02-11-2022, 04:24 PM
(02-11-2022, 04:11 PM)JaggedJimmyJay Wrote: I think this is the problem. I believe this sort of frequentist (or even Bayesian) approach is fundamentally flawed in that it is not (or is barely) data-structure-sensitive, and I believe the entire premise of a model needs to be adapted to one that is minimally- or non-parametric based upon relational structures within the data set not so reliant upon sample size. It's difficult to explain precisely what I mean without writing a dissertation, so oh well.
In any event, I don't think it's possible to accurately project an NFL team's "chance to win". I don't believe probabilities are even applicable.
I'd be interested in hearing what your ideas are. Theoretically, it could be a better approach given that the main issue with NFL analytics is that the sample size (I.E. season) is much more limited than other sports. Sabermetrics is pretty sharp, for instance, because the season is so damn long. The NFL only has 17 games, so it introduces a lot more variance. A team having a 33% chance of victory can still pull out that 33% even if they would theoretically lose a series, if it were a thing.
