This post on StatsBomb (https://statsbomb.com/2015/03/arsenal-score-effects-and-a-season-of-two-halves/) got me thinking on game states. I’ve seen similar stuff before and it makes total sense, however it seems as if many ExpG models don’t really take it into effect. Teams have different strategies for being ahead and behind and often visibly play different, so it seems like any attempt to model the game should take that into account. So I looked through the EPL data to see what effects Game State has.
To start here is the breakdown for how often a certain game state comes up:
And then I looked at how shot rate varies with each of those game states.
So the average team takes just over 13.5 shots per 90 in a tie game. When they go ahead by one, that number drops under 13.5. When they fall behind the number jumps near 15 per game. When the lead stretches to 2 and then 3 or more goals, teams fire more and more shots (I am guessing because this often happens in routs where a great team is facing a poor team).
This finding doesn’t surprise many people. Next let’s look at how the quality of shot varies with game state:
As a team moves from behind to tied to into the league, you can see the average quality of their shot* goes up generally and slightly (though shot quality actually is better down 1 than tied).
*Shot quality is determined by angle from goal, distance from goal, and whether the shot is a header or not
Now lets add in the actual result of those shots:
This is where we can see ExpG breaking down. Teams in the lead are converting shots at rates well above what we would expect based on where the shots come from and what body part they are taken with (I am guessing due to decreased defensive pressure from the losing team going for a goal) and teams that are behind are converting well below what you would expect (likely for the opposite reason). This makes gut level sense for anyone who has looked at shot charts from a game their team lost 2-0 to Chelsea where they had 13 shots against the Blues 11. Combining the shot rate with the ExpG/S and G/S shot charts gives us this one:
And shows just how far off ExpG can be when it doesn’t factor in game states. Up next I will look at the other leagues to determine if this carries over and then adjust the ExpG model to factor in Game State.
EDIT: Added in the other 4 major European leagues and the trend holds across them all. Here is the chart with them all together:
and the shots/90 at each game state: