Excessive piters?

[oli4]

Hi

I am running a multivariate garch model as specified by McAleer with varma variances. Statistics and analysis of univariate models suggest strongly the standardised residuals are non-norm so using robust errors.

My question relates to this, when I increase the number of initial iterations of an initial simplex procedure the significances change on my estimated coefficients. Their P value drops the more iterations used. I'm currently using 150. The log liklihood the bfgs method converges at also doesn't change.

Is there any harm in using excessive piters in model estimation?

Another question I have is with regard to checking specification. I've used the mvqstat to check residual corr in the standardised residuals and their squares. Because the mean equations are specied as univariate attachments would it be appropriate to ensure no resid corr in these using the regular ljung box test also?

Thank you for taking the time to read.

Oli

[TomDoan]

Hi

I am running a multivariate garch model as specified by McAleer with varma variances. Statistics and analysis of univariate models suggest strongly the standardised residuals are non-norm so using robust errors.

My question relates to this, when I increase the number of initial iterations of an initial simplex procedure the significances change on my estimated coefficients. Their P value drops the more iterations used. I'm currently using 150. The log liklihood the bfgs method converges at also doesn't change.

Is there any harm in using excessive piters in model estimation?

From the User's Guide description of BFGS in Section 4.2:

Because the estimate of G produced by BFGS is used in estimating the covariance matrix and standard errors of coefficients, you need to be careful not to apply BFGS to a model which has already been estimated to a fairly high level of precision. If BFGS uses fewer iterations than the number of parameters being estimated, the G will be poorly estimated and, hence, the standard errors derived from them will be incorrect.

The problem with too many PITERS is not that you get to the wrong point estimates, but that you get incorrect standard errors.

Another question I have is with regard to checking specification. I've used the mvqstat to check residual corr in the standardised residuals and their squares. Because the mean equations are specied as univariate attachments would it be appropriate to ensure no resid corr in these using the regular ljung box test also?

Yes. The advantage of @MVQSTAT is that it can catch cross equation serial correlation. The disadvantage is that it doesn't have as much power as the univariate tests for detecting serial correlation within an equation.

[oli4]

Thank you very much for your quick reply. Helped immensely.