It would be great to make it work but I get an error:
univariate_diffuse_kalman_filter:: T does influence the rank of Pinf!
This may happen for models with order of integration>1.
Can you provide me with the codes? And why do you use the univariate filter?
Thank you very much for taking a look at this. Please find the data and mod files attached.
mymod.mod (16.0 KB)
estdata.xlsx (15.2 KB)
I did not explicitly specify for the univariate filter to be used, but my understanding is that it switches to the univariate filter if there are issues with the precision in VCV matrices.
Following the discussion in here, I may not be following some Dynare convention with correctly lagging variables in the UCM model.
It does not look like a precision issue. Rather, the multivariate filter returns an exact 0 in the forecast error variance matrix. That suggests a specification problem. Does your model imply an exact linear combination between the observables?
Thanks for looking into this.
I am almost certain that there should not be an exact linear combination. I checked the rank of the Y matrix in dsge_likehood.m, it gives me the exact number of variables…
I specified a “non-standard” prior for standard deviations to be truncated normal. That has some density close to zero, which may lead to the forecast error variance being close to zero. I tried to revert to inverse gamma or push the lower bound further away from zero, but it did not work.
Could this arise as a result of missing observations?
What do you mean with
The Y matrix is about the data, not the theoretical model properties.
I rushed into thinking that you meant the linear combination across observables, therefore, the data.
No, there should not be any linear combinations as implied by the model. The model features many latent variables, including exogenous shocks. If anything, I would suspect nonidentifiability of some variables.
I have decomposed observables into cycle and trend. But given that each component is subject to exogenous shocks, this should not be a problem, right?
But there must be an issue like that. Have you tried dropping one of the observables at a time to see which one(s) may be an issue?