Model scoring on out-of-samples predictions

Yes, this looks correct. But it’s a bit of a waste to reduce your posteriors to a point estimate. If you’re going to do that, why bother running the expensive and time consuming MCMC algorithm in the first place?

You should consider computing a model score for each posterior draw, and look at the posterior distributions over the scoring functions.