Time series and Gaussian process

[linas]

@anon79882417

Can you please share a more verbose output?

Few minutes ago I did make update-stan. I still got an error when running stanc:

Semantic error in ‘gp_regression_ard.stan’, line 10, column 32 to column 75:

 8:      vector[N_pred] f2;
 9:      {
10:        matrix[N, N] K = add_diag(gp_exp_quad_cov(x, magnitude, length_scale),
                                     ^
11:                                  sigma);
12:        matrix[N, N] L_K = cholesky_decompose(K);

---

A returning function was expected but an undeclared identifier ‘gp_exp_quad_cov’ was supplied.

@anon79882417

No, for a testing sample we want everything to be “unseen” by the model.

Thanks for the reference. However, I have concern if we can use spatial temporal GP at all. If we make some measurements “unseen” by eliminating test trees from covariance matrix then the measurements that we see would be different from the ones we have observed. Something like cat in a box paradox. If we use analogy that each tree is charged then after eliminating test trees the potential field would be different than it was before elimination. I am bit lost… Hopefully we can.

[anon79882417]

Here’s one out of sample time series. The predictions are pretty awful, but that was to be expected. I’m going to spend some time thinking about what other information I can incorporate to make out of sample predictions more accurate.

[avehtari]

@anon79882417

Here’s one out of sample time series.

Looks like it could be useful to add linear component or use some kernel producing saturating functions (like nnet)