Expected Threat (xT) is one of the most important (but least well-understood) statistics in football. ⚽️⚽️
It was used by Liverpool (they call it Goals Added) to do some of their best scouting. 📈
And now, YOU can learn all about it...🧵
The basic idea of xT is to assign a value to a position on the pitch or an action based on the probability it will lead to a goal.
It was proposed (ten years ago) by @srudd_ok who had the idea of using Markov chains to measure how much different parts of the pitch are worth.
This was a mathematically elegant solution and in this video I explain how xT (or probability of a goal) can be calculated step by step.
The name Expected Threat (xT) came form @karun1710, who also described how it works in this amazing interactive blog post. karun.in/blog/expected-…
Lots of different names --- EPV, OBV, G+, VAEP etc. --- have been given to the xT approach.
I distinguish between position-based xT (where place on the pitch is assigned value) and action-based xT (where passes, dribbles etc. are assigned value).
I start by explaining the position-based version of xT. By measuring transitions between different parts of the pitch and the probability of scoring from them, we can iteratively value different places on the pitch.
Then I turn to action-based xT, which better captures the dynamic nature of football.
We can assign value, not just to position, but also how we move between positions.
The action-based version of xT has not been presented in so much detail before, although it is better for professional applications.
How should we measure the performance of a machine learning model? 😕
When I was teaching ML for the first time last year, I was surprised to find there was no agreed upon single number which measures model performance. 🤯
So I decided to look at the question myself... 🧵
Here is the question. Imagine you have created an algorithm which assigns a score to an image based on how likely it is to contain a cat. 🐈⬛
The pictures and the scores from your algorithm might look something like the following.
Now think about setting a threshold for saying 'this is a cat'. For example, in the image above a threshold of r=67 will correctly identify 5 cats (87, 85, 82, 79, 68), but wrongly include 2 non-cats (74, 73) in the images you have labelled as cats. 🐏🐇
In the new spirit of me writing about what I am working on each day, here is a random thread about the 'bespoke football analytics' which we are working on at @twelve_football and featured in the @AnalyticsFC podcast.
The image below is a player radar for Adama Traore from this season in PL before he went to Barcelona. It shows that he is top rated for high-speed dribbles, creating his own chances and maintaining possession. He is also good at cutbacks and through balls.
These kind of stats are important because they move on from simply counting number of actions, to assigning on-ball-value to actions (xT) in terms of ball progression to counting the type of actions a player does and how much value they give (which is what we measure).
It explains why language models like GPT3 are in large part gimmicks: stochastic parrots. They are just generating randomish sequences of sentences from the data put in to them. (1/n)
These can be a bit of fun. But they don't communicate with us: they just repeat random fragments of Reddit, Wikipedia and sensationalised newspaper articles. So it's not clear how they are useful in a wider setting. (2/n)