Pointwise predictive accuracy about celerite2

[dfm]

It's not entirely trivial to do cross validation for GP models since every data point "cares" about whether or not every other data point is included. There isn't really any such thing as a "pointwise" logp for a GP! You can take a look at section 5.4.2 of R&W for some general discussion of how you can compute cross validation scores for GPs, but it's not completely obvious that you can do this calculation in a scalable way, or that it would slot into arviz the way you want.

[bmorris3]

Thanks for that R&W section, I'm going to read it carefully. I agree that it probably won't translate into a few-line arviz call, but I was wondering if it would be reasonable to do this calculation at all. The R&W passage makes it sound feasible, so maybe there is hope!

Perhaps a better question is if I'm being distracted by CV methods, and perhaps there's something better suited to GPs?

[dfm]

I think it's a totally sensible thing to do!

[bmorris3]

I think I've sorted this out and can close the issue, thanks Dan!