Roger wrote:

The smoothing filter seems to be there to iron out the large year on
year variations and give a better idea of the trend at a glance. It uses
a binominal distribution over 11 years with the central year getting a
weight of only 0.176197 and the adjacent years weighted at 0.160179. The
ends have so little weight I am surprised they bother with years 1 - 3
and 19 - 21. On historic data it would appear that it does a pretty good
job but IIUC at the end of the sequence when the 5 forward years are not
available they simply project the final value forward for 5 more years.
The previous year keeps its weight of 0.160179 but the end year then
gets a weight of approximately 0.588* and the 3 previous points on the
smoothed graph are significantly influenced in the same way which will
give totally the wrong impression if the latest figure is on the margins
for any reason.

You might be interested in the following exercise, where I took my BP
data and treated it to a) a 10-point smoothing exercise, and b) a
linear regression, which had a lower variance than a second-order
polynomial (not shown)(all data manipulation courtesy of Excel
spreadsheet).

Basic data, no form of data reduction:

http://farm4.static.flickr.com/3004/...2a5fcf16_o.jpg

10-point moving average:

http://farm4.static.flickr.com/3163/...28b914b0_o.jpg

Linear regression:

http://farm4.static.flickr.com/3230/...900527a5_o.jpg

The doc looks at the first graph, and says "you've got a problem".

The MetO looks at the second graph, and says "your BP is going up"

I look at the third, and note that my BP is falling at the rate of 1
mm Hg per seven days (from the slope of the systolic regression line),
and I say "if I keep this up, I won't need any extra medication, and
my BP will reach 120 systolic after 178 days, or 80 diastolic after
134 days". I'm already 73 days into the trial.

To my mind, there's no comparison between smoothing (which gives me no
figures and is misleading) and a regression exercise (from which I
can calculate all sorts of data). I'll go for the latter every time.

[For Stuart Noble's interest, here's my BP as a function of time of
day, complete with second-order polynomial regression - gives least
value of variance:]

http://farm4.static.flickr.com/3057/...fd1fc9c3_o.jpg