Read on Twitter
An Unexpected Result - Finishing Skill 
Here follows a brief discussion about "finishing skill"
, what it means & how it's used, followed by a short experiment 
.If you think you already know all about it, feel free to skip to the end
!Otherwise, continue
...
Many models combine past xG (expected goals) data with a player's "finishing skill" 
.In theory, a player's xG tells you how many goals their shots were "worth", on average. It considers factors up to the shot, e.g. distance, height, angle & DEF/GK positions
.
The basic idea behind "finishing skill" is some players are better at shooting than others.Therefore such players should get more goals than an average player from the same quality chances
.For example,
Messi (201G from 165.7xG) vs 
Benteke (43G from 55.3xG) ...
So for modelling purposes, you might take a player's past number of goals and divide it by their past xG 
. This gives you the number of goals that that player gets per 1xG.
For example, in games on http://understat.com , Messi has scored 201G from 165.7xG. This gives
...
. This gives you the number of goals that that player gets per 1xG.For example, in games on http://understat.com , Messi has scored 201G from 165.7xG. This gives
...
201 ÷ 165.7 = 1.21 So you'd say Messi's "finishing skill" (at least historically) is 1.21 goals per 1xG, or +21%
.Then if you expected Messi to get 25xG this season, you could apply his "finishing skill"
to see how many goals he might get:25 × 1.21 = 30.25
.
As I said, many models apply this idea. However, @FPLRoosta's reluctance to follow suit in such an extreme way ( https://twitter.com/FPLRoosta/status/1307302307042910209?s=19) inspired me to take a closer look at some of the data 
.
Remember this will only be a small test, further research is encouraged!
The Experiment What we want to do in FPL models is predict a player's season using past data
. Hence I'll look at past seasons & see how predictive earlier data was for each of those seasons 
.I'll compare plain xG to xG with a basic "finishing skill" adjustment
.
As I only want a quick litmus test, I'll restrict to seasons where a player had over 10xG & they had over 25xG in previous seasons (from which to determine "finishing skill") 
.My source will be the top 5 leagues on http://understat.com (data from 14/15 onwards)
.
For example, Zlatan's 15/16 season below wouldn't be counted as he previously only had 18.95xG (less than 25, not enough to determine "finishing skill" 
).Also his 17/18 season wouldn't count as he only achieved 0.3xG (less than 10)
.So we count only 16/17 & 19/20
.
I identified 165 seasons from a total of 79 players. I calculated their previous "finishing skill" in each case & applied it to that season's xG to achieve a "theory G". The idea is to see which of xG and "theory G" is closer to actual goals 
.Sample data
:
Here I'll use the Pearson correlation coefficient to find the correlation between (Goals) & xG and the correlation between (Goals) & (xG × past "finishing skill"). Excel squares this to give the r^2 value, a number between 0 and 1. 0 is no correlation, 1 is perfect 
.
Results Surprisingly to me, plain xG with no "finishing skill" adjustment has a higher r^2, implying a greater correlation
!This small experiment implies that it's actually more predictive (on average) to ignore "finishing skill" altogether
!What now
?
Honestly this was so surprising to me that I'm not sure what's next. I would love ideas for a deeper investigation that could provide more definitive results .For now I'll just thank the analytics community & especially @FPLRoosta for putting me onto it
!Kiwi out
.
