Hey, good question Aidon! As I stressed in the post, the model is super simple, and using the Bernoulli distribution was part of that simplicity.

What the model predicted wasn't whether the team would win, lose, or draw (if you'd like to try that, this becomes a multiclass classification problem). Instead, I just wanted to predict whether the team won or not.

So, if the match ended up as a draw, that would be a 0 (not won). The model basically didn't care about draws or losses: they were treated equally.