Pitfalls in training and evaluating expected goals (xG) models
Expected goals (xG) measures the quality of a shot attempt in soccer based on several variables such as the shot’s angle and distance from goal, whether it was a headed shot or a free kick, etc. Adding up a player or team’s xG can give an indication of how many goals a player or team should have scored on average, given the shots they have taken. It has become the most used and best understood advanced metric in the world of football, but training an accurate xG model is not as simple as it may seem. In series of blog posts, we discuss the technical aspects of data selection, feature modelling and model evaluation.
HOW DATA AVAILABILITY AFFECTS THE ABILITY TO LEARN GOOD XG MODELS
The available training data can affect the quality of an xG model. This blogpost answers 3 questions:
- How much data is needed to train an xG model?
- Does data go out of date?
- Is there a league-specific effect?
See the full post at https://dtai.cs.kuleuven.be/sports/blog/how-data-availability-affects-the-ability-to-learn-good-xg-models
ILLUSTRATING THE INTERPLAY BETWEEN FEATURES AND MODELS IN XG
The shot’s location is among the most important and predictive features in an xG model. How one should encode this location is not trivial and requires a good understanding of some key ML concepts.
See the full post at https://dtai.cs.kuleuven.be/sports/blog/illustrating-the-interplay-between-features-and-models-in-xg