The RATS Software Forum

Thank you Thomas for your amazing answers (as always) !

One more to go: Do you find it problematic if in DCC-process A&B are both small and insignificant, but in ADCC both parameters are significant (with a sum over 0.99) ?

Regressions are run with same data and with same pairwise countries.

Yes. That seems rather odd.

Okey, seems that I have to ask one more and (hopefully) last question.

In this topic, there´s has been some question about the LB Q-stat. in a multivariate environment. I´m in the same process. I´ve got now few outputs and I´m a bit confused: should there be any sign (like * to 5% for example) to show the significance of the result? I just got a output where´s only mv Q and significance level as chi-squared.

Also, in many papers LB Q-stat is done for standardized residuals and squared standardized residuals. How come squared stand. residuals can be done in mv-environment? Also in these papers when the series are jointly tested; the stand. resids. seems to be significant and squared not.

The point is only to check if some autocorrelation remains in the series when jointly tested. What is the best procedure for doing them both? @mvqstat?

BR

Simo Koskelo

There's no simple extension of the McLeod-Li test (the LB on the squares) for multivariate residuals. For that, you want to apply @MVARCHTEST to the jointly standardized residuals. @MVQSTAT should also be applied to the jointly standardized residuals. You shouldn't apply either one to the univariate standardized residuals (i.e. each residual series divided by its own variance). If you do that, you're leaving a time-varying correlation among the series in the test which (a) will likely produce a significant result in the ARCH-test since it's one of the things that it's testing for and (b) will throw off the Q test since it's based upon the assumption of a fixed covariance matrix.

Univariate tests on univariate standardized residuals are OK and can sometimes be helpful in spotting problems, but the main tests should be the multivariate ones since that's what the model is actually trying to explain.

Thank you sir!


Here´s an example of desired output table (in red). Hopefully it will clear the previous question. So that can be done via mvarchtest and mvqstat? The last question, does RATS show in the output when output from Q-test is significant? Since that was one point in this.

The way I read that, those are the tests on univariate standardized residuals---they're not the multivariate diagnostics. See the top of Section 9.4.6 in the User's Guide. @BDINDTESTS does both the LB test, the McLeod-Li test (which is the LB^2) and a few others.

RATS doesn't automatically * the significant results---everyone has their own definitions for what levels need to be highlighted and how.

Regarding the results, it looks like there is a (modest) amount of serial correlation in the Russian series (which I'm assuming is #1 in the first three pairs) which isn't being modeled.

Thank you Thomas!


I have one question about the graphs. How the mean correlation of period under observation can be graphed also in the same graph with dcc -correlation? Is this possible in RATS?

I can make graphs of dcc -correlation, but could not find how to add mean into it?!


Thanks again, and merry Christmas!

BR

Simo

Use the VGRID option on GRAPH.

spgraph(vfields=3,footer="Conditional Correlations")
stats(noprint) jpnfra
graph(header="Japan with France",min=-1.0,max=1.0,vgrid=%mean)
# jpnfra
stats(noprint) jpnsui
graph(header="Japan with Switzerland",min=-1.0,max=1.0,vgrid=%mean)
# jpnsui
stats(noprint) jpnsui
graph(header="France with Switzerland",min=-1.0,max=1.0,vgrid=%mean)
# frasui
spgraph(done)

Hi Tom,

I´ve got one last question about the LB-Q -test. How can I measure the significance level of the test? For example at 5% level?

Here´s just some random output I have:

Multivariate Q(8)= 39.10582
Significance Level as Chi-Squared(32)= 0.18091

How do I know are they significant or not?

Thanks!

See

https://estima.com/ratshelp/significancelevels.html

For your case, your statistic is what you would like to see---you would accept the null of no serial correlation.

hi!


Since I am also a beginner in DCC-GARCH estimation...I have few questions if can anyone help me!

a) I use bond yield spreads in my analysis but they are non stationary. Should I take first differences (that make my variables stationary) before running DCC model?
b) The DCC-GARCH model has to be applied according to "https://estima.com/procs_perl/garchmvdcc2.rpf" code or simply by the menu in rats? For example I estimated the model and then I plotted graphs for conditional correlations as follows
""
GARCH(P=1,Q=1,MV=DCC,VARIANCES=SPILLOVER,rvectors =rr, hmatrices=hh,ROBUST,noprint) / CA FR GM IT JP UK US
set corr_ad %regstart() %regend() = %cvtocorr(hh(t))(6,7)
GRAPH(STYLE=LINE,HEADER="DCC:UK-US") 1

c) and finally, can you help me about how to extract conditional volatility from the model?


I appreciate you help!

Thanks

V

power23 wrote:hi!


Since I am also a beginner in DCC-GARCH estimation...I have few questions if can anyone help me!

a) I use bond yield spreads in my analysis but they are non stationary. Should I take first differences (that make my variables stationary) before running DCC model?

Your mean model has to roughly whiten the residuals, so you have to do something. Typically if you have data with serial correlation, you'll use a low order VAR to remove it. See, for instance, the VAR(1) example that's part of GARCHMV.RPF.

power23 wrote: b) The DCC-GARCH model has to be applied according to "https://estima.com/procs_perl/garchmvdcc2.rpf" code or simply by the menu in rats? For example I estimated the model and then I plotted graphs for conditional correlations as follows
""
GARCH(P=1,Q=1,MV=DCC,VARIANCES=SPILLOVER,rvectors =rr, hmatrices=hh,ROBUST,noprint) / CA FR GM IT JP UK US
set corr_ad %regstart() %regend() = %cvtocorr(hh(t))(6,7)
GRAPH(STYLE=LINE,HEADER="DCC:UK-US") 1

That's fine except for the mean model.

power23 wrote: c) and finally, can you help me about how to extract conditional volatility from the model?

For instance

set hca %regstart() %regend() = hh(t)(1,1)

The diagonal elements of HH(t) are the variances.

Dear Tom


Help you for your help!I followed your recommendations and I estimated the model
However, my sample includes 7 countries and when I applied the changes the program is not responded and closes immediately. However, when I applied 4 countries all they work well! Is this a problem in my rats version or in the model applied?

*model for the mean-VAR(1)*
system(model=var1)
variables CA FR GM IT JP UK US
lags 1
det constant
end(system)

*DCC-GARCH dynamic correlations*
GARCH(P=1,Q=1,model=var1,MV=DCC,VARIANCES=SPILLOVER,rvectors =rr, hmatrices=hh,ROBUST,noprint)
set corr_ad %regstart() %regend() = %cvtocorr(hh(t))(1,2)
GRAPH(STYLE=LINE,HEADER="DCC:Canada-France") 1
# corr_ad
set hca %regstart() %regend() = hh(t)(1,1)
set stddev = sqrt(hca)



Thanks once again

How much data do you have? It sounds like you don't have all that much.

I have 240 monthly observations for each of the 7 countries from 1994 to 2014