estimation of time varying parameter state space model

[akrohit]

Plz advise me on estimation of time varying parameters in state space models or how to use kalman filter for time varying models in eviews. The state space model webpage in eviews gives an explanation for constant coefficient models and not time varying ones.

[EViews Glenn]

For most time-varying specifications, you should use the Auto-specification tool.

http://www.eviews.com/help/helpintro.ht ... ws.html%23

Just choose which type of type-varying parameter specification you want, and enter the name of the corresponding variable in the appropriate box in the Stochastic Regressors page.

[akrohit]

Thank you for your reply. I estimated a state space model with the following equations:
@signal y1 = sv1*y2 + sv2*y3 + [var = exp(c(1))]
@state sv1 = sv1(-1) + [var = exp(c(2))]
@state sv2 = sv2(-1) + [var = exp(c(3))]

The results I get show me a warning signal as well as few other comments which I am not able to understand. Kindly help me with this. Moreover, I am also getting NA in the std errors and p values of my estimated c(1), c(2) and c(3). Is there any mistake in estimation?


Sspace: UNTITLED
Method: Maximum likelihood (BFGS / Marquardt steps)
Date: 08/18/16 Time: 14:50
Sample: 1990M01 2016M03
Included observations: 315
Failure to improve likelihood (non-zero gradients) after 19 iterations
Coefficient covariance computed using outer product of gradients
WARNING: Singular covariance - coefficients are not unique

Coefficient Std. Error z-Statistic Prob.

C(1) -2.736385 NA NA NA
C(2) -4.145503 NA NA NA
C(3) -56.48941 NA NA NA

Final State Root MSE z-Statistic Prob.

SV1 0.715693 0.254579 2.811279 0.0049
SV2 -0.122925 0.027835 -4.416157 0.0000

Log likelihood -113.6310 Akaike info criterion 0.740514
Parameters 3 Schwarz criterion 0.776253
Diffuse priors 2 Hannan-Quinn criter. 0.754793

[EViews Glenn]

Try different starting values. Though the results that you get there suggest that the coefficient on Y3 is not a random walk as the variance estimate is approaching zero.

[startz]

Are you sure you don't want an intercept?

[akrohit]

Try different starting values. Though the results that you get there suggest that the coefficient on Y3 is not a random walk as the variance estimate is approaching zero.

Thank you for your reply. Let me try with different starting values.

Are you sure you don't want an intercept?

I am replicating a res paper in which they have not used an intercept. So, I did not use intercept. Can u plz elaborate so as to when should an intercept be taken and when it should not be taken for a state space estimation. Thank you for your response btw.

[startz]

You should almost always have an intercept unless the data is in deviations from the mean. Same as a simple regression.

[akrohit]

Actually the series y1, y2 and y3 are in the form of (inflation expectations - target inflation) as in the paper by Strohsal et al. (2016) published in Journal of Macroeconomics. They have not used an intercept. Can you plz elaborate. Thank you

[startz]

If the data all has approximately the same mean then a constant probably isn't necessary.

[akrohit]

Ok thank you. there is one more doubt. In the estimation of state space models do we need to account for the non stationarity in may data?

[startz]

Yes. Although nonstationary series are often modeled explicitly as opposed to being detrended before the analysis.