Date: Thu, 1 Dec 2005 13:53:36 -0500
Reply-To: Sigurd Hermansen <HERMANS1@WESTAT.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Sigurd Hermansen <HERMANS1@WESTAT.COM>
Subject: Re: White Noise Test
Content-Type: text/plain; charset="us-ascii"
I haven't done any serious work in time series analysis for a number of
years, but I would be very surprised if any simple litmus test could
assess whether a model is forecasting correctly. As I read through David
Cassell's warnings about the perils of automated model selection, I see
strong parallels between what is happening now in predictive modelling
and what happened about 20 years ago in forecasting. During the early
1980's, complex econometric forecasting models crashed and burned,
'black box' models such as ARIMA and ARMAX replaced them, and 'out of
sample' testing showed that peculiarities in sample data used to fit
models and 'regime changes' at different points in time made forecasts
chronically inaccurate. By the late 1980's, forecasters found that
aggregates of other forecasts produced more consistently accurate
forecasts than did individual forecasts. I lost interest at that point.
The fact that you have not mentioned the purpose and content of your
forecasting problem indicates to me that you are looking for a
forecasting system that knows more about your objectives and context
than you do. Unless you tend to forecast consistently in the wrong
direction, I doubt that an automated forecasting system will do much
more than refine your best guesses and help you understand better what
you understand to some extent already. If that sounds subjective, I'd
agree 100%. That's another way of saying 'How much money would you bet
on a model with blue lines all in the right places if it forecasts a 70%
drop in oil prices during 2006?'.
From: email@example.com [mailto:firstname.lastname@example.org]
On Behalf Of Dimitris polymenopoulos
Sent: Thursday, December 01, 2005 11:43 AM
Subject: White Noise Test
I've been using the time series forecasting system and being a newbie
concerning statistical analysis, i'd like to ask if anyone can explain
to me the following:
For a time series to have been forecast correctly, both the unit root
tests and the seasonal root tests must show a significance probability
of p<0.001 (that is, blue bars all throughout the lags). On the other
hand, the white noise tests must be significantly greater than zero (no
blue bars). Am I right? And if that's the case, what can I do with a
time series whose white noise signifance test jumps from being 99%
significant in lags 1-10 and then suddenly the significance decreases
appear) past the 0.1 (10%) and even 0.01 signifance level! Any ideas
would be welcome. Also, is there any way to model an ARIMAX model using
the automated series forecasting system or do i have to do it manually?
i'm using sas 9.0.
Thanks in advance!