BEKK&DCC

[TomDoan]

1. You have nowhere near enough data to estimate a model that complicated.
2. Your data have very weak (if any) GARCH properties, so the "M" effect won't be identified.
3. The "M" interaction terms won't work the way you have them written---they have to be generated as part of the estimation, not in separate SET instructions.

[TomDoan]

Dear Tom,

Thanks for your reply.

1.What do you mean about "they have to be generated as a part of the estimation".If its possible can you make a bit clear. Is there any program or code example that I can examine as an example. What should ı do to generate as a part of estimation?

You need to use the HADJUST option to create them after the "H" series values get generated. The GARCHMV.RPFexample includes an example of that.

2. What should my sample size be (at least) to estimate a model that complicated. ( my sample size is n=116 begin from 1987:q1 to 2015:q3). Should I work with monthly data rather than quarterly data?

You would certainly be better off. Not only would you have a longer data set, but also the GARCH effects are likely to be stronger at the finer frequency.

[TomDoan]

The SET instructions here:

set k = hhs(1,1)*hhs(2,2)
set x = dk*hhs(1,1)
set y = dk*hhs(2,2)
set z = dk*hhs(1,1)*hhs(2,2)

have the problem that I talked about earlier---the right side depends upon HHS, but HHS is generated as part of the GARCH recursion and so isn't known. SET is a single calculation, not a process. The point of the %() is that you can put multiple calculations within it, so you want

hadjust=%(hhs(1,1)=sqrt(hh(t)(1,1),hhs(2,2)=sqrt(hh(t)(2,2)),k=hhs(1,1)*hhs(2,2),...

Just make sure you close it with the extra ) at the end to match the %(

BTW, you can clean the dummy variable up with

set dk = t>=2003:1

See "Other Dummies" in Section 1.5.9 of the Introduction. (It's also covered by the Trend/Seasonal/Dummies Wizard on Data/Graphics).

[TomDoan]

First note that you're not getting convergence---you need to increase the ITERS option. Program #1 fails because you're defining HHS two different ways. The MVHSERIES option shouldn't be there (and isn't in the other program).

Regarding the two ways to write this, they're equivalent except labeling, but obviously the second is simpler and it's more obvious what it's doing.

[bekkdcc]

Dear Tom,
First of all, thanks for your suggestions.
I increase the ITERS option to 200 and delete the MVHSERIES option.

is they true ?

[TomDoan]

Dear Tom,
First of all, thanks for your suggestions.
I increase the ITERS option to 200 and delete the MVHSERIES option.

1.

I run the two code again called "q1" and "q2". You said the second (q2) is simpler and it's more obvious what it's doing. But when I run the code the first one converge, the other not. If they're equivalent except labeling, why do not give same results? And which one you prefer? or What do you mean by "different" labeling?

The first one "converges" because it's empty---you didn't finish defining your model. (The END(SYSTEM) is missing). The second one needs more iterations. Just give it a 1000 and see what happens. It *is* a big model for the amount of data, so there is no guarantee that it will work.

I already told you that the second one is simpler and more obvious. You should just stop wasting your time with the first one. The second one is set up correctly.

2.

in first model estimation ...... garch(model=mvmean,p=1,q=1,...............) Why there is p=1,q=1 or should I put it in the second model estimation code below, also ?

Aren't those the first two parameters described in the GARCH instruction?

in second model estimation garch(model=mvmean,mv=bekk,rvectors=rd ) Why there is rvectors=rd or should I put it in the first model estimation code above, also ?

That's an output, not an input. If you don't use it (and you don't) you don't need it, but it there is nothing wrong with leaving it in, either.

3.

I am tring to estimate the effects of inflation uncertainty and outputgrowth uncertainty and also their together effect on 4 variables. I am sending the model as an Word document ( MODELSTRUCTURE)...... So whats your thought about the model code that I am trying to use, is they true ?

What I see is that you have two series (INF and EX) that have completely different behavior in the first half of the sample and in the second, and your model gives them only "uncertainty" to explain that. Is that what you believe? The model will do its best to try to explain the difference using what you give it, but it's not clear that the results will be meaningful.

[bekkdcc]

Dear Tom,


4. The model will do its best to try to explain the difference using what you give it.... What do you want to mean by ry to explain the difference using what you give it?

Kind regards,

P.s. I am sending the MODELSTRUCTURE after making some correction on indices.

[TomDoan]

Have you looked at your data? Just one graph should answer #1 and #2.

inflation.pdf

Inflation variable

(3.65 KiB) Downloaded 750 times

The only thing in the model that's different pre and post 2003 are the dummies on the variances. Do you believe that inflation is high pre-2003 only because of uncertainty about inflation? Because that's the only reason you've incorporated into the model.

[TomDoan]

Don't try to force all this into a 4-variable GARCH model. GARCH models were designed for a set of variables with relatively similar dynamics---not four variables as different as these. And even if you figure out a reasonable way to model the uncertainty in inflation, I doubt it will help you much. The problem you have is that the first half of the data set has high inflation, high inflation uncertainty and (I would imagine) rather substantial fiscal issues. If GDP growth is depressed in the first half of the range, which of those is the source? How would you be able to sort them out? You don't have any "natural experiments" for one versus the other, since they all change together. Compare this, for instance, with the Elder-Serletis paper. Oil price had periods when it was high, when it was low, when it was high and variable, when it was high and stable, etc. Thus there is at least some hope that you can separate the level effects from the uncertainty effects. You don't have that.

[TomDoan]

There's no good reason to do a two-step estimator when you have a model with the dynamics linked like that. (And in general, there's no good reason to do two-step DCC unless you have so many variables that a systems estimator isn't feasible).

You would use the same instruction as your BEKK, but with MV=DCC or MV=CC.

[TomDoan]

You're back to the original issue, which is that with quarterly data, you have no "GARCH" effect in the first series (or any of the series actually), so there is no possibility of an "M" effect.

[bekkdcc]

Hi Tom,

Thanks in advance. I have two more questions.

1. Lets say, I estimate the model with CCC and then estimate the same model with DCC. Is there any test that shows CCC is inappropriate? and than means that I can try DCC model?

2. After BEKK, I am doing @mvqstat( serial correlation) and @mvarchtest (remaining ARCH effect) test . Is there any test that ı can do after DCC or I can use the same tests? Is there any other tests, also?

[TomDoan]

Hi Tom,

Thanks in advance. I have two more questions.

1. Lets say, I estimate the model with CCC and then estimate the same model with DCC. Is there any test that shows CCC is inappropriate? and than means that I can try DCC model?

See

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

That's an LM test for CC against an alternative which isn't quite the same as DCC. CC and DCC don't really nest so there is no standard test for CC vs DCC.

2. After BEKK, I am doing @mvqstat( serial correlation) and @mvarchtest (remaining ARCH effect) test . Is there any test that ı can do after DCC or I can use the same tests? Is there any other tests, also?

Same tests. They aren't specific to BEKK.

[TomDoan]

Dear Tom,
Even the model convergence, some ARCH(A coefficients) and GARCH(B coefficients) parameters are negative in results, not only in BEKK model but also in DCC model. Is it normal?

For the BEKK, it's not uncommon. That's described in detail in BEKK model output. For a DCC, it's very uncommon if the GARCH model is appropriate. However, as I've mentioned several times in the thread, at the quarterly level, your data have effectively no GARCH properties. If there is little volatility clustering, the signs of the GARCH coefficients are dependent upon flukes in the positions of residuals. (Theoretically, they would be zero, so small positive and small negative are both possible in practice).

I read somewhere that it can be in that way, because the variance-covariance (H) Matris in BEKK and DCC model is always positive by its structure and the negative coefficient shows that there is a symetric effect. Is it true?

I have no idea what "symmetric" effect means, but that doesn't sound right. Read the link above.

[bekkdcc]

Does negative coefficient can be seen in C matrices on DCC model, or is it also impossible as A and B matirces ?