Quoting Prof Brian Ripley <ripley@stats.ox.ac.uk>:

A seasonal ARIMA model with period 168 is normally unrealistic: how long

is the series? This model has several hundred parameters.

The series is 1000 long.

I suggest you try arima0, as that is likely to use less memory, but either

is going to be inefficient as you are essentially fitting 168 separate

ARMA(1, 2) models for (I guess) each hour of the week.

Correct guess! ARIMA gave me good results using the last 336 samples as input,

so I though I'd try with the whole 1000. I am abviously not an expert

statistician and I am working on a brute force approach by trying all the p, d,

q and P, D, Q seasonal parameters...

(The basic information we ask for in the posting guide such as the version

of R and your platform is missing here.)

Sorry:

R : Copyright 2005, The R Foundation for Statistical Computing

Version 2.1.1 (2005-06-20), ISBN 3-900051-07-0

on

a Linux Fedora Core 3 machine with 2 Xeon processors:

Linux version 2.6.10 (gcc version 3.4.2 20041017 (Red Hat 3.4.2-6.fc3)) #1 SMP

Thanks for your help!

--

Jean-Luc