Currently the quantitative methods for evaluating a players passing are xG build up, xG chains and Passing Progression Value Added. The issue with the first two methods only account for possessions in which a shot is made which means that a players xG chain and xG build up are still dependent on their teammates ability to extend the play after the ball has left the players foot. The main issue I have with Passing Progression Value Added is that it only looks at passes as a change in expected goals value. This method can bring up problems because often a great pass can be something that sets up a goal in that possession. By only looking at shots we limit ourselves from evaluating passes that can set up an attack by only looking at passes that end attacks (via a shot).
My Pass Value Model
To address these issue I’ve devised my own pass value model which attempts to find the goal differential from a pass one can expect a minute after the pass being made.
The strengths of this method compared to the others is that It’s not team dependant because it only needs a pass to come up with a number whereas xG chain requires what happened later in the possession to find a value. The second strength is that it is not focused on shot creation but more the creation of plays that lead to goals.
The time of a minute was chosen because I felt it was long enough that the model starts to evaluate the actual chance created but not too long for the pass to no longer have a significant effect on the play.
The Idea behind this model is that players can often make very valuable passes that allow for the play to develop into a goal scoring opportunity without the pass being directly to a shot or high danger area. The second piece of reasoning is that to win a game a team must score more goals than the other team. So, passes that lead to goals are more valuable than ones that don’t; thus, if we can attach a goal value to a pass we can find what passes actually increase the chance of winning a game.
Using the data from statsbomb.com on the world cup I implemented the model with the following methodology
- find the x,y co-ordinates of every pass, the speed of every pass, and the angle of every pass
- find the goal differential for that team in the following minute and reduce everything to either 1, 0, or -1 (since a pass doesn’t lead to two goals)
- Fit a regression to the model trying to predict the goal differential
- use that trained model to predict the expected value from passes in each game
to solve the issue of training on test data that I had on the expected passes model I used a leave one out technique for every game in the world cup dataset and applied the above methodology for each one using every other game for the training data.
I then added all a players separate pass values together to get an accumulated pass value number for every player
Sorting the players in the world cup here are the following top 5 in pass value.
|Sergio Ramos García||3.06||496||0.00616||Back||4||0.765|
|Jordi Alba Ramos||2.56||414||0.00618||Back||4||0.640|
Below is a map of Pass Value but only using co-ordinates as the inputs to a model
The chart helps visualize how the model works highlighting the best places to pass to when it comes to creating a goal in the coming seconds and the worst places (right in front is -.004) which can often lead to goals against.
The actual model is not far from this but uses three more values for added accuracy.
The model I’ve made presents another understanding of passing and its value. This is important because currently there are only three models in the public that evaluate passing value and all of them are shot-based. My model presents a different perspective and partially answers the question “how much value does the pass add to the possession” a question which hadn’t really had a solid answer before.
The idea for this metric came while I was working on my xPasses model and realized that there are better ways to measure passing than completion rate.