Stationary-diffuse prior during initialization of statespace MLEModel

[as4...@columbia.edu]

Is there an option of giving stationary-diffuse prior, as described in Doan (2010). This prior is implemented in the RATS software under the PRESAMPLE option ERGODIC in the DLM function.

I currently use the initialize_approximate_diffuse option but wanted to know if the results would vary significantly with these priors. 

[Chad Fulton]

On Thu, Jun 8, 2017 at 3:40 PM, <as4...@columbia.edu> wrote:

Is there an option of giving stationary-diffuse prior, as described in Doan (2010). This prior is implemented in the RATS software under the PRESAMPLE option ERGODIC in the DLM function.

I currently use the initialize_approximate_diffuse option but wanted to know if the results would vary significantly with these priors. 

We don't yet have exact diffuse initial filtering and smoothing as Doan uses, but you can combine approximate diffuse filtering with stationary initialization by computing the initial variance yourself each time you `update` via the `initialize_known` method. This is what we do in the SARIMAX and UnobservedComponents models (see the `initialize_state` method).

As far as whether the results would vary, I'll give you two views:

- My experience is that it does not usually matter too much, and in practice lots of people seem to use approximate diffuse initialization.

- Doan (2010) seems to think that it does matter (although of course he took the time to write a paper about his method, so he's going to want to advocate for it)