time series snafu

[Jason Stover]

After much hair-pulling, the time series module is getting close to
being as good as I'm going to make it for a while. I suspect that's
also about as good as it can get.

Here is my summary of how well it works: Badly. I started writing it
partly because the R time series module did not produce very good
estimates and I thought I could write something better.  And for the
tests I have run, the module I wrote does a little a better, but still
not so great. The only reason my own module works as well as it does
is because of its good initial estimates.  I'm starting to think that
maximum likelihood estimation for ARMA processes is just difficult, and
that writing a program to estimate ARMA parameters for an arbitrary, 
user-specified
order isn't a good idea. I wish I had a better result, but after a few
months of work, I don't think the code is going to be what I had hoped
for originally. At least I can get a conference poster out of it.

So now I'm left wondering what to do. The code I have written so far
will work, but the estimates it gives beyond the initial values aren't
so good. I might add this code to the cvs repository with the
understanding that it needs major improvements
later. src/math/ts/innovations.c is potentially a useful routine for
computing orthogonal projections elsewhere, but leaving it by itself means
leaving some mathematical code in the cvs repository without any
procedures that use it.

And perhaps the worst part of the news is this: Not many users need to
estimate the parameters of an ARMA process. This just isn't nearly as
useful to as many potential users as a GLM procedure, or logistic
regression, or one of many other possible procedures. 

So: Should I junk this whole thing, or check it in? I'm hoping
everyone will say "forget it and start on the GLM procedure." In any
case, I think innovations.c should probably stay where it is, since it
could be useful in the future.

-Jason