Adam87inSA said:
Thanks.
That ranking shows that T-Ranketology is middle of the pack on a five year average performance basis, but is rapidly improving. It was better than average for 2024 and significantly better in 2025.
What I like about that site is that he is updating more than once a day here lately.
Torvik's model is not just Bayesian in season, it's Bayesian over multiple seasons. He feeds the Beta-distribution errors of his model back into the model every year, so the errors become part of next year's model. It's kinda a low-level analog to how AI learns over time by looking at the statistical errors of its own statistical guesses.
In an infinite time period, his model would eventually become perfect statistically. But obviously there will always be upsets, human errors, biases (ESPN hates us!!1!), etc.
It's fascinating as a predictive tool, but it's not designed to be as descriptive. It's the same argument some poster (wacarnolds maybe?) used to have about the RPI vs the shiny new KenPom. KenPom rankings are a predictive model, they don't necessarily translate to the best actual results teams
