NewCorkSeth SCIAG NewCorkSeth But the statistical likelihood of a shot going in is complete fairy dust. It's not something that can accurately be predicted based on information already existing.
For example the famous Roberto Carlos free kick. That shouldnt by any logic have gone in.
Also goalkeepers should affect how likely a shot is to go in.
It's completely wrong to say you can't measure how likely you are to score from shooting in a given situation.
Just because an event has a low probability doesn't mean it won't happen. Just because an event happens doesn't mean it wasn't unlikely. The reverse is also true.
By any logic I shouldn't win the lottery, but I might. And it's useful to know that I have a one in 56 million chance of winning the national lottery, a one in twenty million chance of winning the Health Lottery, a one in a thousand chance of winning the Golden Gamble, and a one in fifty chance of winning the tombola at the fayre. (All numbers merely illustrative)
And yes, the quality of the goalkeeper and the quality of the striker (as well as other things not currently measured) explain why XG and G are different numbers. But the quality of the goalkeeper doesn't tell you anything about how good the chance is.
That's what XG is claiming to measure - the quantity of chances weighed against the quality of each chance. (XA is better in some respects because it captures occasions where nobody shoots, but I don't have as much faith in the underlying datasets)
How is it completely wrong? It is literally un-quantifiable.
That's simply not true. Lots of people
have quantified it. All you have to do is look at shots which have already happened and see what factors affected whether they go in. Almost everyone who has attempted this has found that the probability of scoring a goal is explained reasonably well by the distance from goal, the angle of the shot, the height of the ball, and how near the defenders are. Some models, including the ones usually used by media organisations, throw in other factors as well. They might not throw in every conceivable factor, but that can be corrected. Because of the chaos you can only generate probabilities, but these are good probabilities because sample sizes are so high and the thing we're looking at is so simple (we're just looking at one shot, not a whole game).
This works in every other area of life. Why do you think football would be any different?
Seriously Seth, your counter-arguments are making the case
for expected goals. You say you want a system that factors in the quality of chances rather than just the sheer number of shots. You say you want a way of telling whether a player is scoring more goals than you'd expect. Those are both things that expected goals metrics look to do. We can tell that Harry Kane is a good striker because he scores more goals than you'd expect given the chances he gets. Factoring in "Harry Kane is a good striker" would just be circular and would result in everyone having the same expected goals stat.
Criticisms of any given model are one thing. Criticisms of the whole concept of modelling are bankrupt.
Snowflake Royal The problem is the probabilities assigned to the shots aren't reliable.
The best models seem to explain about 20% of the variance. There is obviously a lot of room for improvement, but that's very, very good, particularly when you consider that they don't really consider player quality. There was one model that claimed to explain 97% but that was rubbish.
Legitimate criticisms:
- Models aren't transparent enough. People should know what is and isn't included in any given model, as well as the evidence that such a factor influences the expectation of a goal. The BBC have suggested that things like "leaning back" and "when the ball last touched something" are included in their model, but haven't shown any evidence that these things actually impact goal scoring (it's reasonable to assume they do, but we can't say how much instinctively).
- Related: nobody ever reveals the success or failure of their models. At best we might get "XG vs G" at the end of the season, but not "XG vs G for shots with an XG of 0.25"
- There isn't a single objective way of doing it. We can all agree on what constitutes a "shot" but there are models for XG which do or don't include certain factors. There is therefore a public perception that "XG" is a single measure when actually it is a name given to a lot of similar measures. This doesn't actually undermine any given model but it does help confuse people.
Snowflake Royal It's a shit measure. Just because it's the best anyone's come up with doesn't make it not shit.
It's meaningless to talk about the best thing being shit, particularly in a context like this.
The sorts of models used by professional clubs, which aren't shared publicly, will give them information that increases their chances of scoring a goal. If your information is the best information anyone has, and gives you an advantage over your opponents, then the flaws are secondary. Gaining an advantage is more important than perfect information.