Page 1 of 1
convergence fail, state space model
Posted: Fri Jul 24, 2015 10:16 am
by EviewsUser1
I am estimating a stochastic trend model, which should be straight forward, but I can't get it to converge. Any idea what I did wrong? See attached.
Re: convergence fail, state space model
Posted: Fri Jul 24, 2015 10:27 am
by startz
Are you estimating an AR(1) plus an AR(2)? I don't think the parameters are identified.
Re: convergence fail, state space model
Posted: Fri Jul 24, 2015 10:52 am
by EviewsUser1
I thought I was estimating a random walk for the trend component and an ar(2) for the cylce. The model described by equation (2) in the attached.
Re: convergence fail, state space model
Posted: Fri Jul 24, 2015 10:56 am
by startz
You're right of course. Sorry for the misread. Change the maximum number of iterations to 600 and you should get convergence.
Re: convergence fail, state space model
Posted: Fri Jul 24, 2015 11:51 am
by EviewsUser1
a different error
Re: convergence fail, state space model
Posted: Fri Jul 24, 2015 11:55 am
by startz
If you're using EViews 9, switch the optimization method to EViews Legacy. You get one more iteration and the error message goes away. I'm not entirely sure if this means the problem is solved or if it means it's being masked.
Glenn?
Re: convergence fail, state space model
Posted: Fri Jul 24, 2015 1:42 pm
by EviewsUser1
yes, I was wondering that myself, can I trust the estimates? They do not look like the estimates in the Kuttner paper (even considering that I am using data that has been revised several times).
Re: convergence fail, state space model
Posted: Fri Jul 24, 2015 2:00 pm
by startz
Note that your AR(2) component is essentially just a random walk. What is probably happening is "aliasing." The random walk component is being picked up by what's intended to be the cyclical model and the random walk component has zero variance. These kind of models are very susceptible to this.
You might try different starting values, perhaps something close to Kuttner's. Or get the starting values from the univariate ARIMA representations.
This kind of outcome is a well-known pain in the ass. :(