I love that you put a clickable link on the post shot xG.
What can I say...I like leading horses to water
I’m not sure that stat is as much use as xG actually. The Cunha (delicious) cross to Mbuemo that he put over the crossbar is excluded from that as it isn’t on target. As does the one he missed after the shot deflected off the Bournemouth player’s hand. That stat fails the logic test for me. Those were two good chances by any measure.
The Cunha to Mbuemo one was asthetically pleasing no doubt but he was on the stretch, which made it tough to control the shot. That one had an xG of 0.13 which feels about right. You'd expect that to result in a goal about once every 7 or 8 times.
You could just as easily point to the 0.66 xG which Tavernier got when his header was palmed back into his path by the keeper. Considering the way the ball came back at him it was far closer to a 0 xG chance than a 0.66 xG imho.
Both models have their uses though and each have their relative advantages.
As you say, xG considers every shot (no matter how badly its executed) so is probably superior for evidencing &
predicting chance creation.
Post-shot xG includes crucial information (which xG excludes) about what happens at the key moment after a player takes his shot. So it's a better for
explaning why a certain outcome (on a per shot, per match, per season basis) occurred.
A good xG model will never beat a good post-shot xG model (...or even a crap one tbh) in terms of predicting the likelihood of a given shot resulting in a goal.
Also have you come across Akaike Information Criterion?
Didn't he play up top for AFC Wimbledon a few years back?
You do corrections on overly complex models because overly complex models will always fit data better and more complex is bad. If you can convey information with simple stats there’s probably something to it. If not there probably isn’t.
I've come across the concept alright, as a way to compare models and determine which model best fits the data with the least amount of unnecessary complexity.
More complexity in a model isn't inherently bad, but increased complexity
without a sufficient amount of improved prediction accuracy certainly is.
The post-shot xG model is a bit more complex than regular xG, but not massively so. And crucially the types of attributes it considers tend to pass the sniff test
- Where you place the shot
- How hard you hit it
- The shape and trajectory of the shot
- Where the keeper is positioned
etc
These are the types of data points that any data dinosaur would agree are important factors which influence whether a shot is going to result in a goal or not.
If the post-shot xG model was to consider bringing in features like the ones below, then yes you might start to move into model overfitting territory
- Scoreline at the time of the shot
- Weather
etc.
