Case study: value and context

At Twelve football, we use a method similar to the one implemented in the Python code, in order to value both attacking and defensive play.

The art of taking an expected threat approach is (I believe) using the basic idea of evaluating every action and then thinking about how we can use it in different contexts. It is far to common for companies offering expected threat like solutions to talk about an “AI model of football”. They often try to imply that models based on machine learning are more objective than human judgement, or that their algorithm offers a solution to overall scouting problems.

Such claims are misleading.

It isn’t possible to scout a player using just a single number or output of an AI. For example, some “AI solutions” claim to allow you to transfer in a player in to your squad and tell you the increase of predicted points. This is nothing short of AI snakeoil. Detailed prediction is not possible from machine learning. The reason for this is that machine learning models are really just a simplified representation of the underlying data. We can see this if we think about the stages we have gone through in creating models. We have made a simplfiied assumption about the memorylessness in football and we have used that to help us measure the value of a position on the pitch or of a pass. There is no context or insight from the output of the model. We need to create our own context or insight.

To get something out of these models, we need to work with the output of our models using human insight and judgement.We do need machine learning models to assign values to actions but context remains key to interpreting our results.

In this video I give a few examples illustrating how an Expected Threat model can be used in scouting and performance analysis.