time series in R

Ross Ihaka ihaka@stat.auckland.ac.nz
Tue, 20 Jul 1999 08:49:23 +1200 (NZST)

On Mon, 19 Jul 1999, Prof Brian D Ripley wrote:

> Time Series functions in R

I agree with basically everything Brian says, and would add the
following thoughts:

    1.	Think carefully about defining interfaces before leaping
	into code.  By defining an interface independent of the
	particular code at hand we will make it easier to switch to
	new code.

    2.	Think multivariate even when implementing a function for
	univariate series.  If the interface is defined well, it
	should later be able to generalize to the multivariate case.
	I'd hate to see a second multivariate time-series package
	come along later.

    3.	On the definition question: The existing FFT implementation
	uses a particular definition for the discrete transform which
	is pretty standard.  (Edwards "Fourier Series", Brillinger
	"Time Series" etc.) Using another definition may complicate

    4.	The notation for ARMA models may well be the stuff that
	holy wars are made of, but my personal preference is
	    X[i] + phi[1] * X[i - 1] + ... + phi[p] * X[i - p]
	    = theta[0] + epsilon[i] + ... + theta[q] * epsilon[i - q]
	[ Man to judge: ``It was self defense - I thought he was
	  going to hit me, so I hit him first.'' ]

I'd also add a request for some flexible structural modelling code.
I've pretty much abandonded teaching ARIMA models in favour of
structural models.  (I've written some general code for this, but
there is probably better code about.  There are a couple of TOMS
algorithms for Kalman Filtering which should be checked out.)

Finally: Does anyone know Kitagawa?  He has some very nice time series
and smoothing code.  Could he be persuaded to make it available?


r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-devel-request@stat.math.ethz.ch