[EMAIL PROTECTED] (punky brewster) wrote in message
news:<[EMAIL PROTECTED]>...
> Hi all,
>
> I am constructing a forecasting model for auto sales with predictors
> like advertising spend and macroeconomic conditions. But, I know that
> auto sales have both month-to-month and seasonal trend. What is the
> best way to model this while accounting for the trend and seasonality?
>
> Can I do this?
>
> Y(t) = Y(t-1) + Y(t-12) + predictor1... + e
>
> If not, what would you suggest?
I would definitely not be guessing at the model form or trying to
select a best model from a candidate set of models. What you should do
is to identify a Transfer Function Model with ARIMA structure and
Intervention Variables to accomodate pulses, seasonal pulses ,level
shifts and/or local time trends.
What you don't know is :
1. What lags to put in for the dependent series
2. What lead , contemporaneous , lag structure is needed for each
input series
3. What the effect is of omitted stochastic series whose effect can be
proxied by an efficient ARIMA structure
and
4. What the effect is of omitted deterministic series whose effect can
be proxied by a set of empirically identified Intervention Variables.
Your final model will have an error process [a(t)] that will have(
should have ! )no significant autocorrelations and furthermore will be
unpredictable by any lag of any of your specified input series.
Additionally this error process will have a constant mean throughout
time ( i.e. no intervention variables ) and a constant variance
throughout time.
This model may have the following components
y(t) = constant
+ coeffa(1)*y(t-k) + coeffa(2)*y(t-k-1) ....etc...
+ coeffb(1)*a(t-1) + coeffb(2)*a(t-2) ....etc...
and for each known input series x
+ coeffc(1)*x(t+k) + ... + coeffc(2)*x(t-1) ....etc...
and for each empirically identified intervention variable
+ coeffd(1)*I(t) where I(t) can be a pulse,seasonal
pulse,level shift or local time trend.
Hope this helps....
Dave Reilly
AFS Co.
215-675-0652
P.S. by the way you can go to DOWNLOAD.COM , search on FreeFore and
download a free piece of software called FreeFore which will do this
automatically , at no charge. The program ends up with an equation
where all the parameters are significant and the error process is
Gaussian. Couldn't be better !
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
