|
PureBytes Links
Trading Reference Links
|
Trey Johnson wrote:
>According to Ehlers, FFT shouldn't be used for analyzing price data. Can
>anyone point out why FFT can be used?
I replied privately to Barry Silberman but I may as well reproduce and
expand my reply here.
>-----Original Message-----
>From: Barry Silberman [mailto:barry@xxxxxxxxxxxxxxxxxxxxx]
>Sent: Friday, July 12, 2002 9:09 AM
>To: omega-list@xxxxxxxxxx
>Subject: Request for "FFT or similar cycle measurement"
>
>Gary Fritz recently pointed out that
>
> "Ehlers' HD (Homodyne Discriminator to determine the current
>market cycle) is also extremely noisy and he includes smoothers in
>every indicator to try to deal with that. Still the result is MUCH
>noisier than an equivalent FFT or similar cycle measurement."
>
>Can someone help provide the EL code for FFT or "similar cycle
>measurement" that may be more reliable.
TS already has the FFT function, I recall.
FFT and other cycle measurement techniques (maximum entropy method,
linear predictive coding, wavelet transforms), suffer from requiring
some fixed lookback length in which to measure the dominant
frequencies. Cycles in the market 64 bars ago don't necessarily
have anything to do with the cycles present right now.
The biggest problem with FFT is that it assumes all your data is
periodic. If you have 64 bars of data, FFT assumes the same 64
bars will repeat themselves, i.e. bar 1 will immediately follow bar
64. Therefore you cannot use it for prediction, because it has a
built-in assumption that the data will repeat identically again and
again, interval after interval. This isn't the case in reality.
Ehler's HilbertPeriod, on the other hand, uses only the last 6 bars
to get an estimation of what the dominant cycle period is right now.
It works much like an exponential moving average in that it weights
the most recent activity heaviest. Other techniques won't do that.
It's also extremely fast to compute. I don't think its results are
any worse than from using FFT or other techniques. It measures only
one cycle period though, and several others may be present that, by
themselves, are smaller than the dominant one, but together they
might dominate the "dominant" one.
The biggest problem with ALL these cycle analysis techniques is that
markets don't exhibit cycles all the time. In fact, they exhibit
cycles a minority of the time. Applying a cycle measurement will
always give you SOME answer, but without taking into account the
amplitude of the cycle, you don't know how useful that answer is.
You can apply these cycle measurements to completely random data
and you'll still get an answer that "looks" useful but in actuality
isn't.
--
,|___ Alex Matulich -- alex@xxxxxxxxxxxxxx
// +__> Director of Research and Development
// \
// __) Unicorn Research Corporation -- http://unicorn.us.com
|