Machine learning in trading: theory, models, practice and algo-trading - page 735

[Yuriy Asaulenko]

Nikolay Gaylis:

That's why I asked for the formula...

If you want to work with NS - go to a good professional software, or don't bother with those MQLs, they will be useless.

[[Deleted]]

Probably this one.

[Maxim Dmitrievsky]

[Maxim Dmitrievsky]

Nikolay Gaylis:

Probably this one.

well there is an article on the subject, how to write a neuron, yes, just like that... search works

Don't listen to Asaulenko, he is fed up here with his paradigm, it is unclear why he hangs out at this forum, it's good that he was banned today, this helpless nerd

[Maxim Dmitrievsky]

It's funny :)

https://smart-lab.ru/blog/456953.php

[Mihail Marchukajtes]

quietly rustling the slate my dep grows .....

I'll comment on you later... no time now...

[Mihail Marchukajtes]

toxic:

Well man, Michael..., how you do not understand, already and explained and trolled you to emotionally motivate to evolution, but you are like a rock! It would make sense if you were selling some sort of grail to a sucker, but you don't seem to be, so it's weird, not reasonable. You can not describe the whole market with 40 observations, moreover 3 weeks, it's like describing the faces of thousands of people with 40 pixels, take even for example a photo of Vladimir Lenin and pull out of it as you like, using any data transformation ~40 points and try to recognize the leader of the proletariat in it))) And the whole market is not just one photo, it's hundreds of times larger in capacity. You can not so enthusiastically "pull the wishful thinking" on the reality.

Your problem is that you are trying to describe the entire market. Every bar, as I understand it... Now let's do the math together.

The 15M TF for two weeks is 1920 bars. If you are describing the HUGE market, then you need to feed 1920 bars to the network input, etc., etc.

TF 15M Signals to buy 40 pieces (approximately). Me to describe these two weeks on my TS, you need to submit only 40 values for training, so that the network could learn these two weeks, because I do not analyze the whole market, I analyze it only in the moments of its reversal. The basic TS is counter-trend. That is, it determines the areas of possible market reversals. And it is at this point that the analysis takes place. Which significantly reduces the number of samples during training, but at the same time covers the same time interval (2 weeks)

As I said before, I can't increase the training sample because the data I'm using doesn't allow me to do so. If the input data were better, I would train for 100 and 1000. BUT BUT it's not important, it's the end result that's important and it's so......

[Mihail Marchukajtes]

This picture shows the training section and the EOC.

Here is just a section of OOS from 01.31.2018 To look more clearly

And here is the section from Monday 03.05.2018 TC is the same...

[Mihail Marchukajtes]

And all this is the work of these two babies, which were trained on the necessary data and selected at the maximum VI relative to the input

double getBinaryClassificator1A(double v0, double v1, double v2, double v3, double v4) {
   double x0 = 2.0 * (v0 + 2748.0) / 2951.0 - 1.0;
   double x1 = 2.0 * (v1 + 83.09069) / 154.45321 - 1.0;
   double x2 = 2.0 * (v2 + 71.06971) / 147.16595 - 1.0;
   double x3 = 2.0 * (v3 + 94.29885) / 172.688 - 1.0;
   double x4 = 2.0 * (v4 + 70.91128) / 154.99767 - 1.0;
   double decision = 0.07032014810363377 * x1 * x3
  -0.2709385389305134 * sigmoid(x0 + x1)
  + 0.4766552616529839 * sigmoid(x1 + x2)
  -0.02475017204446986 * sigmoid(x3)
  + 0.6522278547266189 * sigmoid(x4)
  -0.4251146155411889 * sigmoid(x0 + x4)
  + 0.3491339620629828 * sigmoid(x1 + x4)
  -0.11995134291612954 * sigmoid(x0 + x1 + x3 + x4)
  -0.5414699867210747 * sigmoid(x2 + x3 + x4)
  -0.15299357377557646 * sigmoid(x1 + x2 + x3 + x4)
  + 0.3477721452733811 * sigmoid(1.0 + x2 + x3)
  -0.2667852400383829 * sigmoid(1.0 + x0 + x2 + x4)
  + 0.35137296333271945 * sigmoid(1.0 + x1 + x2 + x4)
  + 0.5545211348150159 * sigmoid(1.0 + x1 + x2 + x3 + x4);
   return decision;
}

double getBinaryClassificator2A(double v0, double v1, double v2, double v3, double v4) {
   double x0 = 2.0 * (v0 + 1543.0) / 2763.0 - 1.0;
   double x1 = 2.0 * (v1 + 83.27445) / 157.86037 - 1.0;
   double x2 = 2.0 * (v2 + 96.96413) / 167.20560999999998 - 1.0;
   double x3 = 2.0 * (v3 + 76.54987) / 162.84452 - 1.0;
   double x4 = 2.0 * (v4 + 70.10687) / 136.14457 - 1.0;
   double decision = -1.4629648549243972 * x2 * x3
  -0.24382747582073286 * x2 * x4
  -0.16956988148753577 * sigmoid(x0)
  -0.09466097943059529 * sigmoid(x1)
  + 0.09458009807928075 * sigmoid(x2)
  + 0.5855852404304591 * sigmoid(x1 + x2)
  + 0.5480350088543795 * sigmoid(x3)
  + 0.030113404168369433 * sigmoid(x1 + x3)
  -0.146080234300504 * sigmoid(x4)
  + 0.26372003133088134 * sigmoid(x1 + x3 + x4)
  -0.40493035689960494 * sigmoid(x0 + x2 + x3 + x4);
   return decision;
}

Question: Who is not satisfied with these models?

[Mihail Marchukajtes]

I am absolutely confident in my approach to the market, looking at the results.

Thanks for the help in R, which made the work of the TS times better.....

The approach is labor-intensive and there is a lot to keep in mind so as not to make a mistake, which can fundamentally change the result, but overall I am satisfied, as I wish you....

And now I'm starting to write an article on BO + there will also be a video, so do not miss it. When it is published, I will definitely let you know in this thread.... Good luck!!!!