Page 1 of 2

Diagnostic test and mean model for BEKK-GARCH

Posted: Sat Aug 04, 2018 5:45 pm
by qkvm84
Dear Tom,

I have just estimated a BEKK-GARCH model with the mean model VAR(1) and did diagnostic test to see whether the model is adequate. I did both "univariate" diagnostic and multivariate diagnostic follow the RATS' User Guide.

Here is my diagnostic result:

Code: Select all

MV-GARCH, BEKK - Estimation by BFGS
Convergence in    60 Iterations. Final criterion was  0.0000041 <=  0.0000100
Weekly Data From 1997:07:08 To 2007:06:26
Usable Observations                       521
Log Likelihood                     -2394.7671

    Variable                        Coeff      Std Error      T-Stat      Signif
************************************************************************************
Mean Model(RHK)
1.  RHK{1}                       -0.050488729  0.046670912     -1.08180  0.27934011
2.  RUK{1}                        0.075863528  0.059254207      1.28031  0.20043747
3.  Constant                      0.274979114  0.121417868      2.26473  0.02352904
Mean Model(RUK)
4.  RHK{1}                       -0.026833057  0.030589271     -0.87720  0.38037539
5.  RUK{1}                       -0.082991813  0.052566271     -1.57880  0.11438113
6.  Constant                      0.234809002  0.079567031      2.95108  0.00316661

7.  C(1,1)                       -0.297073107  0.084551821     -3.51350  0.00044224
8.  C(2,1)                       -0.010294091  0.085941519     -0.11978  0.90465728
9.  C(2,2)                       -0.263180811  0.083489962     -3.15224  0.00162020
10. A(1,1)                        0.277588736  0.040689037      6.82220  0.00000000
11. A(1,2)                        0.065113722  0.022653730      2.87430  0.00404918
12. A(2,1)                       -0.126032635  0.042071421     -2.99568  0.00273831
13. A(2,2)                       -0.274239692  0.046336814     -5.91840  0.00000000
14. B(1,1)                        0.960850548  0.013777814     69.73897  0.00000000
15. B(1,2)                        0.021612663  0.008070182      2.67809  0.00740436
16. B(2,1)                        0.012756593  0.016342192      0.78059  0.43504218
17. B(2,2)                        0.946374383  0.016052239     58.95591  0.00000000


Independence Tests for Series Z1
Test            Statistic  P-Value
Ljung-Box Q(40)  33.557567     0.7541
McLeod-Li(40)    41.838360     0.3910
Turning Points    1.769490     0.0768
Difference Sign  -1.971055     0.0487
Rank Test         1.508427     0.1314


Independence Tests for Series Z2
Test            Statistic  P-Value
Ljung-Box Q(40)  40.749492     0.4373
McLeod-Li(40)    15.902004     0.9998
Turning Points    0.312263     0.7548
Difference Sign  -0.454859     0.6492
Rank Test         0.506336     0.6126

Multivariate Q(5)=       9.65744
Significance Level as Chi-Squared(20)=       0.97397

Test for Multivariate ARCH
Statistic Degrees Signif
    72.90      45 0.00530
As you can see the univariate Q test and the univariate McLeod-Li test is insignificant which shows that the univariate standardized residuals are serially uncorrelated. In the mutivariate diagnostics, the multivariate Q is insignificant while the mvarchtest reject strongly. This result is like the example in @GARCHMV.RPF, the results of McLeod-Li test and the mvarchtest are conflicted. But there is no further explaination about that. What is the different between these two tests?

Looking at the RATS' User Guide and your reply in other topics, you said the model is adequate when the MVQSTAT and MVARCHTEST are insignificant. According to that, is it my estimated model above is invalid? Or can I say my model is adequate because of the univariate diagnostics test results? If not, how can I improve my model to be adequate? I have tried to add lags to the VAR mean model, but it is still strongly rejected in MVQSTAT.

Another question: How to decide the mean model before estimating the BEKK model? From the VARMAGARCH.RPF, it says the GARCH with a VAR mean model is common and better than VARMA. In general, it is best to start with a small model and add lags until the model is adequate. So can I use VAR(1) mean model straight away at the beginning? And once the BEKK-GARCH is adequate, there is no problem of using the VAR(1) mean model. In other words, can I say if the estimated BEKK-GARCH model is adequate, which kind of mean model is not of interest?

Thanks,

Haoting

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Sat Aug 04, 2018 9:05 pm
by TomDoan
Most of your questions are answered in greater detail in the ARCH, GARCH and Volatilities Models e-course. However, the univariate diagnostics are doing 40 lags on the McLeod-Li which is way too many to tell you much about residual GARCH---there won't be much power in a test on such distant lags. Regarding the significant @MVARCHTEST see the "Diagnostics on Large Data Sets" in the June 2018 newsletter. While 500 entries isn't quite the type being addressed in that, the reality is that that result is still due to a number of relatively small correlations.

There is no way to tell in advance what type of mean model is needed in a GARCH model because the GARCH residuals will throw off any standard methods of selecting lag length. However, if you use @VARLAGSELECT with BIC, you'll usually get a model which isn't excessively large. Based on your results, I'm going to guess that that would pick 0 lags in the first place.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Sun Aug 05, 2018 6:07 pm
by qkvm84
Thanks for your quick reply, Tom

Could I ask some further questions?

What is the appropriate lags for the univariate diagnostics? Is 10 suitable?

Can I just rely on the results of univariate diagnostics to say my estimated model is adequate if the test result is not significant? Or we must do both univariate and multivariate diagnostics?

If finally, the @MVARCHTEST is still significant, can I still use this BEKK-GARCH estimate results to do empirical analysis? Because I am doing a master dissertation, I was stuck by this diagnostics test

Does 0 lags in the VAR mean model you mention is estimated using the code here? With no mean model defined.

Code: Select all

garch(p=1,q=1,mv=bek,pmethod=simplex,method=bfgs,piters=10,rseries=rs,mvhseries=hhs,stdresids=zu,derives=dd) / RHK RUK
Thanks

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Sun Aug 05, 2018 6:25 pm
by TomDoan
qkvm84 wrote:Thanks for your quick reply, Tom

Could I ask some further questions?

What is the appropriate lags for the univariate diagnostics? Is 10 suitable?
Anything beyond five is unlikely to be very helpful.
qkvm84 wrote: Can I just rely on the results of univariate diagnostics to say my estimated model is adequate if the test result is not significant? Or we must do both univariate and multivariate diagnostics?
You don't really need to do univariate diagnostics with a multivariate model, as an insignificant result doesn't really tell you much. A significant univariate diagnostic can be more helpful than a significant multivariate one in helping you figure out what the problem is (since the failure is simpler to examine in detail) but it's quite possible to have a really bad multivariate model which passes univariate tests.
qkvm84 wrote: If finally, the @MVARCHTEST is still significant, can I still use this BEKK-GARCH estimate results to do empirical analysis? Because I am doing a master dissertation, I was stuck by this diagnostics test
You can, but you need to rule out obvious problems that might produce problems. Again, those are covered in the GARCH e-course.
qkvm84 wrote: Does 0 lags in the VAR mean model you mention is estimated using the code here? With no mean model defined.

Code: Select all

garch(p=1,q=1,mv=bek,pmethod=simplex,method=bfgs,piters=10,rseries=rs,mvhseries=hhs,stdresids=zu,derives=dd) / RHK RUK
Correct. The default is mean only in each equation.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Mon Sep 24, 2018 12:00 am
by sanjeev
Hi,
I want to incorporate dummy variables to account for structural breaks in my model.Could you help me do that? My precise question is how do we incorporate an exogenous variable in our model.
Please reply soon!
Thanks.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Mon Sep 24, 2018 8:52 am
by TomDoan
sanjeev wrote:Hi,
I want to incorporate dummy variables to account for structural breaks in my model.Could you help me do that? My precise question is how do we incorporate an exogenous variable in our model.
Please reply soon!
Thanks.
If it's a variance shift, it's with the XREG option. If it's a mean shift, you just incorporate it into the mean model. It would help if you could be more specific about what the exogenous variable is supposed to fix.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Mon Sep 24, 2018 8:27 pm
by hasanov
Hi Tom,

"If it's a variance shift, it's with the XREG option. If it's a mean shift, you just incorporate it into the mean model. It would help if you could be more specific about what the exogenous variable is supposed to fix".

How about asymmetry dummies represented by D matrix in estimation output? Is it also based on XREG option?

Thank you

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Mon Sep 24, 2018 11:14 pm
by TomDoan
That's not an exogenous variable---the "dummy" multiplier is constructed endogenously.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Tue Sep 25, 2018 12:32 am
by hasanov
Ewing and Malik (2005) added a set of dummies for breaks in BEKK specification (see Eq. 9 on page 2663). The authors used RATS.

Ewing and Malik (2005). Re-examining the asymmetric predictability of conditional variances: The role of sudden changes in variance. Journal of Banking & Finance 29, 2655–2673

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Tue Sep 25, 2018 3:34 am
by sanjeev
TomDoan wrote:Ewing and Malik (2005). Re-examining the asymmetric predictability of conditional variances: The role of sudden changes in variance. Journal of Banking & Finance 29, 2655–2673
So, I wish to incorporate a structural break in my series due to reforms with the help of the dummy(a 0-1 dummy). So, does it come as another variable in the mean model?

Please reply.
Thanks.
Regards.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Tue Sep 25, 2018 6:45 am
by TomDoan
sanjeev wrote:So, I wish to incorporate a structural break in my series due to reforms with the help of the dummy(a 0-1 dummy). So, does it come as another variable in the mean model?

Please reply.
Thanks.
Regards.
No. Absolutely not. Use an XREG variance shift regressor.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Tue Sep 25, 2018 7:34 am
by TomDoan
hasanov wrote:Ewing and Malik (2005) added a set of dummies for breaks in BEKK specification (see Eq. 9 on page 2663). The authors used RATS.

Ewing and Malik (2005). Re-examining the asymmetric predictability of conditional variances: The role of sudden changes in variance. Journal of Banking & Finance 29, 2655–2673
What they're doing is just a bad idea. See the discussion at https://estima.com/forum/viewtopic.php?p=15720#p15720

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Fri Sep 28, 2018 2:28 am
by sanjeev
Dear Tom,
Actually, it is a 0-1 dummy that takes value 1 for the duration of global financial crisis and 0 otherwise. So, can I simply put it as another variables in my model? If so, how would the coefficients be interpreted in this case?
Please reply!

Thanks.
Regards.
TomDoan wrote:
sanjeev wrote:Hi,
I want to incorporate dummy variables to account for structural breaks in my model.Could you help me do that? My precise question is how do we incorporate an exogenous variable in our model.
Please reply soon!
Thanks.
If it's a variance shift, it's with the XREG option. If it's a mean shift, you just incorporate it into the mean model. It would help if you could be more specific about what the exogenous variable is supposed to fix.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Fri Sep 28, 2018 8:07 am
by TomDoan
sanjeev wrote:Dear Tom,
Actually, it is a 0-1 dummy that takes value 1 for the duration of global financial crisis and 0 otherwise. So, can I simply put it as another variables in my model? If so, how would the coefficients be interpreted in this case?
Please reply!

Thanks.
Regards.
TomDoan wrote:
sanjeev wrote:Hi,
I want to incorporate dummy variables to account for structural breaks in my model.Could you help me do that? My precise question is how do we incorporate an exogenous variable in our model.
Please reply soon!
Thanks.
If it's a variance shift, it's with the XREG option. If it's a mean shift, you just incorporate it into the mean model. It would help if you could be more specific about what the exogenous variable is supposed to fix.
To repeat: if it's a variance shift, it's with the XREG option. If it's a mean shift, you just incorporate it into the mean model. You do not make it another variable in the GARCH model.

Re: Diagnostic test and mean model for BEKK-GARCH

Posted: Fri Sep 28, 2018 11:17 am
by sanjeev
Dear Tom,
I have another question. While trying to estimate my BEKK model, I get the following error message if I include more than three variables in the system:
MV-GARCH, BEKK - Estimation by BFGS
NO CONVERGENCE IN 197 ITERATIONS. FINAL NORMED GRADIENT 0.00013
SUBITERATIONS LIMIT EXCEEDED.
ESTIMATION POSSIBLY HAS STALLED OR MACHINE ROUNDOFF IS MAKING FURTHER PROGRESS DIFFICULT
TRY DIFFERENT SETTING FOR EXACTLINE, DERIVES OR ALPHA ON NLPAR
RESTARTING ESTIMATION FROM LAST ESTIMATES OR DIFFERENT INITIAL GUESSES/PMETHOD OPTION MIGHT ALSO WORK

And then it gives some output.Could you please throw some light on it?Please reply soon!

Thanks.
Regards.