It would be even better if your tool could predict every bet correctly. If everyone could win every bet, then the problem of gambling addicts and the companies that take advantage of them would be quickly solved.

But your solution is cool, too.

Funny enough, this was sort of my north star when developing it. I know it's an impossible dream or expectation (predicting sports is extremely hard), but it's very true that this tool is really only as useful as it is correct.

My first iteration of the baseball model took this to an extreme. I broke down every single step from the pitcher deciding which pitch to throw to the pitcher deciding where to throw it to the pitcher throwing it at a certain speed in a certain spot, and on and on. It was something like 17 sequential PyTorch models. My directive to myself was to not care how much compute it would take to run, just get it to the most right version it could possibly go.

Turns out just predicting the end result of each pitch is far more accurate than sequential modeling. I still to this day have a hard time believing it. But the results are pretty stark when I compare the two methods.

Anyway, I really appreciate the kind words and for taking the time to check it out.