Re: [cmap] Fft question
- Hi Sally,
Thanks for your response and interesting discussion.
This was originally researched by J.M. Husrt and
published in his book "The Profit Magic of stocks
Transaction Timing" (about 25 years ago). His research
was done mostly visually and manually at that time and
not mathematically formalized. It shows that cycles
take about 25% of the price movement. The rest -
mostly fundamental events.
But cycles (means repetition) does exist because of
the human phycology.
By the way I've already finished this and so far
cannot make a conclusion if I like the results.
My biggest problem is the time interval - because it
mostly dictates dominant harmonics. Also,
approximation works good on the interval of sampling.
Prediction - takes you outside this interval and
depending on the number of harmonic you sum, the
result outside the interval can jump in both
directions. I think it is a task of extrapolation
rather than interpolation.
Also interesting that the trend of data impacts the
harmonics. It make sense to de-trend the data first,
then look for dominant harmonics.
--- RainbowSally <RainbowSally@...> wrote:
> Let's start here...__________________________________________________
> FFT and stock market prediction...
> You are on the right track but you'd be better off
> some friends in the business-intelligence community
> you want your bets to be a "sure thing." :-).
> Igor Livshin wrote:
> > Hi Sally,
> > Thanks for your response. What I am doing is
> > decomposing the stock market data to get the
> > harmonics, building envelopes and so on. Fft works
> > fine until you get out of the processed time
> > to make some future price projection.
> You are integrating noise, looking for patterns.
> certainly won't get direct hits each time, but if
> get closer than you would by simply guessing, it's a
> success! :-)
> And then there's the psychology of value. The
> stock market thing is based on perception of value,
> on real value. If all the companies on the stock
> were to liquidate, they'd be worth pennies on the
> re: prediction...
> Wavelets would be no help at all. The word
> sometimes comes up in that context but it has to do
> the compression ratio for the very next sample in
> and u-law codecs (compressing 16 bit numbers into 4
> or 8 bits, and so forth).
> Wavelets are better for reproducing noise! That's
> what you want to integrate out. FFT is your best
> tool for that.
> And you may be closer than you think to getting
> accurate data. But accuracy in this context may
> not mean "exact" data. Just "better than random
> chance." :-)
Do You Yahoo!?
Yahoo! Shopping - Thousands of Stores. Millions of Products.