Extreme Value Theory: The Science of Outliers?

Image Credit: Idyllic,

Extreme events often appear to be outliers and are subsequently overlooked, despite their significance in understanding a given phenomenon. Extreme Value Theory (EVT) provides a framework for analyzing extremes, helping us make sense of a world that isn’t always as well-behaved as the average statistician would like. In this article, we’ll look at EVT through practical examples from my research on methods for studying epilepsy, alongside other intriguing models that capture rare and chaotic behaviors.

On Outliers

When I teach machine learning (ML), we always have a lengthy discussion about outliers.

How can we detect them? Once they’re on our radar, what should we do about them? How did they get into our dataset in the first place?

I am firmly on team outlier in that I believe you should only bench (i.e., remove) outliers in the case of sensor or data entry errors. Outliers are exceedingly useful; for instance, they can often provide insight into whether or not a given machine learning model is overfitting:

Image Credit:

Further, in unbalanced datasets — like those encountered in trying to create fraud…