Author's dialogue. Alexander Smirnov. - page 5

New comment
 
Prival:

ASmirnoff 25.01.2008 11:08


Your answers to the following questions are important for me: 1. Whose algorithm is better: mine or Djurica's? How much better? 2. Do you have Djurica's algorithm? 3. How do they differ?

Very glad to have you on this forum. And would like to help with the research. Unfortunately, I cannot assert that here ('Efficient averaging algorithms with minimum lag and their use in indicators') are those 'true' Djuric algorithms. But I think you can use them and any other. I am ready to offer some of my filters for comparison, but more about that below.


My suggestion is as follows. If we need to compare 2 indicators and answer the question which one is better. First, we need to determine the criteria. If there is more than one, we can perform convolution. The main thing is not just a philosophical assertion that one indicator is better than another one, but exactly the mathematical proof of this fact. Plus how exactly did you ask the question how much better ?


Therefore I would like at the first stage to define the following concepts (I take the MA drawbacks from your article)

  1. The time delay of averaging. Where at what point in time it should be calculated and how? To determine if we need interpolation or extrapolation ? Or compare both?
  2. Relatively high volatility ... Further in the article is the Slutsky-Yula effect a Gibbs phenomenon or something else ? (a file explaining the Gibbs phenomenon is attached). If it is something fundamentally different, would like to read, who has a material please share.
  3. Please disclose the concept of " linear frequency distortion".
  4. "Linearisation of non-linear price trends by isolating these trends with a certain bias".

What do you understand by trend, please formula. What do you understand by a non-linear price trend, the formula + what is shifted in relation to what and how, what is singled out.


The thing is that for the comparison of any two indicators we have to give them something for input. And to answer the question about which of them filters better (smoothes, delays, etc.) you need to know the truth, what we are feeding to the input. Let's assume a noisy sine wave whose parameters we know


by filtering (smoothing) this data array we can determine which one of the filters better distributes the truth, as we know it (blue curve).

It is possible to synthesize various input signals, like the ones used by Djuric to demonstrate the excellence of his indicator.(http://www.jurikres.com/catalog/ms_ama.htm#top). The main thing is to determine which ones.

At the final stage of analysis I am ready to output a synthetic one (artificial price series) characterized by AFC, IFR and ACF that coincides with the price series of any selected currency at a period of time like a week. Here, we did something similar in 'Random Flow Theory and FOREX'. We compared Kalman filter and Butterworth filter.


You in the article talk about AFC and FFC of some steady-state mode and some original criteria.

If you managed to achieve some steady state mode plus the AFC and FFI of the signal is stationary, I'm ready to apply all my knowledge of DSP and synthesize a digital filter that is close to optimal. By Bayesian I think will not work, because you need sufficient completeness of a priori data (which rarely happens), but the criterion of maximum likelihood or an ideal observer, I think I can do. It will only be necessary to define what is signal (at least in AFC) and what is noise.

In fact, I don't even need the Djuric algorithm itself. I need a 50-100 bar segment of the price chart and the output product of the true Djuric algorithm. Then we will find estimates of the impulse response and synthesize the algorithm itself.

 
Integer:
ASmirnoff:

...whose articles in the Sun are of great interest to you


Does that mean you are the author of that very "Candidate Commando" that we read while serving in the "Armed Forces"?:-)

1. Whose algorithm is better: mine or Djurica's? How much better?

Tell me, mirror, tell me the whole truth...


2. Do you have Djurica's algorithm?

3. What is the difference between them?


Alexander, interesting questions, after such a loud-sounding title of the article. If you're in the business, why don't you take Juric's code apart and understand the algorithm, since you're in the science of this subject?

There is so much water and few qualified explanations in TAP today that these articles are of interest. A correlation analysis of the indicator algorithms showed that about 10-15% are truly original and independent of each other. Djuric is making money on his "brainchild". I have offered you my ideas for free. I guess I should just say thank you for that.
 
ASmirnoff:
Neutron:

Yes, you're welcome!

I ran a cursory glance at the article. I'm sure I just didn't understand the author!

In the place where it talks about the occurrence of group delay (GD) when using an anti-aliasing algorithm, the author offers a recipe for "getting rid" of the latter using a reverse run. ... Is this a joke? It is known that for casual (working on the right-hand end of BP) systems, GZ is unavoidable in principle. But, of course, if BP is defined and we plan to work with it in the middle of a row (not on the right edge), we can, as the author advises, get rid of the lag by re-averaging with the same parameters in the reverse direction. The author does not mention, however, that the averaging algorithm of this kind will inevitably result in overbidding on the last bar. Has he forgotten about it or does he not know? Or what else?

Here is a quote from the article:

"Thus, with the above proposal we can partially compensate for the m/2 time lag (the first drawback of the traditional sliding average). And the second negative effect is eliminated ... And the third, and the fourth. ...

The use of the proposed averaging algorithm also significantly reduces linear frequency distortion... "

The idea of group delay compensation is not my own, but that of American scientists. However, it "worked" outside the real time scale. Well, in radio astronomy, for example. My achievement is that I managed to propose practically an algorithm in the form of a synthetic moving average. Colleague, you have to study the "material part" before you start a pseudo-scientific argument.


Before you send me to the math again, answer a straightforward question: what do you, "colleague", see the practical value of your algorithm for financial time series analysis if it involves constant redrawing of the right edge of the averaging curve? The very edge at which we (traders, MTS) must make a decision!

And please do not appeal to the Americans, astronauts, etc., they all have solved this problem for their task and I have no doubt that they have solved it, unlike you, with perfection. Also, if you run your algorithm over integrated random noise, it will smooth it out just as perfectly. What, are you going to claim now that you can predict the future where it is impossible in principle? I hope not, because if you look at the right edge of the curve, you can see how it is very different (compared to the rest of the smoothing section) from the experimental points. And only after a lot of fiddling (about m/2 points), it takes a "respectable" position. But it's neither hot nor cold for us. In other words, I want to say that all your harmonious building (almost no FC, LA, low "volatility", etc.) rests on the foundation (the possibility of applying the reverse run) that you seem to have sucked out of your hand and that is not applicable to the analysis of market series.

Gentlemen! I am the same Alexander Smirnov, whose articles in "VS" arouse your great interest. I am a scientist, but in the past I was a practical trader. I am very good at TAR. I am an applied mathematician.
:-)
 
ASmirnoff: The algorithm in WPI is correct, but the programmer who implemented it has no experience with the language. The criterion of truth is practice. This algorithm, regardless of the size of the averaging window, provides a delay of 1 bar in the averaging product.

It is very good that the delay is only 1 bar. Is it possible to see it at first hand? Where can I get a ready-made indicator?
 
LeoV, this has to be seen in dynamics. Any smooth curve with a perfect impulse response is worthless (as applied to our tasks) if it is redrawn.
 
Mathemat:
LeoV, this has to be seen in dynamics. Any smooth curve with a perfect impulse response is worthless (as applied to our problems) if it overdraws.

Does it redraw? Well, then there's no way to apply it to our problems. Algorithm SSA (caterpillar) is also a good algorithm, but unfortunately it redraws......
 
<br / translate="no"> Is it possible to see it in person somehow? Where can I get a ready-made indicator?


I'm trying to translate it from Iasi: 


//+------------------------------------------------------------------+
//|                                                   SmirnoffMA.mq4 |
//+------------------------------------------------------------------+
#property copyright "Smirnoff moving average"
#property link      "www.spekulant.ru"
#property indicator_chart_window
#property indicator_buffers 1
#property indicator_color1 Yellow
//---- input parameters
extern int per=4;   // должна быть кратна 4!(4,8,12,и т.д.)
//---- buffers
double ExtMapBuffer1[];
//----
int n;
double alf=0.0,q1[],q2[];
//+------------------------------------------------------------------+
//| Custom indicator initialization function                         |
//+------------------------------------------------------------------+
int init()
  {
//---- indicators
   SetIndexStyle(0,DRAW_LINE);
   SetIndexBuffer(0,ExtMapBuffer1);
   //
   alf=2.0/(per/2+1);
   n=per-1;
   ArrayResize(q1,per); ArrayResize(q2,per);
//----
   return(0);
  }
//+------------------------------------------------------------------+
//| Custom indicator deinitialization function                       |
//+------------------------------------------------------------------+
int deinit()
  {
   return(0);
  }
//+------------------------------------------------------------------+
//| Custom indicator iteration function                              |
//+------------------------------------------------------------------+
int start()
  {
   int    counted_bars=IndicatorCounted();
   if(counted_bars < 0)    return(-1);
   if(per%4!=0)            return(-1);
   int lm = Bars - counted_bars + 1; 
   //----
   for(int shift=0; shift<lm; shift++)
      {q1[0]=Close[shift];
       for(int j=1; j<per; j++) q1[j]=q1[j-1]+alf*(Close[shift+j]-q1[j-1]);
       for(int s=1; s<per/2; s++)
          {for(j=0; j<per; j++) q2[j]=q1[n-j];
           q1[0]=q2[0];
           for(j=1; j<per; j++) q1[j]=q1[j-1]+alf*(q2[j]-q1[j-1]);
          }
       ExtMapBuffer1[shift]=q1[n];
      }
   //----
   return(0);
  }
//+------------------------------------------------------------------+
Files:
smirnoffma_1.mq4  3 kb
 
zigan: I'm trying to translate from Iasi:

So does this indicator redraw or not?
 
No.
 

Alexander! Is this what it should look like? Blue is what zigan did with your algorithm, navy is JMA from Code Base. Both of them have parameter 12.

Here is another picture - when parameter equals 4 for both curves:

Conclusions (if zigan's algorithm translated correctly):

1. Your curve is "prickly".

2. When you increase the parameter, your curve is underlapped, i.e. Djuric handles strong movements better.

123456789101112...44
New comment