Skip to content

Forecasting GARCH

5 messages · Cristian Gonzalez, Alexios Ghalanos, Arun Kumar Saha +1 more

Alexios Ghalanos
# 
A valid point...will aim to add this functionality to rgarch over the 
weekend (at the earliest) by extending the forecast method to accept a 
specification and data object instead of only a fitted object.

-Alexios Ghalanos
Cristian Gonzalez wrote:
Arun Kumar Saha
# 
In my best knowledge prediction for garch process was already discussed in
this forum. Why dont you have some search here?

Best,
Cristian Gonzalez wrote:

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Arun Kumar Saha
    

    

      
      #
    

  


  

      
Well, please ignore my previous mail. I have misread that. I assumed you are
talking on the garch prediction. However I think that, what you pointed
there should be valid and learned there should this kind of functionality in
R as well.

Best,

        
Arun.stat wrote:

            

      
      
      
    

        

      
      
    

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Brian G. Peterson
    

    

      
      #
    

  


  

      
Cristian Gonzalez wrote:

            

      
      
      
    

        
Arun was correct, this has been covered before. I asked the question 
regarding fGarch, and Yohan answered. Unfortunately, I have not had time 
to complete working this out for fGarch following his excellent 
suggestion. The contents of our interchange are copied below:

   BGP> I've been continuing to examine the fGarch code, and I think
   BGP> that I can probably do most of what I want by fitting a model,
   BGP> overriding.series, and then calling .garchLLH although I've
   BGP> not yet confirmed that this is the case.

        
Yohan Chalabi wrote:

        
Hi Brian,

overriding .series is probably your best option. Note that .series and
other variables stored in the .fGArchEnv environment used to be global
variables. Moving those global variables to an environment was our best
solution to avoid problems with global variables without modifying to
much code.

library(fGarch)
fit <- garchFit(~garch(1,1), dem2gbp)
ls(all.names = TRUE, envir = fGarch:::.fGarchEnv)

you can use .getfGarchEnv and .getfGarchEnv to retrieve and
set new values in this environment. 

Note that .series is scaled by default in .garchFit(). If you override
.series$x, do not forget to change .series$scale because it will be used
in .garchLLH.

Calling .garchLLH with fGarchEnv = TRUE will update the variables
in .fGarchEnv.


As a side note, there is a handy update method for fGARCH object. You
can re-fit the model with new parameters, for example

update(fit, ~aparch(1,1))

HTH
Yohan

   BGP>
   BGP> I understand completely that I can predict by using something
   BGP> like rollapply or apply.fromstart to repeat garchFit and then
   BGP> predict.
   BGP>
   BGP> However, I think that much of the information in a garch
   BGP> model can be extracted without refitting if we simply want 
   BGP> to calculate the conditional variance without refitting the 
   BGP> model.  predict() would likely also be able to be applied 
   BGP> in this way. 
   BGP> Any confirmation on where to look in the code would be
   BGP> appreciated.  
   BGP> As always, any modified code I work up will be contributed back.


Regards,

- Brian