R help - Jul 2011 - Discussion on time series analysis and the use and misuse of Differencing

How does the R module ARIMA account for unspecified deterministic structure
such as seasonal pulses, level shifts, local time trends and regular pulses
without needing to ask the user to intervene to specify this? 

I have attached a Makradakis paper which hammers Box-Jenkins approach to
this problem of nonstationarity. I have also included a recent discussion
from stackexchange which you might find even more interesting. 

http://stats.stackexchange.com/questions/12651/box-jenkins-model-selection/12662#12662

http://www.insead.edu/facultyresearch/research/doc.cfm?did=46900

-----
Tom Reilly
www.autobox.com

--
View this message in context:
http://r.789695.n4.nabble.com/Discussion-on-time-series-analysis-and-the-use-and-misuse-of-Differencing-tp3652268p3652268.html
Sent from the R help mailing list archive at Nabble.com.

tomreilly <tomreilly <at> autobox.com> writes:
> 
> How does the R module ARIMA account for unspecified deterministic structure
> such as seasonal pulses, level shifts, local time trends and regular pulses
> without needing to ask the user to intervene to specify this? 
  It doesn't.
> I have attached a Makradakis paper which hammers Box-Jenkins approach to
> this problem of nonstationarity. I have also included a recent discussion
> from stackexchange which you might find even more interesting. 
> 
> http://stats.stackexchange.com/questions/12651/
   box-jenkins-model-selection/12662#12662
  [split URL, Gmane doesn't like long lines]
> 
> http://www.insead.edu/facultyresearch/research/doc.cfm?did=46900
> 

  My apologies if I'm misunderstanding your point here, but my general
understanding of R's philosophy is that it provides a set of tools
that can reasonably be used to as a component of sensible statistical
modeling, but which in some cases can also be abused (e.g., stepwise
fitting tools). It is rarely prescriptive.

Everyone should look at their data (before, marginally, and
after, residually) to see whether there are problems/important
patterns that are not captured by the particular model being fitted.
One can certainly make a case for simpler (AR) methods that are more
robust to misspecification, or for more complex models that capture
some of these patterns, but in the long run (in my opinion)
an informed user is the best defense. 

If you distrust ARIMA models by all means don't use them, but you'd
have to make a pretty strong case that they were *never* appropriate
in order for anyone to consider removing them from R (and dealing
with the howls from disappointed users) ...

  Ben Bolker