Maybe the keeper played out of his mind. Or - as you point out in part II - maybe good chances are going to players with poor finishing.
Again, I want to make it clear that I enjoyed reading this article, and that I thought it raised good points, and asked the right questions. I'm looking forward to reading the rest of the series. Just a small but important bit of framing - xG is the raw value, and it's up to you to add context to it.
Just want to be clear in case it wasn't. I don't believe "xG doesn't account for who's taking the shot" is a problem with xG. It merely can be a problem with single game xG.
I loved the game-state angle which forced United to take lower xG shots at 0.09. I recently wrote (and still preparing more) deep dives on xG and xGOT, and for the one that is coming out this Thursday I calculated that the average xG/shot at Europe's top 5 leagues for the past 11 seasons stood at 0.11 (data by Understat).
And a great point for who's really taking those chances. This got me interested, and indeed, Leny took a total of 4 shots at United, 3 of which accumulated 0.45 xG but went off (so 0 xGOT, poor execution) and the fourth one is the one you mentioned at 0.3 xG with 0.19 xGOT again poor execution (we judge striker's ability by xGOT - xG).
So this got me thinking, why aren't providers creating xG models for strikers, midfielders, defenders? I know we have xGOT to take care of that, but having a well calibrated xG model that takes account of the player's fundamental role could be valuable (xG is about chance creation, not pure execution) and will deal with this noisy data. Maybe there's something I am missing, I should explore this more.
I quite like that xG ignores the shot taker. That means it can be used to assess the real quality of the chances.
If your style of play regularly creates seemingly gilt edged chances but they fall to a CB, you'll want a CB who can finish, or you'll realise you need to tweak some patterns to get a more natural finisher into those positions.
If your chances are falling to your striker but they're underperforming on xG, it tells you a better striker is needed.
If xG took the player or the position of the player into account, you'd only be measuring them against their own metric, meaning Rasmus's average would look just as good as Ronaldo's. Lowering the expectation based on the players lowers the accepted standard too.
This is exactly what Part II is about
Looking forward to it
Thanks Pauly for your read. Love it as always.
I’ve been reading quite a few stats related articles lately, and you highlighted the most commonly misunderstood thing about xG - it’s a measure of chance, not quality of finish.
And for the latter, that’s why we use xGOT - expected goal on target.
But I’m sure you know that already. So I’m just gonna look forward to your next article!
I was playing around with xGOT (or post-shot xG) calculations for part III looking into measuring how good players turn their chances into - or how much they squander their chances, but ultimately I cut it out because it didn't flow right and didn't really fit
I want to start by saying that I did enjoy reading this piece. That being said, the one thing that bugged me about it is the "xG doesn't account for who's taking the shot" being a problem, - to me, that's part of the point. Stripping away any extraneous information about the player, team, time, and so on, and providing a base value. I also think that using the idea of the same player taking a shot 30 times for the cumulative number of goals is a bit misleading - in practice, it's more like 30 random players taking that shot, or 30 players taking that shot 30 times.
However, I still do like the overall point being raised here. I think xG is designed to be contextualised, and looking at it by itself doesn't really tell you much. That's where the analyst comes in. Effectively, you feed the computer raw data, it tells you what should've happened, and then you have to work out why it did or didn't. Maybe one particularly egregious miss skewed the values. Maybe the keeper played out of his mind. Or - as you point out in part II - maybe good chances are going to players with poor finishing.
Again, I want to make it clear that I enjoyed reading this article, and that I thought it raised good points, and asked the right questions. I'm looking forward to reading the rest of the series. Just a small but important bit of framing - xG is the raw value, and it's up to you to add context to it.
Thanks Dante
Just want to be clear in case it wasn't. I don't believe "xG doesn't account for who's taking the shot" is a problem with xG. It merely can be a problem with single game xG.
Really good point! Well written piece
A great break down of xG blind spots, Pauly 👏
I loved the game-state angle which forced United to take lower xG shots at 0.09. I recently wrote (and still preparing more) deep dives on xG and xGOT, and for the one that is coming out this Thursday I calculated that the average xG/shot at Europe's top 5 leagues for the past 11 seasons stood at 0.11 (data by Understat).
And a great point for who's really taking those chances. This got me interested, and indeed, Leny took a total of 4 shots at United, 3 of which accumulated 0.45 xG but went off (so 0 xGOT, poor execution) and the fourth one is the one you mentioned at 0.3 xG with 0.19 xGOT again poor execution (we judge striker's ability by xGOT - xG).
So this got me thinking, why aren't providers creating xG models for strikers, midfielders, defenders? I know we have xGOT to take care of that, but having a well calibrated xG model that takes account of the player's fundamental role could be valuable (xG is about chance creation, not pure execution) and will deal with this noisy data. Maybe there's something I am missing, I should explore this more.
Anyway, loved the piece!