Jake Taylor

@davmre @jburnim Following [these notes](http://web.pdx.edu/~crkl/readings/Hamilton94.pdf) on Hamilton ARMA SSMs (which it seems `autoregressive.py` was already following), as @davmre suggests, it is easy to modify the `H'` matrix to arrive at...

@davmre, okay, I'd love to contribute (seeing that we use the library quite extensively and really love it!). Here is a [pull request.](https://github.com/tensorflow/probability/pull/1474). I also added a quick check against...

By the way, now that we have a super cool ARMA(p, p - 1) sts model, some users (like me) might be interested in prediction intervals from the marginal distribution...

Okay, yes this is super helpful! We are more in the consideration phase of importing the `sts` module outright (we have heavy use of just the vanilla distributions module), but...

Yes, I think this is the most straightforward way to accomplish this, and best for a first iteration. Additionally, maybe we could directly pass in a `guide` param to `bm.SVI()`...

I think your tutorial solidly covers the first two uses cases in my original post 🥳. Since BM natively covers `fullbayes`, I believe only the `MixedHMC` cases are left 😎....

> @jakee417 Thanks a lot for the PR. I will try to review it ASAP, but might take me some time. Yes understandable, its quite a large PR. Perhaps we...

I can follow up with the Torch errors, was not aware there was a PyTorch backend 😅

Seems there is a bug in [DefaultPredictionStrategy](https://github.com/cornellius-gp/gpytorch/blob/main/gpytorch/models/exact_prediction_strategies.py#L312). Specifically with ``` test_covar = joint_covar[..., self.num_train :, :].to_dense() ``` Once I patched it like this: ```python import torch import gpytorch import numpy...

Actually, I tested setting: ``` ``` in https://github.com/jakee417/jakee417.github.io/commit/2f918c9157f79e65e73ba49974cd82969b8b1d95 and it seems to now have the correct behavior (and even changes with the system theme!). As an added bonus, when enlarging...