CroJack
Key Player
Game after game I discover flaws in the expected goals models. These statistical models suck big time, they are not reliable at all, and I am going to prove it by analysing expected goals for QPR - Swansea game. Not only these models steal our expected goals and points, they also unfairly add them to our opposition, which then results in a totally false expected goals table. It seems that the statisticians who make expected goals calculations don't watch football games.
At first glance you can see something's wrong. Experimental361.com gives QPR 1.2 xG, and Infogol.net gives them 1.46 xG. Not a huge difference when you calculate it for one game. But imagine such a difference over the course of a Championship season. 46 games x 0.26 xG difference = 11.96 expected goals. That's a big difference. It's even worse when we look at Swansea's xG for this game. Experimental361.com gives us miserable 0.9 xG, and Infogol.net 1.54 xG. The difference is 0.64 expected goals, which over the course of a Championship season gives the huge 29.44 xG difference. I mean, this is laughable.
I have criticised these statistical models on Twitter, and the only answer I have got was: It is going to even out over the course of a season. They have not been able to explain how it is going to even out. The difference is not insignificant, it's huge. Their excuses are understandable: they actually don't watch the games and chances, all what they do is to retrieve the raw data from Opta and then apply their "statistical models" to finally calculate the expected goals values. But that's completely wrong. The chances in a football game are real, and they can be exactly calculated without flawed "statistical models". Chances are like offsides. You can use technology to exactly calculate expected goals values. Nothing mystical about it. When I get some spare time I'll explain how.
Now let's see how wrong their calculations are, but first we have to distinguish real chances from could've been chances. A real chance is a chance when a player has an opportunity for a clear shot or header. A could've been chance is not a chance but a "what if a player had an opportunity for a clear shot or header". A could've been chance is a wishful thinking. And two of the the could've been QPR chances have been awarded the highest xG values. The one is the Ball's could've been chance in the 38 minute, and the second one is the Bonne's could've been chance in the 61 minute.
Let's see experimental361.com expected goals for QPR v Swansea game. What you can see below is laughable. Where is Ayew's chance
in the Experimental361.com chart?
Ayew was a couple of yards from an open goal when he scored from one of the easiest headers ever. His chance is according to my
calculation worth 0,95 xG, or translated: Ayew had 95 % chance to score there. His chance only is worth more than
experimental361.com has awarded us for all the chances we had.
Now let's see infogol.net expected goals for the same game. Infogol gives a better graphical description of the chances,
with the exact position of the players, the players' names and also the minute when chances occur.
A bigger circle means a bigger chance, and you can click/tap on it to see which player had the chance and you also can
see the xG value of the chance.
At the bottom you can see the game's xG values.
in the Experimental361.com chart?
Ayew was a couple of yards from an open goal when he scored from one of the easiest headers ever. His chance is according to my
calculation worth 0,95 xG, or translated: Ayew had 95 % chance to score there. His chance only is worth more than
experimental361.com has awarded us for all the chances we had.
Now let's see infogol.net expected goals for the same game. Infogol gives a better graphical description of the chances,
with the exact position of the players, the players' names and also the minute when chances occur.
A bigger circle means a bigger chance, and you can click/tap on it to see which player had the chance and you also can
see the xG value of the chance.
At the bottom you can see the game's xG values.
At first glance you can see something's wrong. Experimental361.com gives QPR 1.2 xG, and Infogol.net gives them 1.46 xG. Not a huge difference when you calculate it for one game. But imagine such a difference over the course of a Championship season. 46 games x 0.26 xG difference = 11.96 expected goals. That's a big difference. It's even worse when we look at Swansea's xG for this game. Experimental361.com gives us miserable 0.9 xG, and Infogol.net 1.54 xG. The difference is 0.64 expected goals, which over the course of a Championship season gives the huge 29.44 xG difference. I mean, this is laughable.
I have criticised these statistical models on Twitter, and the only answer I have got was: It is going to even out over the course of a season. They have not been able to explain how it is going to even out. The difference is not insignificant, it's huge. Their excuses are understandable: they actually don't watch the games and chances, all what they do is to retrieve the raw data from Opta and then apply their "statistical models" to finally calculate the expected goals values. But that's completely wrong. The chances in a football game are real, and they can be exactly calculated without flawed "statistical models". Chances are like offsides. You can use technology to exactly calculate expected goals values. Nothing mystical about it. When I get some spare time I'll explain how.
Now let's see how wrong their calculations are, but first we have to distinguish real chances from could've been chances. A real chance is a chance when a player has an opportunity for a clear shot or header. A could've been chance is not a chance but a "what if a player had an opportunity for a clear shot or header". A could've been chance is a wishful thinking. And two of the the could've been QPR chances have been awarded the highest xG values. The one is the Ball's could've been chance in the 38 minute, and the second one is the Bonne's could've been chance in the 61 minute.
And here is why they are not real chances.
Ball's "chance":
Bonne's "chance":
Here is the Bonne's "chance from another angle:
At first glance it looks like a really good chance, but in reality it isn't.
Ironically, the best chance QPR had in this game was Bonne's chance in the 9 minute, and it was rewarded with only 0.1 xG, or 10 % scoring probability.
In reality, the chance deserved roughly 0.5 xG, or 50 % scoring probability. A clear header with only Woodman to beat. This was the chance where Woodman made a great save, and the only good chance QPR had in this game.
Tomorrow evening we'll have a look at our chances against QPR.
Ball's "chance":
Bonne's "chance":
Here is the Bonne's "chance from another angle:
At first glance it looks like a really good chance, but in reality it isn't.
Ironically, the best chance QPR had in this game was Bonne's chance in the 9 minute, and it was rewarded with only 0.1 xG, or 10 % scoring probability.
In reality, the chance deserved roughly 0.5 xG, or 50 % scoring probability. A clear header with only Woodman to beat. This was the chance where Woodman made a great save, and the only good chance QPR had in this game.
Tomorrow evening we'll have a look at our chances against QPR.
Last edited:
