The original question:
>Where can I find a comparison of different seasonal adjustment methods
>with their advantages and disadvantages?
----------------------------------------
In the following I give a short summary to the replies I got on my post
with questions on "Seasonal adjustment methods" sent the 31th July 1997
to the Time series mailing list.
At first I'd like to thank Chris Chatfield, Brian C. Monsell from the
Bureau of the Census, Ilkka Karanta, Guillaume Leorat, David Carr,
Bircan Erbas and Gianluca Cubadda for answering to my questions.
1. "Seasonal adjustment as a practical problem"
FAG den Butter, and MMG Fase (1991)
Amsterdam:North-Holland.
This was recommended to me a s a good start in most replies.
The price at Amazon is: $ 131.50.
2. "Decomposition of Time Series; Comparing Different Methods in Theory
and Practice"
B.Fischer, April 1995. (EUROSTAT publication)
The focus is on comparing the most widespread ad
hoc method, namely X-12 ARIMA, with the model-based approach,
particularly with SEATS by A. Maravall which uses Box-Jenkins models.
The program SEATS and the companion program TRAMO (which identifies
automatically both the ARIMA model and outliers) can be downloaded
without charge from the web site of the Bank of Spain:
http://www.bde.es/servicio/software/softwaree.htm
(reply from G. Cubadda)
3. Sven Hylleberg (ed.): Modelling seasonality. Oxford University Press,
New York, 476 pp, ISBN 0-19-877317-X (hardcover), ISBN 0-19-8773188
(paperback).
A book containing most of the important papers in seasonal modelling up
to the beginning of 1990's.
This isn't a cookbook, and requires some fluency in mathematical
statistics, econometrics etc.
(reply from I. Karanta)
4. Time series analysis - univariate & multivariate methods: William Wei
(reply from B. Erbas)
>Important for me is to lose the minimum of periods at the end due to
>moving averages, in case they are used.
Most seasonal adjustment methods will either (a) use alternate, shorter
filters at the end of a data set that have the same properties as the
longer central filters or (b) allow you to extend the series with some
type of forecast so that you can adjust the entire series with the
central filter.
(reply from B. Monsell)
>Is it better to recalculate the adjusted series after every new arrival =
>of a piece of data? If yes the computing of the adjusted series should >=
not be too complex.
One does get smaller revisions when adjusting the series after each
incoming observation (called concurrent seasonal adjustment) rather than
using projected seasonal factors (see Pierce and McKenzie's article in
JBES entitled "On Concurrent Adjustment" ).
(reply from B. Monsell)
Recalculation isn't necessary. For example, Kalman filter is recursive.
You could consult
A. Harvey: Forecasting, structural time series models, and the Kalman
filter (Cambridge University Press, Cambridge 1989).
(reply from I. Karanta)
> I read a little bit about X-11 and heard of X-11-ARIMA (Stats Canada)
> and the new X-12 of the Bureau of the Census. Does X-12 use Box-Jenkins
> models?
X-12-ARIMA uses regARIMA models (regressions models with an ARIMA noise
term) to allow the user to extend the series with forecasts and
preadjust
the series for outlier and calendar effects before seasonal adjustment.
The seasonal adjustment module is an enhanced version of the X-11
methodology.
X-12-ARIMA (including a reference manual) can be downloaded from the
Bureau of the Census site:
http://ftp.census.gov/pub/ts/x12a/
or directly via ftp:
ftp.census.gov/pub/ts/x12a/
There is a PC and a UNIX version. Infos on how to download can be found
in the files how2down.pc and how2down.unix.
(reply from B. Monsell)
> At work we are using a home-made multiplicative exponential smoothing
> program which includes a trimming feature for the trend and season
> factor series to forecast Use of uninitialized value in concatenation (.) or string at E:\listplex\SYSTEM\SCRIPTS\filearea.cgi line 455, line 125.
monthly industrial production indices.
> As I have made the experience that simpler models usually perform
> better, at least in short-term forecasting, I ask myself why not use
> those seasonal factors to adjust the series?
Using seasonal factors from a seasonal adjustment package like X-11 to
remove seasonality from a series before forecasting changes the
statistical properties of the series, so I would be careful doing this.
(reply from B. Monsell)
If you want a simple method, the sort of thing you suggest may be
adequate.
It depends on how many series you have, how much residual variation,
how far ahead you want to forecast....
There are no simple answers in forecasting - see Chapter 5 of my book on
Time Series Analysis, Chapman and Hall, 5th edn, 1996.
(reply from C. Chatfield)
--
Alexander FRITZ, database manager at DEBA-GEIE,
1, rue Emile Bian L-1235 Luxembourg
Tel.: (352) 29 77 71-24
FAX: (352) 29 77 71-50