[Rd] Initial covariance matrix in StructTS

[Helske, Jouni]

Dear all,

The definition of the initial covariance matrix P in StructTS function seems to be defined in a somewhat non-standard way without any references. Usually that matrix is defined as a diagonal matrix in case of structural time series models, but StructTS defines this as a singular matrix filled with  1e+06 * var(x, na.rm = TRUE)/100 where x is the time series being modelled.

I wonder if this is a bug or an undocumented feature?

There was more detailed question by Javier López-de-Lacalle in R-help last November without replies:  https://stat.ethz.ch/pipermail/r-help//2014-November/423128.html<https://stat.ethz.ch/pipermail/r-help/2014-November/423128.html>

There is also illustrative figure in his blog post about the effects of this initial covariance matrix (scroll to the bottom of the page):
http://www.jalobe.com:8080/blog/variations-on-a-maximum-likelihood-procedure/

Best regards,

Jouni Helske

	[[alternative HTML version deleted]]