By Daragh O'Leary
Former Republic of Ireland manager, Martin O’Neill recently garnered attention for his critique of the expected goal (xG) statistic on TalkSPORT. O’Neill called the xG statistic useless and claimed that it didn’t mean anything. He further validated his claim by saying that Brian Clough, who once famously claimed David Seaman wasn’t a good keeper because he had a ponytail, wouldn’t start players simply because they accumulated xG.
In previous posts I have discussed my somewhat agnostic view of xG. Essentially, I understand the statistic’s usefulness in certain contexts, but I think most of the commentary surrounding xG appears to either misuse or misinterpret what the statistic actually measures.
Empirical economists have a term which they use to describe how well a variable measures a theoretical concept or phenomenon - construct validity. I think the xG measure has a high level of construct validity when it is used to measure individual chances - it’s intended purpose. This can be seen below with the images of two Bruno Fernandes’ goals.
Former Republic of Ireland manager, Martin O’Neill recently garnered attention for his critique of the expected goal (xG) statistic on TalkSPORT. O’Neill called the xG statistic useless and claimed that it didn’t mean anything. He further validated his claim by saying that Brian Clough, who once famously claimed David Seaman wasn’t a good keeper because he had a ponytail, wouldn’t start players simply because they accumulated xG.
In previous posts I have discussed my somewhat agnostic view of xG. Essentially, I understand the statistic’s usefulness in certain contexts, but I think most of the commentary surrounding xG appears to either misuse or misinterpret what the statistic actually measures.
Empirical economists have a term which they use to describe how well a variable measures a theoretical concept or phenomenon - construct validity. I think the xG measure has a high level of construct validity when it is used to measure individual chances - it’s intended purpose. This can be seen below with the images of two Bruno Fernandes’ goals.
 |  |
The image on the left where Bruno faces an open goal from a yard out with the keeper unable to stop the shot measures as an xG of 0.98 i.e. there is a 98% chance a shot from that position should result in a goal. The image on the right, which also ends in a goal, has an xG of 0.02 i.e. 2% chance a shot form that position should result in a goal.
The purpose of the xG measure is to indicate quantitatively that there is a far lower likelihood of a goal coming from the situation on the right than the situation on the left. When used to assess individual chances, I think it’s hard to argue that the xG statistic doesn’t have a decent level of construct validity. xG in that instance informs you broadly as to the goal threat associated with one chance relative to another.
The issue is that most people don’t use or interpret xG in this fashion. xG is commonly reported as an aggregate measure which adds up the total xG associated with each team’s chances throughout an entire game i.e. Team A xG of 1.45 vs Team B xG of 0.75. This doesn’t really make sense.
A team can accumulate an aggregate xG of around 0.75 by being awarded a single penalty (xG = 0.75) or by taking 10 long range shots which each have a very low likelihood of being converted into a goal – i.e. xG = 0.1 + 0.7 + 0.12 + 0.06… Meaning you can accumulate a high aggregate xG without creating a single ‘good’ chance. This means that, when aggregated, xG loses its construct validity because it may be high due to one or two great chances or 10 terrible “chances”.
Common criticisms of the xG statistic, like Martin O’Neill’s, occur because people point out that the aggregate xG stat often doesn’t predict the scoreline. While this is an understandable criticism, it is simply a case of statistics being used for a purpose they are not intended for. xG quantifies the probability that an individual chance will result in a goal. Using the statistic for any other purpose than this results in inappropriate conclusions being drawn from measures that aren’t fit-for-purpose.
The purpose of the xG measure is to indicate quantitatively that there is a far lower likelihood of a goal coming from the situation on the right than the situation on the left. When used to assess individual chances, I think it’s hard to argue that the xG statistic doesn’t have a decent level of construct validity. xG in that instance informs you broadly as to the goal threat associated with one chance relative to another.
The issue is that most people don’t use or interpret xG in this fashion. xG is commonly reported as an aggregate measure which adds up the total xG associated with each team’s chances throughout an entire game i.e. Team A xG of 1.45 vs Team B xG of 0.75. This doesn’t really make sense.
A team can accumulate an aggregate xG of around 0.75 by being awarded a single penalty (xG = 0.75) or by taking 10 long range shots which each have a very low likelihood of being converted into a goal – i.e. xG = 0.1 + 0.7 + 0.12 + 0.06… Meaning you can accumulate a high aggregate xG without creating a single ‘good’ chance. This means that, when aggregated, xG loses its construct validity because it may be high due to one or two great chances or 10 terrible “chances”.
Common criticisms of the xG statistic, like Martin O’Neill’s, occur because people point out that the aggregate xG stat often doesn’t predict the scoreline. While this is an understandable criticism, it is simply a case of statistics being used for a purpose they are not intended for. xG quantifies the probability that an individual chance will result in a goal. Using the statistic for any other purpose than this results in inappropriate conclusions being drawn from measures that aren’t fit-for-purpose.
RSS Feed
