PLEASE HELP!! No Convergence in DCC

[mbfakhry]

Dear all,

I am new to RATS but have been using econometrics models for approximately 5years, however my work have been in univariate models. So in essence I bought RATS to model the covariances and correlations in a multivariate environment.

I am using WinRATS Standard Edition 8.21 on a six core 12gb ram Alienware computer with Windows 8

I am attempting to do a 7 variate DCC to establish the time varying conditional correlation. I have attempted different options to the estimation methods and all resulted with the following error

NO CONVERGENCE IN 86 ITERATIONS
LAST CRITERION WAS 0.0000000
SUBITERATIONS LIMIT EXCEEDED.
ESTIMATION POSSIBLY HAS STALLED OR MACHINE ROUNDOFF IS MAKING FURTHER PROGRESS DIFFICULT
TRY HIGHER SUBITERATIONS LIMIT, TIGHTER CVCRIT, DIFFERENT SETTING FOR EXACTLINE OR ALPHA ON NLPAR
RESTARTING ESTIMATION FROM LAST ESTIMATES OR DIFFERENT INITIAL GUESSES MIGHT ALSO WORK

The series is the daily price differential in 7 sovereign debt markets.
The number of observations for the 7 variate it is 2479 each

I have attached the data files and the output.
Data files:
Multivariate2002 contains the 7 variates and start in 2002

output files:
DCC2002 contains the output for DCC estimations using the multivariate2002 datafile

NB: SORRY FOR THE MISUNDERSTANDING IN USING THE RPF EXTENSION FOR THE OUTPUT

Many thanks in advanced for the help.

[TomDoan]

It seems to work fine if you fit the data through 2007. After that, a number of the series show radically different behavior each other, and from themselves earlier in the data set---since the data aren't described by a single multivariate GARCH process over the full data period, it's rather unlikely that you will find a reasonable multivariate GARCH model which will fit it.

[mbfakhry]

Dear Tom,

Firstly, I'd like to thank you for your reply,

Secondly, it seems to work with the BEKK model which i used to model the volatility spillover effect.

However I do agree with you that the behaviour of the series do change dramatically after 2007. As you said there is an issue, and when I modelled the series in a univariate garch I found that each series had to be estimated using a number of garch models to find the best fit model.
So I knew that finding a multivariate garch model representing all seven series was going to be almost impossible. However that doesn't explain why it worked with the BEKK.

So I'm going to have to estimate the series with a sub period of 2002 to 2007 for the DCC which isn't a problem.

Just one more question how can I interpret the DCC(1) and DCC(2) does high dcc(2) mean that the lagged volatility is correlated with the current volatility in all markets and does a high DCC(1) in relation to the data means that the lagged returns in all markets are correlated with the mean returns.

Do excuse my ignorance if it is not true because I have just started looking at the DCC and CCC models.

Once again, many thanks for your help

Bachar

[TomDoan]

DCC has only 23 parameters to explain the dynamic variances and covariances while BEKK has 126. It appears that DCC in this case is simply too inflexible to handle the apparent break in 2007. A high DCC(2) means a high level of persistence---in the limit DCC(2)=1.0 would mean a CC model (though they don't nest because DCC doesn't estimate the "CC" part). A high DCC(1) means more rapid changes to the correlation matrix.

[mbfakhry]

Dear Tom,

Thanks ever so much for your help.

Bachar.