VECM-GARCH Model

[TomDoan]

Just change the MV option on the GARCH instruction---you don't have to make any change to the VECM mean model.

[faaequah13]

i did this but along with the output table i found one more msg which is as(## OP9. Options VECHMAT and your choice for MV May Not Be Used Together) can you help me what is this?

[TomDoan]

That's correct. CC and DCC do not admit a VECH representation. See the description in the User's Guide. 9.4.1 describes the MV options which are restricted VECH and 9.4.2 describes the ones that aren't.

Note that that is just a warning, and has no effect on the GARCH instruction itself---it just warns that you can't use the output from the VECHMAT option for further calculations.

[faaequah13]

thank you Tom Doan, i have one more query as my variable are integrated at different level and to test the co integration i had applied ARDL model, and then I found co integration and then applied VECM model and now i need to apply garch bekk model, all i have done in eviews, now i am using rats for garch model, can you help me out how can i do? am i need to run ardl, vecm and then garch? please help me.

[TomDoan]

An ARDL isn't helping you---you need a model where all the variables are endogenous. The VECMGARCH example shows how to do a GARCH model with a VECM mean model properly. Just fix the options on GARCH to match what you want.

[faaequah13]

it mean i have to start my model with the same codes (@johmle(lags=10,det=rc,cv=cvector) as u share VECMGARCK rpf file. and then carry on with garch bekk model.

[TomDoan]

thank you Tom Doan, i have one more query as my variable are integrated at different level and to test the co integration i had applied ARDL model, and then I found co integration

What does that mean? If you have an I(1) and an I(0) variable, then they cannot (basically by definition) be cointegrated.

At this point, I think it makes sense for me to ask you to attach your program and data. I'm not seeing where you're getting the model you are out of the data that you describe.

[faaequah13]

as i told u i am working on bivariate data on different companies closing price and currency exchange rate risk, among 10 companies four are stationary at level and exchange rates are stationary at first difference. in this case i cant run johansen cointegration beacuse for johansen all the series must be stationary at first difference. because of this limitation i used ARDL model to extract cointegration, and i found that the series are cointegrated. then i came to VECM model on those which are cointegrated. for the volatility purpose i want to run garch bekk model. now i am stuck how run garch bekk model for ARDL cointegration. as u shared with me johansen vecm then garch bekk model. i think now you get a point.

[TomDoan]

as i told u i am working on bivariate data on different companies closing price and currency exchange rate risk, among 10 companies four are stationary at level and exchange rates are stationary at first difference.

There is no way that (log???) closing price will be stationary. In the unlikely event that they may reject a unit root in a sample, that will be because of sampling error.

in this case i cant run johansen cointegration beacuse for johansen all the series must be stationary at first difference.

That's not true. A stationary series is, in the Johansen methodology, "cointegrated" with itself. If you run @JOHMLE on a pair, one of which is I(0) and one is I(1), you will get a rank of one. It's just with a trivial "cointegrating" vector which includes only the I(0) series (in population---in practice, it will have a small coefficient on the I(1) variable). The VECM representation produced is perfectly valid, and is valid if both series are I(0), if one is I(0) and one I(1), both are I(1) and not cointegrated, or both I(1) and cointegrated. The difference among those is the rank and shape of the PI matrix.

because of this limitation i used ARDL model to extract cointegration, and i found that the series are cointegrated. then i came to VECM model on those which are cointegrated. for the volatility purpose i want to run garch bekk model. now i am stuck how run garch bekk model for ARDL cointegration. as u shared with me johansen vecm then garch bekk model. i think now you get a point.

There is no "ARDL" cointegration. There are ARDL tests for cointegration, but the series are cointegrated or they aren't. And the ARDL representation is worthless for GARCH estimation.

[faaequah13]

according to you, whether my series are stationary at level or first difference i can use johansen test and whatever output is like cointegratied or not cointegrated in that case too i can use VECM model. am i right?

[TomDoan]

according to you, whether my series are stationary at level or first difference i can use johansen test and whatever output is like cointegratied or not cointegrated in that case too i can use VECM model. am i right?

That's correct, except that you need to be careful about interpreting the cointegration test since it will show rank 1 if one of the series is I(0).

[faaequah13]

if i do not get rank 1, does it means some error or simply no co integration.

[TomDoan]

Rank 2 means they're both stationary. Rank 0 means they're I(1) but not cointegrated. Rank 1 means they're cointegrated if they're both I(1), or that one is stationary.

[faaequah13]

hello Sir, during the univariate diagnostic test, set std1 = rr(t)(1)/sqrt(hh(t)(1,1))
## MAT15. Subscripts Too Large or Non-Positive
Error was evaluating entry 4550, I got this error. can u help what is this and how to rectify it

[TomDoan]

Restrict it to the estimation range:

set std1 %regstart() %regend() = rr(t)(1)/sqrt(hh(t)(1,1))