Evaluating xG
Published:
How to Evaluate Expected Goals Models?
What is this article actually about?
First thing’s first, this article isn’t really about expected goals (xG) models but rather about evaluating and optimizing any model that assigns probability values to discrete events. This could apply to many things. In the context of sports, this article could just as easily be written about game prediction models and outside of sports it applies to any number of things from evaluating a weather channel’s ability to predict whether it will rain to election forecasts. What this article is actually about then is (spoiler alert): log loss. Log loss is a metric that is often used is discussions of hockey analytics for example here it is used to compare game prediction models. In the past when I have seen “log loss”, I have known “small number good” but did not know where it comes from or why I should care about it. Statistics that are not understood can be counter productive and confusing so in this article I hope to provide that understanding. I am going to start off with some statistics basics and from there slowly build up to the most reasonable evaluation method and in doing so derive the meaning of log loss.