In my experience, you can get extremely good performance in this style of sports prediction with extremely basic methods. E.g. doing some sort of basic Elo rating system + choosing the higher rated competitor to win every game. This is simply because almost all human competitors will frequently try to predict upsets (as its more fun), but it will harm their performance overall.
I would be cool to see if something like this can do can outperform a benchmark Elo system but even that is an extremely high bar that most sophisticated models can't beat.
I would be cool to see if something like this can do can outperform a benchmark Elo system but even that is an extremely high bar that most sophisticated models can't beat.