Two sets of std. residuals in mgarch estimation...

[kula]

Hi,
I try to estimate two variable BEKK MGARCH model. The related wizard lets us save the standardized residuals under a name, say "xx". Following the estimation of the model, when I intend to apply @MVARCHTEST to the standardized residuals, I see that there are two sets of std residual series, namely "xx(1)", and "xx(2)" appearing under "View/Series Window". Performing mvarchtest on both gives conflicting results, rejecting the null on the former, while nonrejection on the second. Which one should be picked as standardized residuals? Or, is this not the correct way of applying mvarchtest?

This is the way I apply mvarchtest following the estimation of the model by means of the wizard:

@mvarchtest
# xx(1)

Many thanks for your replies in advance...

[TomDoan]

No. XX is a VECT[SERIES] will the full set of jointly standardized residuals. You want to do

@MVARCHTEST
# xx

not separate tests on each. See the section on diagnostics in the User's Guide.

[kula]

Hi Tom,
Many thanks for the kind reply. This time, I have performed as in the User's Guide. The attached output reads that both multivariate q test and multivariate arch test are significant. Does this mean that the model is not a fit, proper one? If so, what should be done?

bekk output.RPF

(2.44 KiB) Downloaded 885 times

[TomDoan]

Until you clean up the serial correlation, it's not clear whether the BEKK model is adequate. You should probably look at a mean model with lags (like a VAR(1)) to see if you can do something to fix the significant Q.

[kula]

Hi Tom,
As for BEKK model, the Q test turns out to be insignificant at the 12th lag (until 12th, it is significant), though the mvarch test result is still significant. Now, does the model seem to be proper, fit this time?

12th lag bekk model.RPF

(20.96 KiB) Downloaded 961 times

[TomDoan]

What kind of data are those? The mean model for the HACADET looks just completely wrong for anything that would be fit by a GARCH model.

[kula]

Hi Tom,
HACADET is for the log value of the daily transaction volume change (in lots) in BIST100 index. It appears that, in case of using several lags, the mean value is significantly affected from the several lags in row. Many thanks for your comments.

[TomDoan]

The concern I would have is that there's some type of timing issue with those two series. For (by far) the most significant coefficient in the VAR to be the first "cross" lag looks a lot like that has contemporaneous rather than lagged information.