|
PureBytes Links
Trading Reference Links
|
Monday, April 26, 2004, 6:21:11 PM, Alex wrote:
AM> I have a problem figuring out exactly how to measure the quality of
AM> fit for a low-lag smoothing filter applied to market data. I need
AM> such a measurement for the purpose of an optimization I'm doing in
AM> Excel.
I have two suggestions: (a) measure the filters on synthetic data,
and (b) measure the filters as they will be applied.
No doubt you've seen John Ehlers and Mark Jurik apply filters on
synthetic data. It offers a very controlled environment so you can
effectively measure different aspects of the given filter.
I use sine, triangle, square, and saw wave data in CSV format, both
in "pure" form and with a sharp up/down mixin as Jurik does. Now you
can controllably measure overshoot, undershoot, and convergence (if
any). BTW: very interesting to measure Mesa this way ... you see
instability an little convergence.
The second, and probably more important, is to consider the context in
which the filter will be applied. For instance, most low-lag filters
do not combine well with price cross-over, and some do not do well
with fast/slow filter cross-over. Most low-lag filters DO perform well
with some sort of direction change (perhaps +/- delta). Given this,
one thought is to generate 1000's of price curves, using something
like a travesty algorithm along the lines of Monte Carlo analysis
(you're selecting segments from real data and re-splicing). Apply the
set of filters to each generated price curve and bin the results.
If one filter is more effective than another, I believe it should show
up statistically.
--
Cheers,
Kevin mailto:ksberg@xxxxxxxxx
|