robwhickman/epv

R package to implement various Expected Possession Value models

title	output	pagetitle
README	github_document	README

EPV

R package to implement various Expected Possession Value models

Written by Robert Hickman, October 2019 and January 2020

## What motivates this package

How do we quantify the effect soccer players have

The motivating idea behind Expected Possession Value models of football events is that each action (or 'event') a player makes has some impact upon the expected final result. The simplest and most dramatic example of this is a spectacular long range drive suddenly changing the score differential. For example:

In this example James Rodriguez controls the ball and scores a very low xG chance to give Colombia the lead over Uruguay in the 2014 World Cup Round of 16. If we want to quantify the effect this one movement has had on the game, we can calculate it as the change in score from the event minus the pre-action expected change in score. Let's be charitable and say one out of 10 shots Rodriguez attempts from here would be scored, this resolves to:

xG added: 1 - 0.1 = 0.9

A fairly substantial contribution given the total post-game xG for any side in any football match is likely to be between 0 and 3!

How do we quantify non-shot events

However, most actions during a match are trying to progress the ball into an area where a better shot can be taken from. When in possession this is done by either passing or carrying/dribbling the ball, for example:

Here Robin van Persie scores an excellent diving header to level vs. Spain in the same World Cup. However, while it is a spectacular finish, van Persie has scored a chance we might expect him to finish; given a header in that location, we are perhaps not surprised he scores. For the sake of this example, let's say we give him a 75% chance of scoring (0.75xG).

When Daley Blind picks up the ball on the halfway line however, we might expect a vanishingly small chance that the ball will be in the net within 4 touches of the ball. Let's approximate this as 0. The xG he has added with two controlling touches and a ranging pass is therefore equal to 1 (the goal is fully 1xG) minus the xG added by van Persie (1 - 0.75 = 0.25) minus the xG added by whatever happened before he got the ball (~0 - ~0 ~ 0):

xG added: 1 - (1 - 0.75) - (~0 - ~0) = 0.75

Or when out of possession, by pressing/tackling the opposition to win the ball back and allow a chain of passes/carries to begin.

Bibliography

1. Singh, Introducing Expected Threat (xT), 2019 - the blog post introducing the expected threat (xT) metric and the theory behind it (to be added: Statsbomb conference video)
2. Hickman, Considering Defensive Risk in Expected Threat Models, 2019 - My own paper presented at the 2019 Statsbomb conference playing with some ideas around xT (to be added with talk video)

epv's Issues

implementation?

This isn't really an issue... just wanted to comment and say that I'd love to see the code for all of this. (Your StatsBomb presentation is fascinating!)