DCC bootstrap

[fabio fornari]

Hi Tom
I saw there is a code to bootstrap or forecast from mvgarch, but not dcc for obvious reasons.
However, would you have anything ready, based on sensible assumptions on how to forecast from dcc as for example in the original engle sheppard paper, to be used to bootstrap from dcc, i mean taking a string of past errors of length k and run the dcc forward to t plus k getting simulated returns (whatever is in the mean equations) and correlation matrix (i.e. variances and covariances)?
thanks a lot for any help on this topic

[jones381]

I was also wondering if there is a way to bootstrap confidence intervals in a MVGARCH DCC model? Any help is greatly appreciated.

[TomDoan]

An example of bootstrapping with multivariate GARCH models is provided at http://www.estima.com/forum/viewtopic.php?f=8&t=1605.

[jones381]

Thanks so much for your help. Any suggestions on how to modify the code to produce confidence intervals around time-varying correlation coefficients?

[TomDoan]

It would be very, very, very time-consuming. Inside the bootstrapping loop you would have to generate a full data set, not just 10 data points, re-estimate the model, apply the new coefficients to the original data to get the bootstrapped dynamic correlations. The re-estimation of the GARCH model with each bootstrapped sample is the part that takes the most time.

[jones381]

Tom,

Thanks again for all your help. I modified the code but I'm just estimating over the same data set multiple times. How do I get my model to estimate over a different data set?

[TomDoan]

It's a lot more complicated than what you're doing. Because the DCC correlation is associated with the specific ordering of the original data, you have to bootstrap a data set, which needs to have the proper mean model as well, use it to estimate a new set of GARCH coefficients, then evaluate the DCC GARCH model on the original data set with the bootstrapped coefficients. I would recommend doing the Metropolis-Hastings method shown in:

http://www.estima.com/forum/viewtopic.php?f=8&t=1591

It will probably be over 100 times faster since it just has to do one function evaluation per sweep rather than needing full estimation of the GARCH model for each bootstrap sample.