How do you achieve a qualitative leap in market analysis? There is an option: - page 6

 
SergNF:
Reshetov:
... "Maybe someone will come in handy and maybe someone will tell you how to do it better?"
"Classics of the genre" suggest using "time lags" as neural network inputs, i.e. essentially recognizing a "time pattern" (what is now in ArtificialIntelligence.mq4). IMHO, sometimes it turns out to be interesting to recognize AND.... say, a "situational pattern", i.e. to input values of several "indicators" (in quotes!!!!) on the last bar (for example, Fourier spectrum or "whatever it is called," again "Arbitrage" .

I've tried both. And what's called more than one week the computers were humming. The time patterns, or indicator values in dynamics always give more profitable results compared to sets of values of different indicators on the last bar. It is also important that by the time of optimization these very temporary patterns are faster in MT4. And with the selection of different indicators we obtain something similar to Krylov's fable "Kvartet": you guys are not good musicians yet.

As for wavelets - this is a scam. If we take any function, decompose it into a Fourier series and restore it, it falls under the definition of a wavelet with respect to the zero harmonic level, since the integral of the function histogram at this very level is 0. The wavelet operators only invent that their "inventions" supposedly contain more information than the Fourier transform. Fucking lobbyists are lying.
 
Folks, does anyone here really need to play smart? Reshetov, I understand that there are people who are fundamentally critical of everything they see. Read carefully what I said about your system, there is not a word of criticism. In fact, I didn't feel lazy and ran it through the history yesterday. Honestly, I didn't spend much time optimizing it, because I've got 5 parameters (plus 4 shifts relative to current bar in iAC). Didn't get any definite result. But it would be a mistake to claim that nothing will be achieved. We should at least perform a full optimization for all input parameters and analyze the resulting 6-dimensional surface. I do not know what considerations guided you when you decided to take 4 weights. I tried, like in my own case, to disable support in your system for pure flips. But couldn't get a full view. Again there are a lot of parameters for full optimization. Using genetics does not give an objective result as everything depends very much on which parameter is being optimized. This can be seen just from the example of your expert.
 
getch:
We should at least do a full optimization on all input parameters and analyze the resulting 6-dimensional surface.
We have to, Fedya. We have to. Otherwise it will turn out to be an empty phrase: we ran one parameter on the history with the other, and quotes and contract specifications from the third brokerage company.

As Zoshchenko said: I would be surprised if a lady put the half of her coat in a bucket with paint. And I would be surprised if a system that was designed for one thing was run on another and you could see the result.
 
My research has also shown that using time patterns is more effective. I just don't understand why we should input the changes of indicator values instead of the price. In the end there is pattern recognition by indicator values (which is often wrong), but not by price. I suppose that the use of the neural network is most effective when it is run through prices. If you believe in self-similarity of time series of quotes, you'd better use the smallest timeframe. Because the system will give more signals and there will be much more ineffective sets of weighting coefficients.
 
"Wavelets are a scam" is a bold statement when you consider the significant improvement in the compression ratio of some data when using them.
 
getch:
My research has also shown that using time patterns is more effective. I just don't understand why we should input the changes of indicator values instead of the price. In the end there is pattern recognition by indicator values (which is often wrong), but not by price. I suppose that the use of the neural network is most effective when it is run through prices. If you believe in self-similarity of time series of quotes, you'd better use the smallest timeframe. Because the system will give more signals and there will be much more ineffective sets of weighting coefficients.
Well, who forbids you to replace all iAC() with corresponding shift Close[] in the code?
 
getch:
"Wavelets are a scam" is a bold statement when you consider the significant improvement in the compression ratio of some data when using them.
If one considers the loss of quality when using this very compression, one will find out where the "redundant" information went.

Why bother people with wavelets and other innovations when it is possible to use the same data in Fourier transform, cut off a part of harmonics with small amplitudes, reconstruct them relative to level 0 and obtain thereby what is called a wavelet?
 
A small digression: if the change of quotes is a completely random process, then creating a profitable system is not possible (otherwise there is pseudo-randomness). But there is no task to create a profitable system. Because even with a random behavior it is possible to create any profitable and stable system on the finite time interval. And this finite time interval can be measured in minutes, or years or decades.
 
getch:
A small digression: if quote changes are a completely random process, then it is not possible to create a profitable system (otherwise there is pseudo-randomness)
Pseudo-randomness and randomness are different categories. At stock markets, quotes are driven by market makers and their behavior is not random. The currency and other markets do not toss dice and pennies when they move pips either.
 
Yes, we can discuss the criticality of losing some information for a long time. Because everyone perceives the rest of the information differently. You have to look at the value (amount of information) / (perception). I have not come across such studies.