Demean or not to demean

Hi all

I was hoping someone could shed light or direct me to a resource (or two)
regarding a "demean" question.

As I understand, QMLEs estimated on "demeaned" log return data vs straight
"log return" data behave quite differently in finite samples (particularly
for nonlinear MA models where the MA parameter is of interest). Apparently,
for linear AR models, demeaning data does not seriously affect estimation
of non-intercept parameters (refer: Yong Bao "Should We Demean Data?").

For monthly financial log return data, I find ARMA specifications are not
significant, but some sample *means *ARE significant, while others are not.
In either case, I add the GARCH model specification with various error
distributions from the "rugarch" package.

Code example:
x.log.ret = diff(log(price.x) #i.e. not "demeaned"
spec <- ugarchspec(variance.model=list(model="sGARCH",garchOrder=c(1,1),
             submodel=NULL,external.regressors=NULL,variance.targeting=F),
             mean.model=list(*armaOrder = c(0,0)*,* include.mean = T*,
external.regressors=NULL),
             distribution.model="std")
tempgarch <- ugarchfit(spec=spec,data=x.log.ret,solver="hybrid")

I work through the steps necessary to fitting a *t*Copula from which to
simulate and ultimately work my way back to simulated returns.

The goal here is to extract from the matrix of simulated returns those
groups of returns coinciding with certain pre-determined "scenarios". These
are then used for portfolio optimization.

In the "global respect" of the methodology, can anyone shed light on the
merits/demerits of not first demeaning the data? I haven't found any
glaring problems, but it bothers me that the "rugarch" package operates on
demeaned data.

Hmmm, I did, initially (to R-sig-finance email address).

You then replied (which I thought would address the R-sig-finance group),
and me to your reply (also thinking it would address the group).

I'll double check..

On Wed, Aug 12, 2015 at 10:34 AM, alexios ghalanos <alexios at 4dscape.com>
wrote:

You should post your question to the mailing list.

A.

On 12/08/2015 09:30, Gareth McEwan wrote:

Sorry to keep bothering you Alexios (I know you're busy).

Referring to my most recent response: if one specifies "include.mean =
TRUE" in "ugarchspec", does the "ugarchfit" function use that
(unconditional mean?) estimate to center/demean the innovations?

I'm trying to understand how the "include.mean = T" is used in fitting
the model.

Many thanks
Gareth


On Tue, Aug 11, 2015 at 1:22 PM, Gareth McEwan <mcewan.gareth at gmail.com
<mailto:mcewan.gareth at gmail.com>> wrote:

    Hi Alexios

    Is it correct then to say that by specifying "include.mean = TRUE"
    in "ugarchspec", the "ugarchfit" function uses that (unconditional
    mean) estimate to demean the data before subsequently fitting the
    GARCH portion of the model?

    Many thanks
    Gareth



    On Tue, Aug 11, 2015 at 12:02 PM, alexios <alexios at 4dscape.com
    <mailto:alexios at 4dscape.com>> wrote:

        On 11/08/2015 10:33, Gareth McEwan wrote:

            Hi all

            I was hoping someone could shed light or direct me to a
            resource (or two)
            regarding a "demean" question.

            As I understand, QMLEs estimated on "demeaned" log return
            data vs straight
            "log return" data behave quite differently in finite samples
            (particularly
            for nonlinear MA models where the MA parameter is of
            interest). Apparently,
            for linear AR models, demeaning data does not seriously
            affect estimation
            of non-intercept parameters (refer: Yong Bao "Should We
            Demean Data?").

            For monthly financial log return data, I find ARMA
            specifications are not
            significant, but some sample *means *ARE significant, while
            others are not.
            In either case, I add the GARCH model specification with
            various error
            distributions from the "rugarch" package.

            Code example:
            x.log.ret = diff(log(price.x) #i.e. not "demeaned"
            spec <-

 ugarchspec(variance.model=list(model="sGARCH",garchOrder=c(1,1),

submodel=NULL,external.regressors=NULL,variance.targeting=F),

                         mean.model=list(*armaOrder = c(0,0)*,*
            include.mean = T*,
            external.regressors=NULL),
                         distribution.model="std")
            tempgarch <-

ugarchfit(spec=spec,data=x.log.ret,solver="hybrid")

            I work through the steps necessary to fitting a *t*Copula
            from which to
            simulate and ultimately work my way back to simulated

returns.

            The goal here is to extract from the matrix of simulated
            returns those
            groups of returns coinciding with certain pre-determined
            "scenarios". These
            are then used for portfolio optimization.

            In the "global respect" of the methodology, can anyone shed
            light on the
            merits/demerits of not first demeaning the data? I haven't
            found any
            glaring problems, but it bothers me that the "rugarch"
            package operates on
            demeaned data.


        The rugarch package does NOT operate on demeaned data. It offers
        the option (default=TRUE) through "include.mean" on whether to
        demean the data or not. In case you are not demeaning, then the
        data are passed straight to the GARCH routine and assumed as the
        zero-mean residuals.



            Thank you very much for the help
            Gareth


        Alexios