In his lecture series, @avehtari explained how ELPD works on metric data and made me understand it better than ever. That said, I am still not clear on how ELPD is used to evaluate ordinal predictions such as ratings from 1 to 4.
I am trying to find out how Loo is comparing models in detail and wonder if anyone has some resources on the actual math behind them. For instance, if a model predicts a 42% likelihood of a rating of 3, how would this be compared to the predictions of another model?
I would greatly appreciate any resources or insights that could help me delve deeper into the mathematics behind these concepts. Thank you in advance for your assistance.
Best regards.
Simon