ROBUSTERRORS

[jimm306]

Hi Tom,

I know we select the robusterrors option for a BEKK-GARCH (or other) model estimation assuming a Normal distribution (QMLE).

When estimating a BEKK-GARCH model with a t-distribution, do I need to select the robusterrors option?

[TomDoan]

You don't need to. You can if you want. It's probably less common to use ROBUSTERRORS when you do t-distributed errors since it explicitly models the failure of conditional Gaussianity.

[jimm306]

When we estimate VIRFs through simulations we should check the stability, eigenvaues of %%vech_a+%%vech_b+%%vech_d? Because although the eigenvalues are (1.050,-0.000) (1.029,-0.000) (1.005,0.000) the volatility impulse responses have no explosive behaviour.

Thanks a lot

[TomDoan]

If you run the VIRF's long enough, the 1.029 would eventually cause (slowly) "explosive" behavior.

[jimm306]

I have four assets returns with the goal of implementing VIRFs.

1) The four lag selection criteria agree that i should have a constant mean, but after estimating the model the multivariate Q statistic is significant. So, i choose a VAR(3)-BEKK GARCH model that successfully eliminates autocorrrelations and lagged cross correlations. Is it a valid move?

2) Also, the prices of these assets are cointegrated. Should i implement a VECM-GARCH model?

3) What is the difference between t distribution and ged? when i use ged usually i get better results regarding the elimination of heteroskedasticy in the residuals and stability of the model. It's just which one make the job?


Thanks a lot for taking the time to consider my questions.

[TomDoan]

I'm very confused about the combination of #1 and #2. If the series are cointegrated, then they have to have a (very) different mean model than just a constant. Was your first go at estimating models on first differences (returns)? If so, and the log prices are cointegrated, then yes, the overall model is misspecified and you would expect that a multivariate Q would pick up the cross correlations in the residuals. A VAR *can* cover the types of cross variable dynamics from a VECM, but if you have a strong reason to believe the series are cointegrated, then a VECM-GARCH (a VECM being a constrained VAR) would be appropriate.