It's not clear from the article what the input data is. I had a look in the repo, it wasn't clear from that either. Is it the normal fare you'd expect?
I've heard of prop betting firms using stuff like "distance travelled by the away team" and other logical things like that.
We give a high level overview in footnote #4 of the post, but more detail can be found in the code [1]. We tried to pick a relatively small set of features (and kept those picks constant) in order to isolate the model tuning gains from pure feature selection. You can fork the code and try any other interesting features you think would make it better (there is a ton of data we didn't have the chance to look at).
I've heard of prop betting firms using stuff like "distance travelled by the away team" and other logical things like that.