Closed vladiki closed 7 years ago
Not yet. This weekend I'm going to sit down and do some back-testing. One can easily leave values out of the known matrix, and see how well they are predicted in the rank-reduced SVD matrix. I can say, however, that I have tested it against the last ~100 or so games, and I've beat the line about 73% of the time so far.
Thanks for the feedback!
That's awesome. I think the consensus is that the naive baseline (predicting home team to win every time) is 60% and state-of-the-art (using vegas line) is ~70%. Would also be interesting to see how this performs not only picking winners, but versus the o/u line.
Oh, sorry. I meant, I've been putting it against the spread (not just picking the winners). Although, I have been accurately predicted a few upsets, which has been nice!
To be honest, the 73% will likely drop. I can't image it preforms that well asymptotically.
Oh in that case 73% is even more impressive. Does this approach extend to more dimensions as opposed to just OR and Pace?
The method (SVD Factorization with rank reduction) is completely blind to what metric you are putting in. Here, I used pace and OR, but there is no reason why whatever incomplete matrix you wanted to look at couldn't be inputed.
Added cross_val.R
Hello
This looks like a really cool project and a novel approach to this problem. Do you have any results to report on how accurate this method is for historical data?