Expost prediction for ARMA or GARCH

Perhaps following thread will be of your benefit:

http://r.789695.n4.nabble.com/Forecasting-GARCH-td930240.html#a930243

Thanks,

View this message in context: http://r.789695.n4.nabble.com/Expost-prediction-for-ARMA-or-GARCH-tp3094155p3094391.html
Sent from the Rmetrics mailing list archive at Nabble.com.

Dear researchers

I would like to ask, if there is a way how to make in R
(fGarch,tseries,..) EXPOST prediction(quasi future). Correct me, if I am
wrong, but the command predict(x,n.ahead) makes EXANTE predictions(true
future prediction) only. In contrast with Eviews, where forecast

function

produces expost results. Another idea, is to divide the testing set into
training and validation sets, estimate for example ARMA coefficients for
training set and use them with validations set to produce expost
predictions. In pseudocode it looks like this:

testing_set (1:100)
training_set(1:90)
validation_set(91:100)

arma_coef<-fit(training_set,~arma(1,1))

expost_prediction<- arma_coef * validation_set

Is this a way how to produce EXPOST predictions, or my approach is
completely wrong? I beleive, there is much easier way how to do it in R,
but I just cant find the satysfying solution by myself. Thank you any

help.

I believe what you are asking is:

"How do I use a pre-set (trained) set of ARMA or GARCH coefficients on a
new data set without changing the model?"

This is not ex post 'prediction'.  The model fitting is ex post.  predict
is *always* using the current model to predict future movements (ex ante)
on observations in the time series that come in after the model is fit.

In fGarch, at least, there is no way to do this.  See a list post from me
approximately one+ year ago asking about predict and pre-fitted values for
fGarch, and Yohan's helpful reply on the modifications that would be
necessary to fGarch.

In any GARCH model, the conditional volatility may be (and usually is)
displayed on the prior data.

for arma(), I don't know the answer, but I suspect it will involve some
coding on your part.

Regards,

  - Brian

Brian G. Peterson
http://braverock.com/brian/
Ph: 773-459-4973
IM: bgpbraverock

On Sun, 19 Dec 2010 04:15:28 -0800 (PST), "Arun.stat"

Perhaps following thread will be of your benefit:

http://r.789695.n4.nabble.com/Forecasting-GARCH-td930240.html#a930243

Arun,

This is precisely the thread I was talking about in my reply.  Thank You.

I did not check after that thread to see if Alexios added the
functionality to the rgarch package or not, but it looks like he did, from
this quote on the rgarch website:

"Forecast (from spec or fit), Simulation (from fit or spec)"
http://rgarch.r-forge.r-project.org/

To the OP's question, this would be forcast from specification.

Regards,

  - Brian

Brian G. Peterson
http://braverock.com/brian/
Ph: 773-459-4973
IM: bgpbraverock

Dear Brian, Arun and Alex

Thank you all  for helping, I am going to check all your suggestions. At the end, I would like to provide few lines of code, just in case, if someone can utilize it in future:

library(quantmod)
library(tseries)
getSymbols("XPT/USD",src="oanda")

# 2 time series, 
x<-XPTUSD
y<-Lag(x,k=1)

length(x)

500

## fitting arma model on first 450 observations,
fit <- arma(x[1:450], order = c(1, 0))
fitted_model<-fit$fitted.values


## lets calculate it all, on a full length, 1:450 are same, 451:500 would be expost
for (i in 1:500)
fitted_manual[i]<-fit$coef[1]+(fit$coef[2]*y[i])

## for comparision, I show, that values fitted by model (fitted_model) are exactly the same as those, that were calculated with coefficients and lagged time series (fitted_manual), what its ##not a suprise :)))

comparision<-cbind(fitted_mode,fitted_manual[1:450]);comparision   ## same

## so I thing, the last 50 values could be considered as expost predictions, because arma coeficients were estimated only on first 450 observations
comparision_expost<-cbind(x[451:500],fitted_manual[451:500]);comparision_expost

plot(comparision_expost[,1],type="l")


Best regards

Dear Brian, Arun and Alex

Thank you all ?for helping, I am going to check all your suggestions. At the end, I would like to provide few lines of code, just in case, if someone can utilize it in future:

library(quantmod)
library(tseries)
getSymbols("XPT/USD",src="oanda")

# 2 time series,
x<-XPTUSD
y<-Lag(x,k=1)

length(x)

500

## fitting arma model on first 450 observations,
fit <- arma(x[1:450], order = c(1, 0))
fitted_model<-fit$fitted.values


## lets calculate it all, on a full length, 1:450 are same, 451:500 would be expost
for (i in 1:500)
fitted_manual[i]<-fit$coef[1]+(fit$coef[2]*y[i])

## for comparision, I show, that values fitted by model (fitted_model) are exactly the same as those, that were calculated with coefficients and lagged time series (fitted_manual), what its ##not a suprise :)))

comparision<-cbind(fitted_mode,fitted_manual[1:450]);comparision ? ## same

## so I thing, the last 50 values could be considered as expost predictions, because arma coeficients were estimated only on first 450 observations
comparision_expost<-cbind(x[451:500],fitted_manual[451:500]);comparision_expost

plot(comparision_expost[,1],type="l")

I tried running the code but it's having trouble early on with the
getSymbols command. Are downloads from Oanda supposed to be free and
open or are they subscription based?

getSymbols("XPT/USD",src="oanda")

[1] "XPTUSD"
Warning message:
In readLines(tmp) :
  incomplete final line found on
'C:\DOCUME~1\Mark\LOCALS~1\Temp\RtmpDJGlT6\file7cac6eba'

The 'LOCALS~1' looks strange as the directory doesn't exist in my
directory. (Unless it's hidden by default.)

- Mark