Impact of Game State to my Expected Goals/ExpG model

So, I forgot one thing when I made my ExpG model and posted the results here. I guess this is what is so great about trying to contribute to a community and be open about what one is doing.  11tegen11 gave me some feedback and I got curious too. Game State, the current state of the game regarding score should make an impact, I mean who hasn’t seen a player missing out on some clear opportunities when the game already is decided? Tonight I added game state and this is what happened…

First a reminder of how the fit between actual goals and Expected goals looked with the “old” algorithm, not taking game state into account.

An average fit. Last time I did it the fit it gave me an r2 of 0.63 so the latest rounds of Allsvenskan seem to have had negative impact on the fit. Örebro and Halmstad are underachieving a lot according to this model and are the ones making the fit bad. I guess we´ll see at the end of the season if they have normalized their values.

Now I did the same plot but with Game State, GS taken into account and this happened!

The r2 increased to 0.86, well matching the the model of  Martin Eastwood . So all of a sudden I have a model using a Gradient Boosting Regressor that seem to do just as well as the SVM model Martin Eastwood uses.

And, to answer the question I got on twitter from 11tegen11 I can now list the feature importances as before but now with GS/Game State added:

print clf.feature_importances_
[ 0.22372978 0.23444513 0.01263396 0.00831069 0.04321135 0.1731689
0.04730928 0.25719092]

xpos is valued at 22%
ypos is valued at 23%
finish from corner at 1%
finish from freekick at 1%
finish from penalty at 4%
game time at 17%
possession at 5%
GS 26%

GS is by far the second most important feature behind the shot location.

To sum this post up I can also say that I remade the ExpG plot on Lasse Vibes 2014 season.

I’m not sure what this says. I’ll have to look in to his goals and more exactly look at how game state affects the scoring probability. One theory I have is what I could call the classic “goalscorer syndrome” and that may be what we se on Lasse Vibes increase in ExpG here above. That is: When a game already is settled – a goalscorer keeps on trying to score while the defenders may loose some edge… Just one theory..

Published by

Ola Lidmark Eriksson

Football analyst/programmer https://blog.stryktipsetisistastund.se/

3 thoughts on “Impact of Game State to my Expected Goals/ExpG model”

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s