Fourier connoisseurs... - page 6
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Vladimir - then how does your Extrapolatore shape the continuation process of the transformed function...?
Suppose the first trigonometric function is fitted to prices x[n]:
y[n]=m+A*Cos(w*n+a), n=0,1..N-1
x[n] - real prices, y[n] - model. The extrapolation is simple. We take the same function and the time index n is incremented forward, into the future
y[n]=m+A*Cos(w*n+a), n=N,N+1,...
Then a second trigonometric function is fitted to the remainder of x[n]-y[n]. And so you continue until you have fitted and extrapolated a given number of functions (HarmNo):
is not redrawn. But it's not in good shape, although I haven't seriously tested it. Although I understand that for those who work with neural networks this indicator will be very valuable.
I wish I had a good partner for research, to run it in the NS.
What are you going to use as inputs for NS, and what is the supposed dimension of the inputs?
What is intended to be used as NS inputs, and what is the intended dimensionality of the inputs?
This induced input, it is guaranteed to lie between -1 and 1.
I only know the theory, but I haven't used NS practically, so some questions may be confusing. If you're interested in it, I'd be only too happy to. I'd prefer to skype. It will be more convenient and faster. Although I don't believe in these NS.
The only input parameter is the depth of history. 60 minutes is selected in the figure, i.e. 60 last minutes are analyzed. The interpretation of more than 0.5 + indicator goes up is a buy, less than -0.5 and falls is a sell, this is from the theory (logic of its construction). But I haven't run it in the tester, since I can see its further improvement and am working on it.
It's better to divide it into at least two ranges (Fast/Slow)
and interpret their combinations
The examples you gave of y=k*x+c or very large period, it is a failure of Kotelnikov's theorem, the spectrum is infinite.
Exactly, if you subtract the spectrum that PF finds from the quotes, the remainder is always the long-period component,
which the PF can't extrapolate correctly.
I tried to avoid this problem in two steps,
by compressing the quotes (1st), i.e. every 10th bar in the result out of 5000 obtained. 500 bars and extrapolated
I then changed the result obtained by stretching and subtracting from the price and obtaining a market slice without extra LF component (2 Sts),
but because of the fractality of the market I got the same problem but in 1 st and so on to infinity.
Hence, the moral for extrapolation of curves not satisfying Kotelnikov's theorem requires another method.
P.S. Which one I do not know yet. >> Good luck.
This input induced, it is guaranteed to lie between -1 and 1.
I only know the theory, but I haven't used NS practically, so some questions may be confusing. If you're interested in it, I'd be only too happy to. I'd prefer to skype. It will be more convenient and faster. Although I do not believe in these NS.
I don't believe in anything myself.
Is there only one way in? It's not an idle question. The matter is that NS essentially searches for an optimum of functionality (any) at necessary quasi-stationarity on input parameters. Before to bother with construction we need certainty of stationarity by these or those parameters. The more such parameters, the more advantageous the NS is in comparison with parametric analysis methods.
I don't believe in anything myself.
Is there only one way in? The question is not idle. The point is that NS is essentially looking for an optimum of a functional (any) with the necessary quasi-stationarity on the input parameters. Before to bother with construction we need certainty of stationarity by these or those parameters. The more such parameters, the more advantage of NS in comparison with parametric methods of analysis.
It is possible to have a lot of them. If we consider as a new input, the same inductor but with a different depth of analysis
here's a picture of N=60 and N=480.
I just think that NS will find patterns faster than I do when searching for variants of this indicator in the tester.
But your figure shows a straight line, which is tied to the formula y=ax+b
I'm showing a function that through a Fourier transform (green line)
has its function based on cosines, i.e. we can observe the continuation of the curve...
after transforming it, we get the pre-curve and after transforming it we get the smoothed
price
That's the thing: as long as you extrapolate the generated lines, everything's fine, but when it comes to the market, ouch.
I've tried the most complicated combinations with frequency hopping, and frequency hopping, and everywhere the PF does the trick
except when Kotelnikov's theorem is violated. Try subtracting the long-period MA from the quotes
and extrapolate the result and the last point effect will disappear.
But you cannot extrapolate the long-period MA itself because of this very theorem.
What do you mean you don't believe in anything, Sergei? I still take you for a nerve-racking enthusiast.
P.S. I decided to refresh an unfinished project from a year and a half ago. Fibs. And I'm afraid my code will be very similar to AIS code in style of naming functions (only them though)...