My xG models

I as re-start my weekly updates of the stats that I post on Tableau I wanted to get around to writing about my different xG models.

I have already written about where I get my data from (although that doesn't include StrataBet data but that isn't in the xG model right now or any of the stuff on Tableau) and wanted to go into depth on how these numbers are derived.

First is that I have 4 different xG models, they each vary slightly and are dependent on the information at hand. They are all derived using a logistic regression based on Premier League Data from 2015-16 and the big 5 European leagues + Champions League from 2016-17.

The first and most used is what I started out calling Chance Quality last season and then gave up and realized that name will never catch on and switched to following everyone else calling it xG. In Tableau this will be labeled xG.

Here is the formula for this:

(1-(1/(1+((e^(-2.5+(Feet*-0.29)+(Head*-0.71)+(Very Close Range*1.28)+(Six Yard Box*0.39)+Center of the Box*0.45)+(Wide Box*-0.24)+(Outside Box*-0.86)+(Long Range*-1.09)+(More than 35 Yards*-1.59)+(Difficult Angle*0.16)+(Set Piece*0.88)+(Direct Free Kick*1.91)+(Corner Kick*0.71)+(Throughball Assist*0.31)+(Cross Assist*-0.25)+(Headed Pass Assist*-0.28)+(Fast Break*1.59)+(Opta Big Chance*1.94)

All of the variables are either 1 for yes or 0 for no. It is in the simplest taking a look at where a shot was taken based on fairly large sized buckets. Was it set or open play, how was it assisted and was it a fast break or big chance. Overall I have found that this pretty accurately reflects the work done by others and gives numbers that are very reasonable.

The next xG model is looking at Shots on Target which uses the same variable as above but it is based on only shots on target (for this it is very simply saved shots or goals no blocks). I created this because I believe that shooting accuracy is a skill (how much of a skill compared to luck or random variation ¯\_(ツ)_/¯) and also I way to get a rough measure of goalkeepers.

The formula for this is:

On Target*(1-(1/(1+((e^(-1.64+(Feet*-0.54)+(Head*-1.23)+(Very Close Range*1.99)+(Six Yard Box*1.13)+Center of the Box*1.15)+(Wide Box*0.45)+(Outside Box*-0.07)+(Long Range*-0.95)+(More than 35 Yards*-1.28)+(Difficult Angle*0.46)+(Set Piece*1.33)+(Direct Free Kick*1.95)+(Corner Kick*1.32)+(Throughball Assist*0.05)+(Cross Assist*-0.004)+(Headed Pass Assist*-0.37)+(Fast Break*2.78)+(Opta Big Chance*1.33)

The next xG is for when I do not have Big Chance data. Collecting the big chance data is a big pain in the ass, it is very helpful for building the model but it is all collected by hand and it isn't always available at the exact moment. There are also competitions that it isn't published and I wanted to be able to do analysis on those competitions as well. This uses the same variables expect for no big chances.

The formula for this is:

(1-(1/(1+((e^(-2.5+(Feet*0.09)+(Head*-0.46)+(Very Close Range*2.58)+(Six Yard Box*1.18)+Center of the Box*084)+(Wide Box*-.32)+(Outside Box*-1.07)+(Long Range*-1.29)+(More than 35 Yards*-1.8)+(Difficult Angle*0.11)+(Set Piece*0.86)+(Direct Free Kick*1.88)+(Corner Kick*0.51)+(Throughball Assist*1.19)+(Cross Assist*-0.33)+(Headed Pass Assist*-0.29)+(Fast Break*1.9)

The last model is the newest one that I have created and I hope to be able to replace it with the very first one (or even combine the two) and this is used to create the xG Chain stat that I just introduced. This is a very rough model with just six variables but it is good enough that I felt okay publishing and continuing to work on it in the mean time.

The formula for this is:

(1-(1/(1+((e^(-8.26+(Feet*-0.54)+(Head*-1.23)+(X Coordinate*0.06)+(Y Coordinate*0.000074)+(Fast Break*0.93)+(Opta Big Chance*2.37)

At the very least I think that I should be changing the coordinates to distance from center of goal and maybe add the angle. We will see how things work out and I am open to suggestions.