[R-SIG-Finance] estimation difference between ugarchroll and ugarchfit‏

[alexios ghalanos]

Hi Guoshi,

I've posted a fix on r-forge:
(see https://r-forge.r-project.org/projects/rgarch/).

Its version 1.0-9 revision 401,  and should be available to download 
once the check/build cycle completes.

Regards,
Alexios

On 24/05/2012 00:18, linpack wrote:
> Hi Alexios,
>
> Thanks a lot for your reply and also your enormous effort on the
> development of this pakage.
>
> 1. I see the reason now: I forgot to reset the specification in each
> step of the loop.
> Hope others who see this wont make such mistake anymore.
>
> 2. Really appreciate that, I guess I should just use
> myfit = ugarchfit(spec,
> logexr[(i*refit.every):(fitindx.start[i]+refit.every)], out.sample =
> refit.every, solver.control=list())
> ldpx=length(dpx)
> ugarchforecast(myfit, n.ahead=1, external.forecasts = list(mregfor =
> dpx[ ldpx], vregfor = dpx[ ldpx ]) )
> # where dpx[ldpx] is the last element of dpx which is already supplied
> to R when out.sample =1
>
> Thank you again.
>
> Best regards
> Guoshi
>
>
> At  2012-05-24  04:05:55,"alexios  ghalanos"  <alexios at 4dscape.com>  wrote:
>>Hi  Guoshi,
>>
>>1.
>>#################################
>>#  part  of  the  code  is  straight  out  of  the  'rugarch-rolling.R'  file
>>#  which  I  suggest  you  take  a  closer  look  at.
>>forecast.length  =  400
>>refit.every  =  1
>>n.ahead  =  1
>>#  your  data  size
>>T  =  750
>>fwindex  =  t(rugarch:::.embed((T  -  forecast.length  -  n.ahead  +  2):T,
>>refit.every,  by  =  refit.every,  TRUE  )  )
>>fitindx.start  =  c(fwindex[1,]-1)
>>fwindex.end  =  fwindex[refit.every,]
>>nf  =  length(fwindex.end)
>>forecast.length  =  floor(forecast.length/refit.every)  *  refit.every
>>####################################
>>
>>#  Do  10  fits  to  convince  yourself  it  is  the  same:
>>
>>lik  =  rep(0,  10)
>>for(i  in  1:10){
>>dpx  =    matrix(dp[(i*refit.every):(fitindx.start[i]+refit.every)],  ncol=1)
>>spec  =  ugarchspec(variance.model  =  list(model  =  "eGARCH",  garchOrder  =
>>c(1,1),
>>external.regressors  =dpx  ),
>>mean.model  =  list(armaOrder=c(0,0),  include.mean  =  T,
>>external.regressors  =  dpx),  distribution.model  =  "norm");
>>myfit  =  ugarchfit(spec,
>>logexr[(i*refit.every):(fitindx.start[i]+refit.every)],  out.sample  =
>>refit.every,  solver.control=list())
>>lik[i]  =  likelihood(myfit)
>>}
>>
>>#  use  your  the  'uGARCHroll'  object  to  compare:
>>all.equal(lik,  test at roll$LLH[1:10])
>>[1]  TRUE
>>
>>
>>2.  Correct.  The  rolling  method  does  not  take  into  account  the  presence
>>of  external  regressors...most  likely  an  omission  on  my  part  which  I'll
>>look  to  rectify  during  the  next  release.  Do  keep  in  mind  that  the
>>ugarchroll  method  is  just  a  convenience  wrapper  for  fitting/forecasting
>>with  the  existing  methods.
>>
>>
>>-Alexios
>>
>>
>>
>>On  23/05/2012  19:54,  linpack  wrote:
>>>  Hi  there,
>>>
>>>  I  have  two  questions  (which  might  be  related).
>>>  I  am  trying  to  predict  mean  and  variance  of  monthly  stock  excess  return
>>>  using  dividend  price  ratio  with  a  eGarch  model.
>>>
>>>  Data  format  is  the  following
>>>  column  1  column  2  column  3
>>>  gross  S&P  return  gross  t-bill  return  lagged  dividend  price  ratio
>>>  #data  dimension  is  nrow=750  and  ncolumns  =3  and  the  "redp.dat"  file  is
>>>  attached.
>>>
>>>  code  is  following
>>>  library(rugarch);
>>>  dat=read.csv('C:\\redp.dat',header=F);
>>>
>>>  logexr=log(dat$V1)-log(dat$V2);  #  log  of  gross  excess  return
>>>  as.matrix(logexr);
>>>  dp=log(dat$V3);  #log  of  dividend  price  ratio
>>>  as.matrix(dp);
>>>
>>>  spec  =  ugarchspec(variance.model  =  li!  st(model  =  "eGARCH",  garchOrder  =
>>>  c(1,1),  external.regressors  =  as.matrix(dp)),
>>>  mean.model  =  list(armaOrder=c(0,0),  include.mean  =  T,
>>>  external.regressors  =  as.matrix(dp)),  distribution.model  =  "norm");
>>>  #  dividend  price  ratio  are  used  as  external  regressor  in  both  mean  and
>>>  variance  equation
>>>
>>>  test=ugarchroll(spec,  n.ahead=1,  data=logexr,  forecast.length=400,
>>>  refit.every=1,  refit.window="moving");
>>>  as.data.frame(test,  which='LLH')
>>>  #  where  I  did  a  rolling  window  forecast  and  display  all  the  likelyhood
>>>
>>>  #then  I  did  a  double  check  on  the  rolling  estimation  by  doing
>>>  for  (j  in  1:10)  #  I  am  justing  printing  the  first  ten  loop  step
>>>  myfit  =  ugarchfit(spec,  logexr[j:349+j]);
>>>  likelihood(myfit)
>>>  end
>>>
>>>  #  Question  1:  the  first  likelihood  from  ugarchfit  exactly  equals  that  in
>>>  ugarchroll,  while  the  values  are  different  from  then  on.
>>>  #  When  I  modify  the  spec  to  exclude  any  external  regressor,  the  two  sets
>>>  of  likelihood  are  then  exactly  the  same.
>>>  #  reducing  the  forecast  length  (and  therefore  increase  the  length  of  the
>>>  estimation  window)  seems  to  mitigate  the  estimation  difference.
>>>
>>>  #  Question  2:  I  then  compare  the  density  forecast
>>>  as.data.frame(test,  which='density')
>>>  #  this  seem  to  produce
>>>  for  (j  in  1:10)  #  I  am  justing  printing  the  first  ten  loop  step
>>>  ugarchforecast(myfit,  n.ahead=1);  #  and  again  when  the  likelihood  are
>>>  different,  the  forecast  is  different.
>>>  end
>>>  #  however,  I  am  a  bit  confused  since  I  thought  it  should  produce
>>>  for  (j  in  1:10)
>>>  ugarchforecast(myfit,  n.ahead=1,external.forecasts  =  list(mregfor  =
>>>  dp[350+j],  vregfor  =  dp[350+j]  ));
>>>  end
>>>
>>>  looking  forward  to  your  help.
>>>
>>>
>>>  Guoshi
>>>
>>>
>>>
>>>
>>>
>>>  _______________________________________________
>>>  R-SIG-Finance at r-project.org  mailing  list
>>>  https://stat.ethz.ch/mailman/listinfo/r-sig-finance
>>>  --  Subscriber-posting  only.  If  you  want  to  post,  subscribe  first.
>>>  --  Also  note  that  this  is  not  the  r-help  list  where  general  R  questions  should  go.
>>
>
>
>