FIR filters - page 10

 
FION >> :
You guys are arguing about the wrong thing, Prival demonstrates a classical approach and Sabluk a practical approach. Better let's talk about the applicability of spectral methods to the market. A trending market is low frequency, a flat market is more high frequency. That's understandable to the hedgehog, but how is it "much better" than the same wagons? You can also make the flappers long or short, in principle the flapper is close to a first order Butterworth filter with quality factor 0.7. In most applications it is quite satisfactory. It's also not a fact that fast response is a good thing, it all depends on the TC.

If you think a little bluntly about what grasn has pointed out, there is some light at the end of the tunnel.

Because MA, Djuric, FIR and other filters by themselves are a dead end. So it is a bit faster, a bit smoother curve. But in essence they are the same.

Fourier, there are no sinusoids.

But if you estimate the probability of fast or slow changes, then you can think about how to filter.

After all, before you filter, you should know what you want to filter and what you will get as a result.

And then the delay of the filter is not so important, because we know it and we can estimate its ability to distort the trend direction.

 
begemot61 >> :

Yeah, yeah, yeah, I'm the one who's confused by SYNUS.

If you can email me: eugene_dvoskin@yahoo.com

Attachments up to 10 Meg.

Catch

 
begemot61 >> :

If you think a little bluntly about what grasn has pointed out, there is some light at the end of the tunnel.

Because MA, Juricks, FIRs and other filters on their own are a dead end. So it is a bit faster, a bit smoother curve. But in essence they are the same.

Fourier, there are no sinusoids.

But if you estimate the probability of fast or slow changes, then you can think about how to filter.

After all, before you filter, you should know what you want to filter and what you will get as a result.

And then the delay of the filter is not so important, because we know it and we can roughly estimate its ability to distort the trend direction.


I don't know how blunt one should look, but it seems to me that the only way to apply this field is through adaptive filtering. And in that it all comes down to correct identification of the model. And this is not a very easy task.

 
ssd >> :

It's beginning to become a little clearer what this is all about. If you have time, please comment on my reasoning.

Let's say we called 2000 bars in the terminal and we want to analyse the "wave" pattern.

May I say that I deal with a wave frequency F=1/T= 1/(2000*timeframe_in_minutes* 60) or with a period of 2000 bars?

It turns out that I can.

Then what can be done with this wave ?

I take it, represent it as a Fourier series and see that this wave with the period of 2000 bars actually consists of a number of harmonics.

Each of the harmonics also has a different frequency/wavelength/period, amplitude.

In other words, each harmonic is again a wave with a period, which is again measured in bars.


If for the filtering process I set a bandwidth for waves from the frequency range,

say from 200 bar to 600 bar, would that mean ? What?


The reasoning seems to be correct, but I do not quite understand the question.

In general:

1. You have shifted the cut-off frequency down by a factor of 3. And what has changed at the output depends on what was in the input signal spectrum. I.e. in some cases, the output signal may be virtually unchanged.

2. You have estimated the spectrum. Are you sure you have done it correctly? In order to estimate the spectrum of a signal, you have to have a good idea of the properties of that signal. This is not sophistry. Then you can estimate the error of a particular method. Just otherwise you can easily get hilarious pictures that have nothing to do with the real spectrum.

3. If you're using my genius work, maybe there's a mistake somewhere? I'm not much of a programmer.


 
grasn >> :

It seems to me that the only way to apply this field is adaptive filtering. And in it all depends on correct identification of the model.

As far as I understood from afftar description of JMA on his website, this filter works well up to Cauchy distribution model. And this distribution, as we know, does not have not only the second, but even the first moment (i.e. m.o.).

Djuric even says, whoever shows the filter that works better on data subject to Cauchy distribution by returns, will get a money prize.

Seryoga, is that what you mean by correct model identification?

 
begemot61 >> :

The reasoning seems to be correct, but I don't quite understand the question.

In general:

1. You have shifted the cut-off frequency down by a factor of 3. What has changed at the output depends on what was in the input signal spectrum. I.e. in some cases, the output signal may be virtually unchanged.

2. You have estimated the spectrum. Are you sure you have done it correctly? In order to estimate the spectrum of a signal, you have to have a good idea of the properties of that signal. This is not sophistry. Then you can estimate the error of a particular method. Just otherwise you can easily get hilarious pictures that have nothing to do with the real spectrum.

3. If you use my work of genius, maybe there's a mistake somewhere? I'm not much of a programmer.


The program is fine, draws a good line, it's hard to find a MA that draws such a line.

I'm not talking about the programme yet. Let me try again in my simple language.

Suppose the spectrum of the wave I mentioned with a period of 2000 bars, in addition to all the other harmonics,

there's a harmonic with a period of, say, 50.

(I can not imagine it physically, and imagine it only as an element of the Fourier series for this wave with a period of 2000 bars,

although intuitively I understand that such a harmonic is a kind of a fine rattle, which should be discarded).

Suppose further that some ideal filter is set up to allow the entire spectrum of the said 2000 bar wavelength to pass through to the output,

Except for that one harmonic, which is perfectly suppressed.

Now the question which concerns the "physics" of the filter operation.

As I see it, the filter, using various methods and techniques, finds in the input wave with a period of 2000 bars

all possible combinations of consecutive 50 bars and what does it do with them?

 
ssd писал(а) >>

The programme is OK, draws a good line, it's hard to find a MA that draws a line like that.

I'm not talking about the program yet. Let me try again in my simple language.

Suppose the spectrum of the wave I mentioned with a period of 2000 bars, in addition to all the other harmonics,

a harmonic with a period of, say, 50.

(I can not imagine it physically, and imagine it only as an element of the Fourier series for this wave with a period of 2000 bars,

although intuitively I understand that such a harmonic is some kind of a fine rattle, which I should get rid of).

Suppose further that some ideal filter is set up so that it allows the entire spectrum of the said 2000-bar wavelength to pass through to the output,

except for that one harmonic, which it perfectly suppresses.

Now a question which concerns the "physics" of the filter's operation.

According to my understanding, the filter, using various methods and techniques, finds in the input wave with a period of 2000 bars

all possible combinations of 50 consecutive bars and what does it do with them?

You will not understand it until you get acquainted with the Fourier theorem. You can't just jump into it. You have to learn a little bit.

 
grasn писал(а) >>

I don't know how blunt one should look, but it seems to me that the only way to apply this field is through adaptive filtering. And in that it all comes down to correct identification of the model. And this is not a very easy task.

That is my point exactly. And it seems this way lies in a self-adaptive mesh, like the one Neutron is talking about in his thread.

 

to Mathemat

Серега, ты на это намекаешь, говоря о корректной идентификации модели?

Alexei, see the private message.


to FION.

That's what I'm saying. And it looks like this way to a self-tuning grid, like the one Neutron is talking about in his thread.

"Shura, no theft - only robbery!!!" (s) (something like that, I don't remember verbatim). You won't believe it, but perceptrons are multi-layered and "the like" are the same filters. Not an expert at all, but I find the application of self-organising stochastic control systems and filtering theory (especially the adaptive part) more tempting. These are two big and related theories, moreover, more elaborated and practical for BP than for NS. Of course, there are subtleties and I'm not against NS at all, moreover, I use such a thing. Anyway, we'll see.

 
ssd >> :

The programme is OK, draws a good line, it's hard to find a MA that draws a line like that.

I'm not talking about the program yet.

For https://www.mql5.com/ru/users/begemot61

Now for the program.

Found out today that it's overdrawing the indicator line.

It's clear that it's in here somewhere:

int start()
{
int limit, i;
int counted_bars=IndicatorCounted(); //amount of bars shanged
if(Bars<=(FilterLength+1)) return(0); //not enough bars for calculations
if(counted_bars < 0) return (0); //eror protection
if(counted_bars > 0) counted_bars--;
limit=Bars-counted_bars-1;
for (i = limit;i>=0;i--) // Cycle for uncalculated bars
{
FilterBuffer1[i] = FilterResponse(i); // Value of 0 buffer on i-th bar
}
return(0);
}
----------------------------

It turns out that the program changes not only the i-th buffer element but also the elements already generated by ....