The problem:
I’ve written before about decision-making in sports, and how difficult it is to quantify.
It’s actually something I’ve been thinking about for a long-time, evidenced by this tweet in early 2020.
The Solution:
With 50 or so odd games of 360 at my disposal, I present to you my newest metric:
Ross Barkley or RB’s.
A Ross Barkley is how much xG you gave up by shooting instead of passing to a better positioned teammate.
(xG of Potential Shot * xP of pass) – Actual xG of shot = Ross Barkley’s
I built a passing model with my limited 360 data and combined it with SB’s freeze frames to quantify:
Actual xG = xG of shot
Expected Pass(xP) = Probability of completing a ground pass to teammates in freeze frame given defenders position (built by a sample of 360 data)
Potential xG = pxG if the receiver of potential pass shot immediately upon getting the ball – and assuming no-one else moved in the freeze frame!
More RB’s are obviously not good as it means you are consistently ignoring your better positioned teammates.
Results:
Here are the five biggest RB’s in the first 15 weeks of the EPL:
Stuart Dallas with an extra touch instead of pass to Bamford – this one works although not sure the pass is 80% expected completion by the time he shoots
Actual xG: 22% , Potential xG: 87%, xP: 80%, RBs: 47%
Saka with a bad decision here – Bellerin pass is on – it’s good to see Bellerin’s bad reaction here confirm what we’re looking for – although I think the model wants him to pass it to Laccazette (and that is definitely not a 78% pass)
Actual xG: 6% , Potential xG: 81%, xP: 78%, RBs: 57%
Pretty good example here – Fabio Silva has a tap-in if the cross goes through
Tough pass here for Bowen to Antonio but probably worth it – pass model looks a little too ambitious though!
More frustrated teammates from Grealish here- a great sign!
In general, it looks like the concept works well despite xP being too high for my intuition – they’re all shots in chaotic situations – after a dribble or coming from a first time pass that probably should’ve been passes.
Some musings on what this all means:
A quick look at some aggregates confirms that RBs have an inverse relationship with xG/shot – which makes sense – if you’re already shooting high quality chances you probably don’t have a bunch of better options.
The biggest offender and outlier to this rule?
Jack Grealish – who has combined RB score of 4 xG – meaning he could’ve created 4 more goals if he wasn’t so selfish.
Interestingly Grealish has the highest xG/Shot of all the top offenders – and given what we know about the load he carries (this great piece by Euan on Villa 20/21 is a must-read) perhaps it’s natural that he ends up taking a few shots at the end of high difficulty carries.
The most selfish players:
| Jack Grealish |
| Grady Diangana |
| Mateusz Klich |
| Alex Iwobi |
| Wilfried Zaha |
| Arthur Masuaku |
| Josh Brownhill |
| John Egan |
| James Philip Milner |
| Matty Cash |
On the other side of the spectrum – the only players with a xG/Shot of less than .2 and still making good decisions are Harry Kane, KDB and Firmino – and interesting group of superstar attacking players all known for good decision-making!
The least selfish players:
| Patrick Bamford |
| Mohamed Salah |
| Tammy Abraham |
| Dominic Calvert-Lewin |
| Trent Alexander-Arnold |
| Harry Kane |
| Alexandre Lacazette |
| Kevin De Bruyne |
| Callum Wilson |
| Ollie Watkins |
Apologies to all England fans for that list – Harry Kane really should’ve squared it to Sterling:
Obviously this is all somewhat linked to team style, context, and the models (specifically the pass model) are still pretty weak and can be much improved by people with better Math #skillz.
Also important to note that the good people at StatsBomb I work with will probably produce something far superior to this – this is just me playing around with some data!
Methodology:
The passing model had two features in my efforts to keep it simple.
One was the average distance by all defenders in the frame to the pass path.
The other was the sum of the distances of defenders in ‘relevant distance’ defined as the players who if at top speed (30 km/h) could make it to the interception point of the pass before the pass crossed it.
For xG numbers – I just stole from our great Data Science team their model in API format.
Flaws/future improvements:
The biggest flaw in this model is the assumption that by the time the pass gets there the defenders and keepers won’t move to adjust for the potential shot. Could handle by building some sort of optimization for the defenders to try to minimize the xG of the shot and assume they have a top-speed of 30km/h or so. Quite a challenging problem though because defending for the shot could open up other easy passes.
Passing model is quite weak – the proxies obviously not working very well. It also does not deal with pass height, run-on/through-ball passes at all as some passes are quite easy if you incorporate height or empty field space. It also does not handle offsides. Both are solvable problems that need more thoughtfulness. Voronoi diagrams and exit velocity and angle of pass perhaps a better modelling input.
Assumption of first time shot is a weak one, sometimes the pass is coming from straight behind the potential shooters and so his body position is all wrong. This could be improved by angle of pass or proxy for body position maybe.
Fikha's Blog 
That’s a really clever idea, nice to see stats work that shows an intuitive understanding of playing the game.
LikeLike
Cool post, I think ranking decision making ability of players is probably the hardest thing to do, and this is an interesting way to try to start doing that. 2 minor nitpicks: in your explanation of Ross Barkleys, you say “pass to a better positioned opponent” when I think you mean teammate (passing to opponents sounds like a bad way to create chances), and I think it’s kind of obtuse and confusing to list Tammy Abraham as “Kevin Oghenetega Tamaraebi Bakumo-Abraham” in the chart. Took me a minute to realize who that even was, especially considering he doesn’t use his full name anywhere that I’m aware of. Otherwise, great piece! Hope you can figure out some ways to tweak this even further, would love to see an even more in depth version.
LikeLike
Thanks for the feedback! Fixed both of them – Tammy’s name was just copy pasted off our DB.
LikeLike