Skip to content

MAR-ARCH

6 messages · Jeff Ryan, Ajay Shah, Dirk Eddelbuettel +1 more

Original
Jeff Ryan
# 
While Dirk is correct in that the mgarchBEKK is _NOT_ MAR-GARCH, it
_IS_ an mgarch implementation... (which is worth something)

And I did get it to install and run with two fixes.  I have NO idea if
the results it produces are worth anything (and it chews up resources,
so I won't be spending any time to figure out).

But for those interested - windows users will need to adapt the
following (or switch to unix/linux/mac)...actually it *may* only work
on *nix boxes.


1. Download the most recent version from

http://www.vsthost.com/vstDocs/mgarchBEKK/release/mgarchBEKK_0.07-8.tar.gz

2. untar/zip
tar xzf mgarchBEKK_0.07-8.tar.gz

3. Fix _at_least_:
mvBEKK.est
line 475 has a trailing comma in the list of returned values after
residuals. - get rid of

mvBEKK.sim
line 252 has a trailing comma in the list of returned values AS WELL.
- just delete it

4. tarball back up for R CMD
tar czf mgarchBEKK_0.07-8.tar.gz mgarchBEKK

5. R CMD INSTALL mgarchBEKK_0.07-8.tar.gz


and happy mgarchBEKKing...

Jeff

My guess is it was breaking because previous R versions *may* have
been more lenient with the extra comma.  Too lazy to further
investigate though : )
On 10/21/07, Ajay Shah <ajayshah at mayin.org> wrote:
Ajay Shah
# 
Jeff,

Thanks a lot! The mistakes you have identified (line 475 and line 252)
are obvious syntax errors. It's a wonder that R tolerated them
earlier.

As is well known with MV Garch models, there are difficulties with
convergence even with the example supplied with mvBEKK.est().

What will be great is if someone with access to other MVGARCH codes
will compare and contrast what mgarchBEKK reports for some standard
problems. E.g. imagine a recent 1000-day time-series for USD/EUR,
USD/JPY and USD/GBP. All three series are easily accessible from the
US Fed website. It would be great to have reference estimates for what
results come out for this problem from a few different codes with
different starting values and algorithms. Unfortunately, I don't have
any other MVGARCH codes so I'm not able to do this experimentation.
Patrick Burns
# 
Comparing estimators is a good idea.  But a good comparison
is more complex than stated.

What matters with garch is prediction, not estimation.  So we
want to test how good the predictions are.  For univariate we
can just compare the realized variance to the predicted variance
for a number of time periods.  It is best if this includes periods in
different volatility regimes.

For multivariate the comparison of realized to predicted is more
complex.  A couple of choices are to look at the maximum absolute
value of eigenvalues of the difference.  Another is to look at the
minimum variance portfolio.

If you do such testing, it becomes apparent that 1000 daily observations
is about as small as you want to go.  2000 observations is better, and even
more is probably good.


Patrick Burns
patrick at burns-stat.com
+44 (0)20 8525 0696
http://www.burns-stat.com
(home of S Poetry and "A Guide for the Unwilling S User")
Ajay Shah wrote:

            

      
      
      
    

  
  


  


      

    

    

      
      

        


  

        

      Ajay Shah
    

    

      
      #
    

  


  

      
On Tue, Oct 23, 2007 at 09:50:49AM +0100, Patrick Burns wrote:

            

      
      
      
    

        
I'm sorry I was not clear. For starters, I was only after software
testing. Does this code replicate the numerical values obtained for
standard datasets with standard codes?

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Dirk Eddelbuettel
    

    

      
      #
    

  


  

      
Hi Ajay,

        
On 23 October 2007 at 19:47, Ajay Shah wrote:

        
| On Tue, Oct 23, 2007 at 09:50:49AM +0100, Patrick Burns wrote:

        
| > Comparing estimators is a good idea.  But a good comparison
| > is more complex than stated.
| 
| I'm sorry I was not clear. For starters, I was only after software
| testing. Does this code replicate the numerical values obtained for
| standard datasets with standard codes?

Well are there standard datasets and results for volatility estimation?

It's been a (longish) while since I looked closely at this, but isn't
volatility still an unobservable?  Short of a Monte Carlo study with metrics
such as the ones suggested by Pat, what do you suggest one looks at?
Realized vol?  Implied vol?  "Traded" vol from variance or vol contracts?

I'm sure there are good answers to be had for this, so let's hear them :)

Dirk

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Patrick Burns
    

    

      
      #
    

  


  

      
Ajay is talking about something different than what you (Dirk)
are and I was talking about.

The latter topic is: how good does my model capture the
financial phenomenon in which I'm interested.

Ajay's topic is: I have a model and at least two implementations
of that model.  I don't care how good it is for interpreting reality,
but what I do care about is if I can say anything about the quality
of the implementations.

Garch is a particularly fertile ground for the second question.  It
is probably somewhat of an exaggeration, but there may be problems
for which you get a unique answer from each implementation you try.
Even for univariate garch(1,1) assuming Gaussian errors.

As for standards on this, Bruce McCullough wrote a paper on garch
implementations, and Dietmar Maringer and Peter Winker wrote a
paper on the difficulty of getting the optimal estimate.

Pat

        
Dirk Eddelbuettel wrote: