Page 4 of 9

Re: DCCGARCH11

Posted: Tue Mar 31, 2015 3:49 am
by trubador
Delia wrote:Hi trubador. Can u offer an example which used dccgarch11? I don't know what should be put in four empty boxes. Thanks so much.

What do you expect dccgarch11 add-in to do? What data do you have? And most importantly; what is your research question?

Re: DCCGARCH11

Posted: Fri Apr 24, 2015 6:51 pm
by jimi
Hi trubador, i am using panel data in my estimation but the add-in does not allow. Is there a way to go about this? Thanks.

Re: DCCGARCH11

Posted: Sat Apr 25, 2015 3:38 am
by trubador
jimi wrote:i am using panel data in my estimation but the add-in does not allow. Is there a way to go about this?

No, unfortunately not.

Re: DCCGARCH11

Posted: Thu Apr 30, 2015 7:16 am
by pedpereira
akash27 wrote:Before estimating a DCC-GARCH(1,1) model, time series have to be filtered to assure zero expected (mean) value of the time series. Usually, a bivariate Vector Autoregressive (VAR) model used to initially remove potential linear structure, then the residuals of the VAR model are used as inputs for the DCCGARCH model.Is this procedure involved in the add-in or do I need to do this separately and give residuals as input to the add-in?

The add-in only allows univariate AR-X models for mean equations. VAR model needs to be specified at the outset.
[quote="akash27"]

Dear Trabadur, I pretend to apply the dccgarch11 add-in for my research (contagion between bonds and stocks). I already filtered the financial series (through log 1st dif) and now I'm estimating the best ARIMA that fit the filtered series in order to obtain the residuals that "feed" the dcc garch. I believe this is the way to go before applying the add-in, can you confirm if this is it please. Does the add-in only works if the model is AR(P)?

Thanks in advance

Re: DCCGARCH11

Posted: Thu Apr 30, 2015 11:29 am
by trubador
pedpereira wrote:Does the add-in only works if the model is AR(P)?

Unfortunately, yes. However, you can build a separate ARMA mean model at the outset as you did and use the residuals from that model. Then it essentially becomes a three-step method.

Re: DCCGARCH11

Posted: Thu Apr 30, 2015 12:08 pm
by pedpereira
I was going to use RATS to estimate the GARCH-DCC buf if the add-in works with this 3-step method I will try it.

Just for confirmation:
Step 1: "build a separate ARMA mean model and use the residuals from that model"
Step 2: "perform the univariate GARCH and obtain the residuals, cov and etc..."
Step 3: "perform the DCC-GARCH with using those results from the univariate GARCH"

with step 2 and 3 being performed subsequently at once with the add-in, right?

Thanks a lot Trubador

Re: DCCGARCH11

Posted: Thu Apr 30, 2015 12:19 pm
by trubador
Yes, that is correct.

Of course, it is always better to carry out simultaneous estimation if/where possible. Unfortunately, this is something the add-in cannot do at the moment due to some technical difficulties. But it is in my mind...

Re: DCCGARCH11

Posted: Mon May 04, 2015 12:29 pm
by pedpereira
trubador wrote:Of course, it is always better to carry out simultaneous estimation if/where possible. Unfortunately, this is something the add-in cannot do at the moment due to some technical difficulties. But it is in my mind...


Hello again, regarding this particular issue of "quality" i believe i will lose as much with this add-in as with the RATS estimation of garch dcc since the estimation is deferred anyway, am I correct?

After running the add-in on the residuaIs of the ARMA models, I am struggling with the issue of non-stationary since I got theta(1)=0.047192 and theta(2)=0.921872, as stated by the Eviews output the "Stability condition: theta(1) + theta(2) < 1 is met", so the results are solid, right?
However, since the sum of thetas amounts to 0,969064 (very close to 1) i wonder if i shouldn't prefer a IGARCH-DCC estimation? I believe this add-in does not support that type of model estimation just Garch, Tarch and Egarch, am I correct?

Just one more thing, regarding the error distribution option of the add-in, should i test the ARMA residuals for normality and then select if normal or t-student's?

Many thanks for helping me trubador!

Re: DCCGARCH11

Posted: Mon May 04, 2015 1:06 pm
by trubador
Stability condition is just one of the diagnostics and it simply indicates that the estimated model is not explosive. It does not confirm that your model is robust.

The number 0.969 may appear to be "very close" to 1, but in the end it is the standard deviation to decide. And for large samples it may well be significantly different than 1. The standard errors in the output will give you some hint, but you can perform a hypothesis test to check it. Unfortunately you'll have to do it manually as no such option is available at the moment.

And yes, it is always safer to do the diagnostic test (if available) first and then decide what to do next in light of its results.

Re: DCCGARCH11

Posted: Mon May 04, 2015 3:22 pm
by pedpereira
trubador wrote:Stability condition is just one of the diagnostics and it simply indicates that the estimated model is not explosive. It does not confirm that your model is robust.

I believe I have to test the residuals of each univariate garch for (1) arch effects and (2) autocorrelation and also the residuals of garch dcc model for (3) normality, in order to test robustness, am i wrong?
trubador wrote: The number 0.969 may appear to be "very close" to 1, but in the end it is the standard deviation to decide. And for large samples it may well be significantly different than 1. The standard errors in the output will give you some hint, but you can perform a hypothesis test to check it. Unfortunately you'll have to do it manually as no such option is available at the moment.

Coefficient Std. Error z-Statistic Prob.
theta(1) 0.047192 0.014863 3.175180 0.0015
theta(2) 0.921872 0.039586 23.28799 0.0000

The coefficients are significant but i'm not sure about the hint that the SE is giving, can you shed some light please?
trubador wrote: And yes, it is always safer to do the diagnostic test (if available) first and then decide what to do next in light of its results.

Are you referring to the normality of the residuals of the univariate garch or the ARMA residuals?

I'm a bit confused and could really use your help regarding the validity of the dcc approach using this add-in. Should I place my doubts under other more adequate topic? Should i upload my workfile?
Thanks (again)

Re: DCCGARCH11

Posted: Tue May 05, 2015 1:53 am
by trubador
pedpereira wrote:I believe I have to test the residuals of each univariate garch for (1) arch effects and (2) autocorrelation and also the residuals of garch dcc model for (3) normality, in order to test robustness, am i wrong?

Yes, usual diagnostics.
pedpereira wrote:The coefficients are significant but i'm not sure about the hint that the SE is giving, can you shed some light please?

Significance is a test against 0. Assuming coefficents are approximately normal, your hypothesis test against 1 can be formulated as follows:

Code: Select all

((0.92+0.047)-1)/@sqrt(0.0395^2+0.0149^2)

Which indicates that the sum of coefficents is significantly different from 1.
pedpereira wrote:Are you referring to the normality of the residuals of the univariate garch or the ARMA residuals?

GARCH residuals.
pedpereira wrote:I'm a bit confused and could really use your help regarding the validity of the dcc approach using this add-in. Should I place my doubts under other more adequate topic? Should i upload my workfile? Thanks (again)

DCC add-in does not include any diagnostic tools specific to this type of models, like EViews' regular objects. You need to figure it out on your own. If you have any doubts on DCC-type models, the add-in is not the right place to start. You need to study the background first.

This topic pretty much includes all of my suggestions for the possible problems that you may encounter during estimation. You can always refer to previous discussions...

Re: DCCGARCH11

Posted: Tue May 05, 2015 3:13 pm
by pedpereira
trubador wrote: DCC add-in does not include any diagnostic tools specific to this type of models, like EViews' regular objects. You need to figure it out on your own. If you have any doubts on DCC-type models, the add-in is not the right place to start. You need to study the background first.


Thanks for your insights trubador, I am trying hard to understand the model more deeply.

I thought that if applied the usual diagnostics to the garch univariate residuals, I could rely on the dcc estimation output of the add-in, but i suppose the best is to test the dcc model itself then.

If i choose normal distribution the step 1 garch univariate estimation does not converge (after 1000 iterations) for one of the series and the arch term is not significant, which i believe means that the model is a GARCH (0,1) but the add-in gives the output of the dcc anyway (and values do seem reasonable: alpha=0,052 and Beta=0,92)


Example: output with t-student distribution is this:

Coefficient Std. Error z-Statistic Prob.

theta(1) 0.061770 0.040861 1.511716 0.1306
theta(2) 0.347048 0.443426 0.782651 0.4338

t-Distribution (Degree of Freedom)

theta(3) 3.649305 0.148316 24.60491 0.0000

Log likelihood 7872.354 Schwarz criterion -10.82304
Avg. log likelihood 2.722114 Hannan-Quinn criter. -10.85277
Akaike info criterion -10.87048

* Stability condition: theta(1) + theta(2) < 1 is met.

The values of Alpha and Beta don't seem in line with I've seen (typically Beta is around or higher than 0.90). The p-value for both coefficient is >0,05 so they aren't significant does this mean that the dcc model is not valid under the assumption of t distribution of the errors?

Trubador I'm sorry to bother you but your help is invaluable, and believe me i'm trying really hard to find the answers.

Re: DCCGARCH11

Posted: Sun May 10, 2015 6:19 am
by trubador
Please read one of my earlier posts in this thread: viewtopic.php?f=23&t=9677&start=15#p36929
If you look closely to the results from your first garch model, you'll see that the parameters violate the stability condition. Moreover, the garch effect does not seem to be particularly strong. Results suggest that the residuals are not stationary and are therefore not appropriate for garch analysis. You should really make sure that residuals are stationary, before moving forward. After that, you can also try IGARCH and variance targeting to see it they fit into your data better.

Re: DCCGARCH11

Posted: Wed May 13, 2015 2:32 pm
by pedpereira
Thanks for your answer Trubador. Unfortunately I could not overcome the problem...only assuming normal distribution the dcc parameters seem to have reasonable values but the problem is that the first uni-variate GARCH does not converge and therefore I can't really rely on these results, can I? (I've uploaded the updated eviews file)

Regarding the residuals of the ARMA models, I am sure they are stationary however regarding the auto correlation I'm not so sure...I executed the Breusch-Godfrey Serial Correlation LM Test (p-value = 0,215) and the Q stat and the latter has given me some doubts:

AC PAC Q-Stat Prob
1 0.002 0.002 0.0052
2 0.025 0.025 0.8838
3 -0.054 -0.054 5.1666 0.023
4 -0.002 -0.002 5.1716 0.075
5 0.006 0.009 5.2320 0.156
6 0.024 0.021 6.0385 0.196
7 0.036 0.035 7.9245 0.160
8 0.028 0.027 9.0455 0.171
9 -0.011 -0.011 9.2284 0.237
10 -0.054 -0.052 13.446 0.097

Despite the "smallness" of all the values, only lag 3 falls below 5% (I am evaluating Q at 10 lags wich is below 10%) but I believed that just one lag below 5% was not critical ...so do you think that auto correlation of residuals could be the source of the dcc estimation problem given these results?

I suspect I have to try the Igarch model since I have seen similar studies using Igarch estimation, so perhaps that's the way. Should I try to make the code in Eviews or do you suggest any other software?

Re: DCCGARCH11

Posted: Thu May 14, 2015 1:13 am
by trubador
There is nothing you can do, but continue experimenting. These are complex methods and are sensitive to the data under study. You cannot expect them to produce the desired output each time you run the model. You should also consider the fact that your data might not have a dynamic correlation after all.