Anyone one know how adslguide work out their 'predicted trigger dates' -
from one or two I've monitored it looks likely they may just be making a
straight line extrapolation from the last two data points! (It doesn't
appear to say on their site, and they haven't replied to my e-mail...)

The overall result is sadly rather meaningless - a bit like the 'how
fast is your ISP' exercise where we're meant to read significance into
the difference between 450 & 455 kbps data rates determined by a short
trial download... when for most of us other overheads dominate small
transfers, (and long ones are, or should be, all much of a muchness)

Bob

(Yes, I _do_ like the site - even though it still breaks in Netscape! -
and Happy Xmas to you all)
--
robert w hall

On Tue, 23 Dec 2003 19:47:37 +0000, robert w hall wrote:

No idea. The problem is that the start of a campaign, or even just
the announcement of a target can significantly effect an exchanges
registration rate, by

For example take the Riding Mill exchange

http://www.adslguide.org.uk/availabi....asp?code=NERM

On the 3rd December any reasonable prediction of the trigger date would
have been months away, yet two days later it hit the trigger. This is
perhaps an extreme example, but illustrates predicting trigger dates
like all extrapolation problems is fraught with pitfalls.

You could always take their data and fit your own lines to it and
calculate your own prediction date if you want something more robust.
Certainly something like the Berkeley exchange looks just perfect
for a linear least squares fit.

JAB.

--
Jonathan A. Buzzard Email: jonathan (at) buzzard.me.uk
Northumberland, United Kingdom. Tel: +44 1661-832195

In article <pan.2003.12.23.22.50.06.667090@uk.me.buzzard>, Jonathan
Buzzard <jonathan@uk.me.buzzard> writes
agreed - so can Christmas lassitude :-)
or Torpoint, where they got 200 registrations in 3 days, and then... :-)


error - they both went through a hundred in mid April and are on
198-ish now. But one has a predicted target of June,and the other has
had a March date and now has an October one. Hmm.

/Long Boring Aside/
** Technically though, I was always told to differentiate noisy data
with the d/dx of a smoothing function, like exp(a*x*x)
ie convolution with 2x exp(a*x*x).
However,
The data are easily expressed as T(N) (time T to reach N registrations),
which is equally spaced.
But even then, convolution is slow,
So adslguide really need to have an FFT routine handy...
//
Bob

(* By the way, try googling on (Market)-Lavington - it appears to be
quite a place!)
--
robert w hall