Don't tell me then that TA doesn't work - page 14

[[Deleted]]

artikul:

I'm shocked too )))) Probably because TA doesn't work after all ))))

One thing is clear - the proposed method of optimisation really works.

All that is needed is to mathematically demystify the method and break it all down.

But we need a simpler Expert Advisor, with more understandable signals to clearly see all BPs in the analysis

Which signals stay and which ones get cut off and why.

[Vladimir Gomonov]

Reshetov:

The most useful thing here is normalization of perceptron output for comparison with a threshold value. I haven't applied it in my EAs yet, but I need to try it.

I normalize all the oscillators. And not on the maximum (it's not wise, no matter how you look at it), but on the variance.

//--- Расчёт коэффициента нормализации -----------------
double GetCn(const double &source[],long len) // Возвращает коэффициент нормализации
  {
   double qsum=0;
   for(uint i=0; i<len; i++)
     {
      double t=source[i];
      qsum+=t*t;
     }
//   if(Prints)Print("QSum == ",qsum,"; Len == ", len);
   qsum=sqrt(qsum/len);
//   if(Prints)Print("CNorm == ",M_SQRT1_2/qsum);
   return(M_SQRT1_2/qsum);
  }

Comment on the application of the root of 1/2 as a normalising factor. This is chosen for the reason that the variance of the sine wave = 0.5, so the standard deviation = sqrt(1/2).

I.e. the standard deviation of the resulting oscillator will be the same as that of a sinusoid.

--

I also normalize the coefficients of perceptrons. Optimization (fitting) is faster. Here is my version of "lattice" in mql5.

// I don't attach the tested indicator due to modesty. If you want to test it, you can add any of your indicators.

[Vladimir Gomonov]

Ah yes! It won't work without a market driver either. Putting it in the trailer. No comment (it's simple).

[hrenfx]

MetaDriver:
I normalise all oscillators in general. And not by the maximum (which is not wise, anyway), but by the variance.

There was no normalization by maximum.

[Vladimir Gomonov]

hrenfx:
There was no rationing at the most.

I'm sorry. :)

[hrenfx]

MetaDriver:
Sorry. :)

Now I don't know... Happy Holidays!

About rationing:

You need to be clear about what you're doing and what you're doing it for. For example, it is important to know the range of values of the normalized value. The conditions for achieving the extreme values of the interval. And so on.

It is because of rationing considerations that relative rather than absolute returns work. And in general, the view of price VRs carries a certain universality.

This is not an elucidation, just a thought out loud.

[Vladimir Gomonov]

hrenfx:

1) Now I don't know... Happy Holidays!

2) About rationing:

You need to be clear about what you're doing and what you're doing it for. For example, it is important to know the interval of values of the normalized value. The conditions of reaching the extremes of the interval. And so on.

It is because of rationing considerations that relative rather than absolute returns work. Well, in general, the view of price VRs carries a certain universality.

3) This is not an elucidation, just a thought out loud.

1) Likewise! :)

2) There's total agreement.

3. I see.

Anyway, I'm glad that you're interested in combining TC. (I probably still have some thoughts on cooperation). I intuitively agree with alsu on the theme that regardless of how the quotes are combined, their properties are unlikely to change much. Just a little bit. :) But with TS it's a completely different matter. BPs of trade signals may have quite different and very different characteristics. There is a great scope for constructing quite meaningful combinations.

By the way, for a long time now (about a year and a half) I have been building and testing only systems that modify the net position on every bar, i.e. with "continuous periodic output".

// This does not mean that I will place them in such a version on the market. It's important to clearly understand what is being done and for what purpose.... :-))

[Alexey Subbotin]

MetaDriver:

BP trading signals can have very different and very different characteristics. There is a huge scope for building very meaningful combinations.

Only theoretical basis is needed - lately something like this has been glimmering here ... When several people's minds start working in the same direction at once, it's not good))

By the way, for a long time now (about a year and a half) I have been building and testing only systems modifying net positions on every bar, i.e. with "continuous periodic output".

Well yes, those are the easiest to analyse.

[Vladimir Gomonov]

alsu:
1) I just need a theoretical basis - there's been a glimmer of something lately... When several people's minds start working in one direction at once, it's not good)).
2) Well yes, those are the easiest to analyse.

1. Let's just cut the theoretical bullshit already oops, I mean of course! ;) Well... I have some thoughts. But it's the greed that's stifling, as usual... :) Maybe some time later...

2. It's not just about that (although that's fair enough). It allows you to select sets of TCs that have "good" additive properties. I.e. they are more prone to summing profits when signals are added together.

[Alexey Subbotin]

MetaDriver:

xre....

I'm waiting for an article from our Mathematician on dependence-independence research as a starting point. In my mind I'm already screwing the theses there to the subject of this thread... It looks like it should be the theory that will explain the sifting out of "true-fit" signals.