[ppml] Combining Forecasts

[michael.dillon at bt.com]

> except for how amusingly accurate frank solensky was over a 
> decade ago.

Once upon a time, Frank said this:

   Frank Solensky: ...(my) estimates for
   maximum address space utilization
   have risen about 8% over the last 3
   years. If one were to argue for
   extrapolating this over time as well,
   the resulting statement would be that
   "in the year 2019, the trend line will
   suggest that we will eventually run
   out of IPv4 addresses".


Given that current projections point to 2010 as the date, what was so
amusingly accurate about what Frank said?

Note that given 10 years of quotes to trawl through, I expect that I
could find "amusingly accurate" predictions on just about anything you
care to mention.

In any case, this thread is about combining forecasts which is an
established way of dealing with multiple forecasts which give different
answers to the same question. It's not about finding pinpoint accuracy,
but about dealing with the variations between the different forecasts.

One of the ways to deal with two forecasts (Geoff's and Tony's) which
give different answers is to simply average the two answers. This isn't
just an off the cuff suggestion to sweep the differences under the
carpet and avoid the fact that one or both of them can't produce an
accurate forecast. Instead, this is established forecasting best
practice that is the result of numerous studies over the years.
Averaging two forecasts may not give a very large increase in accuracy
(about 12%), but it does narrow down the time period in which we will
run out of the global free pool.

Given the fact that combining forecasts is known to work best when using
both METHODS and DATA that differ substantially
http://www.forecastingprinciples.com/paperpdf/Combining.pdf should we
actively try to solicit additional forecasts that can profitably be
combined with the Tony/Geoff work?

More on forecasting principles here for anyone who wants to follow up on
this:
http://www.forecastingprinciples.com/papers_page.html#Full_text_papers
http://www.forecastingprinciples.com/researchers.html

--Michael Dillon