The RATS Software Forum

I am trying to fit VAR-multivariate garch models to the returns series of three exchange rates.I am having intervention as the exogenous variable affecting both mean and the variance of exchange rates.I have got reasonably good results from the VAR- BEKK model.For the alternative model specification , i am trying to fit VAR-DCC model with variances=varma but it is showing non-convergence.Changing initial values might help me but I am unable to understand the option through which we can provide initial values of parameters to rats in multi-variate GARCH models .Please tell me how to provide initial values as a vector in multivariate garch models.Also, kindly tell me the procedure/codes for fitting the VAR- DCC two-step estimation to my returns series data with exogenous variable/s in both mean and variance equations.

Thanks
Sanjeev

Have you tried a simpler DCC model (that is, without VARIANCES=VARMA)?

You can't do a two-step DCC with VARIANCES=VARMA because the two-step procedure requires you to be able to estimate separate GARCH models on each series. VARIANCES=VARMA links the variance equations, so it requires a full system estimate.

Thankyou so much for the quick reply! Yes, i have tried simple VAR-DCC model , it is converging and giving good results . Since one of my hypothesis is to check for the variance spillovers across exchange rates,I need to fit DCC with variances = varma in order to confirm my results with alternative specification.Please help.

VARIANCES=VARMA was really designed to be used with CC, not with DCC so it might be quite hard to get it to fit with DCC. However, you would have to post your program and data for us to see if there's anything that can be done differently.

Thankyou for the instant reply! I am attaching my data file consisting of two exchange rate return variables rd and re respectively and net (intervention variable).VAR(1)-BEKK specification works out well with the following code:

GARCH(P=1,Q=1,mv=bekk,XREGRESSORS,REGRESSORS,method=bfgs,PMETHOD=simplex,RVECTORS=RA,HMATRICES=HH,PITERS=10,subiters=1000) / RD re
# Constant RD{1 } re{1 } NET{1}
# NET{1}

But as soon as i choose mv=DCC in the above command , the system do not converge even in this simple model.Even the constant correlation model is showing non-convergence.I am unable to understand why only VAR(1)-BEKK is giving results with my data.Please help.

Thanks,
Sanjeev

1. You don't have much data.
2. Neither of your series has particularly strong "GARCH" effects. RE, in particular, is barely distinguishable from a basic white noise process.
3. RE and RD have very different dynamics. RD seems to have a rather clear structural break between the first five years and the rest of the data and the "NET" variable doesn't really do anything to explain that.

Both DCC and CC are designed to work with a number of fairly closely related series. These don't qualify for that.

Many thanks for the reply. Can you tell me the test that we can use to test for GARCH effects in a series in RATS (other than the Engle ARCH -LM test)? Also, which test we can use for testing the structural break in the series in RATS ?

sanjeev wrote:Many thanks for the reply. Can you tell me the test that we can use to test for GARCH effects in a series in RATS (other than the Engle ARCH -LM test)?

@ARCHTEST (which in the Engle test) and @MCLEODLI. Ordinarily, you would apply those to the residuals from a preliminary linear regression (to get rid of any serial correlation in the mean), but RD accepts lack of "ARCHness" even without that---as I said, it looks like a textbook example of plain old white noise.

sanjeev wrote: Also, which test we can use for testing the structural break in the series in RATS?

In this case, look at the data. There are many different tests for structural breaks because there are many different types of structural breaks, each of which requires a different test. However, you don't need a test when the difference is as striking as it is in this case.

Does anybody know what to do about the following-obviously, getting convergence in DCC-GARCH models is difficult. But when I set the number of iterations to say, 800, as below,

garch(p=1, q=1, model=ar1, mv=dcc, method=bhhh, iterations=800, rvectors=resid, vcv, hmatrices=hh, robust)

I get a message saying that there was no convergence in 338 iterations, or other times a lot less than that, although again I specified 800. In other words, the procedure cuts off well before the number of iterations specified without yielding convergent estimates.
Thank you very much for any input.
Bill

If it stalls short of achieving convergence, there's nothing much it can do with extra iterations. See Section 4.15 of the User's Guide. However, I can suggest the following:

1. Add PMETHOD=SIMPLEX,PITERS=10 options
2. Use METHOD=BFGS rather than METHOD=BHHH

BHHH is not a good choice for the primary method of non-linear estimation.

Hello, what is the difference between DCC with variance spillover as compared to normal DCC .

If we are trying to check DCC method for robustness to support BEKK method for volatility spillover between oil and stocks, should i chose DCC with spillover in variance or not ?