I’m going to give it to you straight, this is about as unique and in-depth a formula as I’ve ever seen from FiveThirtyEight. If you aren’t aware, FiveThirtyEight is an ESPN owned sports and politics driven analytics site. It’s basically a home base for number junkies.
Yesterday, they posted an article, showcasing a formula they’ve been using to project upsets, but it’s not your typical algorithm. It was actually inspired by results from image recognition. If you want the dirty details of how it calculates this stuff, then just give it a read,
but basically, they say it is trained to find upsets using team-to-team similarities instead of raw statistics such as offensive rebounds or turnover rates.