Convergence issues EGARCH Model

[lsal0106]

Hi I am new here so please be nice .

I am having some issues with regards to my EGARCH models converging, and the convergence issues seems to arise at Random.

The issue I have is that for example, using exactly the same dependent variables for a particular country, but change the dependent variable (e.g. CDS premia with 1 year maturity, to CDS premia with 10 year maturity), the model no longer converges.

I have tried various cvcrit (from 0.1 to 0.0000000000001), various piters from 10 to 950, and various subit from 5 to 30. And combinations in between. I have also tried the manual maximize code instead of the built-in GARCH command, and the model still does not converge.

Does anyone have any tips? Any advice would be greatly appreciated.

Thanks.

[TomDoan]

Using a really small CVCRIT definitely won't help. Did you include the ASYMMETRIC option? That sometimes helps when the base EGARCH doesn't converge.

[lsal0106]

Thanks for your reply. The asymmetric options is already included in the code. Even when I specify a cvcrit of 0.01 there is no convergence unfortunately. It is only for a few of the estimations, but still frustrating.

[hasanov]

Hi,

In case, attempts with different conditional mean specifications do not help, you may try with derivative-free methods (i.e. simplex) which are made available in RATS and are very useful for initial guesses. This method can simply be incorporated in GARCH command (for example, pmethod=simplex,piters=5).

Derivative-free methods for ML optimization are described in some depth in the following books.

Green, W., 2012. Econometric analysis, seventh ed. Pearson Education, Inc., Prentice Hall.
Press, W.H., B.P. Flannery, S.A. Teukolsky and Vettering, W.T., 2007. Numerical Recipes, third ed. Cambridge University Press, New York.

Hope this will help.

[TomDoan]

Thanks for your reply. The asymmetric options is already included in the code. Even when I specify a cvcrit of 0.01 there is no convergence unfortunately. It is only for a few of the estimations, but still frustrating.

You would have to post your program and data so we can see.

[lsal0106]

Thank You everyone for your fantastic feedback.

I will try the PMETHOD=SIMPLEX method with different convergence criteria.

In the mean time I have attached an example of the code I am using (for one of the countries that did not converge), as well as the underlying data series in excel.

The interesting thing is that this problem only occurs in a very few of hundreds of estimations!

Thanks

[lsal0106]

Using a really small CVCRIT definitely won't help. Did you include the ASYMMETRIC option? That sometimes helps when the base EGARCH doesn't converge.

With a couple of the estimations, when I excluded "asymmetric" from the code - the model converged? Are there any wide ranging implications of doing this?

[TomDoan]

First, is there a reason you are putting so many regressors (which isn't likely to be a problem), and xregressors (which might be) in the model? Have you tried models with a more modest number of shifts?

Second, your data at the start of the sample is apparently pegged at specific rates for long stretches of time (I'm assuming that these are CD rates which were covered by Reg Q or something like that). That is not the same data generating process as the second half of the data set, and no type of GARCH model will be able to fit that satisfactorily---after all, the GARCH means that shocks are followed by a period of noisiness, while there, by construction, shocks are followed by a flat period. At least as far as fitting GARCH models, you'll have to restrict yourself to the second half of the data set.

[lsal0106]

Thanks for your feedback. I thought the issue may have been because of the nature of the CD data itself. In that for countries such as the US and AU, there was not much volatility in the data (i.e. quite flat) for long periods of time.

I might try a sub sample and see if there is convergence. Many thanks for your help - this forum is great!