Likelihood of individual data points

[odemangeon]

I have been trying to perform model comparison of models which include celerite (and/or george) kernels using the leave one out (LOO) and widely applicable information criterion (WAIC) (see for example https://arxiv.org/abs/1507.04544).
These algorithms require the computation of the ln(likelihood) of individual data point. Is there a way to compute this with the current interface ?

[dfm]

I haven't directly implemented anything like this and it will be somewhat more complicated with GPs than with independent data, but the basic procedure is discussed in Section 5.4.2 of R&W. I bet it would be possible to implement this using the features already available (in particular the apply_inverse method), but there might also be clever ways to speed it up that I haven't thought of. I'd be happy to chat more if you put together a demo!

[odemangeon]

Thanks for the pointers ! I will have a go at it when I have some time and go back to you.