Activity
Dennis Hackethal posted idea #4953.
This could be a promising approach to formalize HTV:
Frans Zdyb (@FZdyb) on X
@dchackethal I had Claude code up a demo + blog post: https://t.co/LZW3t37WAQ To be sure, this is not my idea, it's a standard tool for Bayesian model selection. Haven't heard anyone connect it to HTV though.
GitHub - deoxyribose/hard_to_vary_posterior_predictive: Exploring the correspondence between posterior predictive evaluated in held out data, and Deutsch's Hard To Vary criterion for explanations
Exploring the correspondence between posterior predictive evaluated in held out data, and Deutsch's Hard To Vary criterion for explanations - deoxyribose/hard_to_vary_posterior_predictive
It’s AI generated, so not eligible for the bounty. And I’m not familiar enough with the probability calculus to evaluate it. But bookmarking it here for the future.
(One of my first criticisms was that HTV has nothing to do with likelihood, which the author granted but addressed by saying it maps onto marginal likelihood. See https://x.com/FZdyb/status/2051004605601898561 and surrounding discussion.)