On 26/07/2012 16:34, stoyan.stoyanov wrote:
> Hi all,
>
> A question for Alexios or someone else who has some experience with rugarch.
> I am having a hard time understanding the functionality of the ugarchfilter
> class in the rugarch package. My understanding is that you should be able to
> use it to apply the estimated coefficients of an old fit to a dataset
> including new points (fixing the already estimated parameters and only
> adding the effect of the new data points to arrive at updated parameter
> estimates). I am not entirely sure that this is correct.
No. It does not "arrive at updated parameter estimates". The
ugarchfilter method simply "filters" the new dataset with the existing
estimated parameters to generate an "updated" conditional mean and
variance. It is exactly like performing a 1-step ahead rolling forecast
(since the filter generates t+1|t..)
>
> Either way, the new resulting object produces a list of coefficients (among
> other things) which are different from the fit coefficients.
That is NOT what happens. It does not estimate any new coefficients so
the list of coefficients is EXACTLY what you provided through the
uGARCHspec object (with fixed parameters).
> If i am
> correct, these are the coefficients I want to be using for forecasting when
> I have added a single data point to my series and I am looking for speed.
> However, the ugarchforecast and ugarchboot functions can only be applied to
> ugarchspec and ugarchfit classes.
You are not forecasting coefficients. You are forecasting the
conditional mean and variance.
> Does that mean that in order to forecast, I need to transform my filter
> object into a spec object and then into a fit object (through using the
> fixed.pars method in ugarchspec)?
> I get the feeling that this is a bit complicated to be true, so I am
> probably misunderstanding the functionality intended for the filter
> function.
Definitely misunderstanding it. I hope the previous comments clarify.
You forecast from either a uGARCHfit-class object or a uGARCHspec-class
object (with fixed parameters).
> Also, any additional clarification on the n.old method would be appreciated
> - I am having a hard time interpreting the explanation in the package
> documentation (which might be owed to my faulty interpretation of the
> object's functionality).
>
GARCH models require an estimate of the t_0 variance (usually the mean
of the squared residuals) to initialize the recursion. If you are
providing a dataset (length T) which is simply the old dataset (length
N) augmented by new data (length M), then you might want to use the
original t_0 value to filter the data (n.old=N) so that:
1. You do not have look ahead bias by using n.old=T (the bias is usually
very small, but it does not hurt to control for it).
2. Conditional variance values agree exactly (used internally in a
variety of cases where this is important to control for).
- Alexios
_______________________________________________
[hidden email] 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.
Locked
