Date:         Fri, 19 Nov 2004 17:27:32 -0800
Reply-To:     cassell.david@EPAMAIL.EPA.GOV
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "David L. Cassell" <cassell.david@EPAMAIL.EPA.GOV>
Subject:      Re: Applying Neural Nets in SAS EM for continous target variable
In-Reply-To:  <446DDE75CFC7E1438061462F85557B0F0613E7D1@remail2.westat.com>
Content-type: text/plain; charset=US-ASCII

Sigurd Hermansen <HERMANS1@WESTAT.COM> sagely replied:
> I would not rule out using neural nets to predict really difficult
sequences
> in time series forecasting (such as turning points); nonetheless, the
usual
> sources of variation in time series (trends, seasonality, and cycles)
don't
> fit particularly well into a classification model. For example, serial
> correlation typically explains everything and nothing in a time series
> model.

Standard neural nets are generally equivalent to multivariate
statistical
techniques of one sort or another.  Projection pursuit analysis, etc.
There has been some info on this in SAS-L in years past.  So I wouldn't
recommend neural nets for a clear time series problem.

> The fact that SAS has a separate product for analyses of time series
(ETS)
> suggests that you should look at that first, and other time series
analysis
> packages such as RATS/CATS as well. Time series analysis may be the
most
> complex problem in statistics. I recall from way back that my
econometrics
> instructor advised his students to delay looking at time series until
we had
> a better grounding in cross section analyses. He did not mention
neural nets
> as a possible short cut. I have not heard anyone else recommend neural
nets
> either. The only common ground that I see for neural nets and some of
the
> frequency domain programs used to estimate cycles is that both fall
into the
> general class of 'black box' modelling tools.

Also look at Autobox.

David
--
David Cassell, CSC
Cassell.David@epa.gov
Senior computing specialist
mathematical statistician