Estimation of an Unobserved Components Model using the Multivariate Diffuse Kalman Filter

[czuber]

Dear all,

I get the following message (many times) during Bayesian estimation of a simple multivariate UC model (which is an extended version of Harvey (1985) and Clark (1987)):

univariate_diffuse_kalman_filter:: T does influence the rank of Pinf!

Missing values at the beginning seem to be no problem here (as pointed out in this discussion).

Any ideas on how to get Dynare to use the multivariate filter? Or does anyone have experience in reformulating the problem to a univariate treatment as Durbin & Koopman (2012, p. 125) suggest?

The code attached.
problem.zip (2.7 KB)

[jpfeifer]

I am not sure I understand your problem. Regarding your warning, see

Univariate diffuse kalman filter: T does influence the rank of Pinf!

This warning (not sure it was me who made it) is related to the fact that once Pinf is for example numerically null rank (according to the tolerance set), then the updated matrix TP_\infty T' should also be rank deficient. when the warning is issued, in most cases this is due just to numerical issues, but it should give the message not systematically. Otherwise, it may be an indication of lack of some issues in how the model is defined (whether unit roots are properly projected onto observabl…

So that may not necessarily be a problem.

In terms of setting up the problem, did you have a look at

tests/kalman_filter_smoother/test_compute_Pinf_Pstar.mod · master · Dynare /...

[czuber]

Thanks. I have seen the discussion you are referring to before but the second link helped. It was about the timing of the cyclical and trend component and when I correct for this, the warning vanishes.