EViews.com

Hello,

I am beginner in estimation of such a model.
I want to estimate dynamic factor model with 2 unobserved factors where I have 1 quarterly variable "hph" (my object is its forecast) and 6 monthly variables (dtr, ipp, idp, sid, spi, tpc). I am a little confused how to specify the model because in literature there are examples of only 1 factor case.
Quarterly variable is known every third month.

I have the following specification and I would like to ask you to check it whether it is good or how to respecify it to obtain some reasonable results.

@signal dtr = c(1)*f1+c(8)*f2+ [var = exp(c(15))]
@signal ipp = c(2)*f1+c(9)*f2+ [var = exp(c(16))]
@signal idp = c(3)*f1+c(10)*f2+ [var = exp(c(17))]
@signal sid = c(4)*f1+c(11)*f2+[var = exp(c(18))]
@signal spi = c(5)*f1+c(12)*f2+[var = exp(c(19))]
@signal tpc = c(6)*f1+c(13)*f2+[var = exp(c(20))]
@signal hph = c(7)*f1+c(14)*f2 +[var = exp(c(21))]

@state f1 = c(22)*f1(-1) + [var=exp(c(24))]
@state f2 = c(23)*f2(-1) + [var=exp(c(25))]


EViews gives me the following statements:
Failure to improve likelihood (non-zero gradients) after 299 iterations
Coefficient covariance computed using outer product of gradients
WARNING: Singular covariance - coefficients are not unique

All std. errors, z-statistics and p-values are NA

Where is the mistake?

My last question is how to set the starting values for the coefficients?

Thank you for any comments

Looks like an optimization gone numerically wrong.

I have few suggestions that might (or might not) work.

- Setting priors for state values and covariances can help and can be done by:
@mprior v1
@vprior m1

- Initial parameter values can be set (at least for some objects, maybe also the state space object) with
param c(1) 153 c(2) .68 c(3) .15

- Finally, using another form of the variance of disturbances have been proven to work for me. E.g. you could try
[var = c(15)^2] instead of [var = exp(c(15))]