Machine learning in trading: theory, models, practice and algo-trading - page 455

[Mihail Marchukajtes]

Well, charge your algorithms. Set up your turnkey systems, for there are two datasets coming, which will show the whole essence of them. The verdict on them will be interesting, and preferably the built models, so that you can test them in real work.

[Mihail Marchukajtes]

Vizard_:

OH... Sensei, thank you))) Haven't laughed like that in a long time))))))))

XM.... I don't get you...... if your model is some kind of secret, you don't have to post it. Or do you think that your model will work forever and you kind of give the grail in the wrong hands???? What makes you laugh so much????

In fact, I personally adhere to the following postulate "The architecture of the network plays an insignificant role in model building. The main thing is to be able to properly train at least one neuron".

Of course, I didn't express it correctly, but I can't think of anything sensible to say since this morning. I just wanted to say that you can get a model in the form of a polynomial, without any secrets, etc. ordinary polynomial .... BUT it will carry in itself generalizability and workability in the future. It is just trained correctly, but this is an art, it is not the NS itself, but its environment that is responsible for it.....

If I post my model, the safety of my NS will not be violated, because it is not the polynomial itself that is important, but the method of obtaining it....

double getBinaryClassificator1(double v0, double v1, double v2, double v3, double v4, double v5, double v6, double v7, double v8) {
   double x0 = 2.0 * (v0 + 854.0) / 1708.0 - 1.0;
   double x1 = 2.0 * (v1 + 6025.0) / 12050.0 - 1.0;
   double x2 = 2.0 * (v2 + 83.31175) / 166.6235 - 1.0;
   double x3 = 2.0 * (v3 + 94.01706) / 190.59075 - 1.0;
   double x4 = 2.0 * (v4 + 79.89835) / 162.32691999999997 - 1.0;
   double x5 = 2.0 * (v5 + 8139.72596) / 16279.45192 - 1.0;
   double x6 = 2.0 * (v6 + 44728.87178) / 89457.74356 - 1.0;
   double x7 = 2.0 * (v7 + 16744.5244) / 33489.0488 - 1.0;
   double x8 = 2.0 * (v8 + 3571.0) / 7142.0 - 1.0;
   double decision = -0.3081635711326393 * sigmoid(x4)
  + 1.0249861163756249 * sigmoid(x2 + x4)
  + 0.09986072300626747 * sigmoid(x1 + x2 + x3 + x4)
  + 0.5413533668890985 * sigmoid(x5)
  + 0.6997159366377829 * sigmoid(x0 + x2 + x5)
  -0.04588868418500921 * sigmoid(x0 + x1 + x2 + x4 + x5)
  + 0.16781775869820087 * sigmoid(x1 + x2 + x3 + x4 + x5)
  + 0.4607985948890632 * sigmoid(x0 + x6)
  -0.8671700766023466 * sigmoid(x1 + x6)
  -0.5270267887837945 * sigmoid(x4 + x6)
  -0.027936937492837814 * sigmoid(x0 + x4 + x6)
  -0.6617354089719066 * sigmoid(x1 + x3 + x4 + x7)
  -0.19638937616247806 * sigmoid(x1 + x5 + x7)
  + 0.8684769002935395 * sigmoid(x6 + x7)
  -0.5967137681478805 * sigmoid(x0 + x2 + x5 + x6 + x7)
  -0.2097815643098296 * sigmoid(x0 + x1 + x4 + x5 + x6 + x7)
  -1.5340457322179422 * sigmoid(x8)
  + 0.7646273899667675 * sigmoid(x0 + x8)
  + 0.27679539504420725 * sigmoid(x2 + x3 + x8)
  -0.37855134296518955 * sigmoid(x1 + x6 + x8)
  -0.0311654310975556 * sigmoid(x0 + x1 + x2 + x6 + x8)
  -0.6036203203370856 * sigmoid(x0 + x1 + x2 + x3 + x5 + x6 + x8)
  + 0.04123987376920568 * sigmoid(x3 + x7 + x8)
  + 0.8450984194705711 * sigmoid(x0 + x1 + x2 + x3 + x4 + x7 + x8)
  -0.8578008338989624 * sigmoid(x2 + x3 + x5 + x7 + x8)
  + 1.059103470465344 * sigmoid(x1 + x3 + x6 + x7 + x8)
  + 1.0514388283102527 * sigmoid(x0 + x1 + x2 + x4 + x5 + x6 + x7 + x8)
  -0.06758350008374249 * sigmoid(1.0 + x0 + x3 + x4)
  -0.24213702035383408 * sigmoid(1.0 + x4 + x5)
  -0.8011798876969051 * sigmoid(1.0 + x0 + x1 + x3 + x5 + x6 + x7)
  + 0.7506445968459932 * sigmoid(1.0 + x0 + x3 + x8)
  + 0.3049328737780207 * sigmoid(1.0 + x0 + x1 + x2 + x3 + x6 + x7 + x8);
   return decision;
}

[Dr. Trader]

Why are there two tables?
One has target 1 and the other 0? Why does each table already have target 0 and 1?
Can I merge them into one?

[Mihail Marchukajtes]

Dr. Trader:
And why two tables?
One has target 1 and the other 0? Why do we have target 0 and 1 in every table?
Is it possible to combine them into one?

One for buy signals, the other for sell signals.....

To combine them into one, you need to make an inversion for one of the tables. Then one and the same model can be used for buy and sell signals.

And so you get two independent models, one to buy, the other to sell....

[Mihail Marchukajtes]

Vizard_:

It's a disease))))
You realize that fuck, at least for the sake of respect should first correct the cap))))) that the Target does not have the property of completeness and people use not only kv and so forth ..,
but also stupidly divide in half.)) Do you realize that this post isn't just a stele either? No ... so in the next 13 years, you will write the same rubbish)))
For laughing, let me remind you for the tenth time that Reshetov's infernal machine has two errors, one of which can be partially leveled... but this contraption
will always lose at least 2%... on Doc's dataset, you can get 55.3% on the OOS (test)... but of course these are just parrots...

Do you know what these errors are???? Since you speak with such confidence.....

[Dr. Trader]

I tried Buy15MALL, the model found some correlation between ST9, AD5, Volum7, VVolum7, other inputs are completely ignored. The accuracy is 55.6%. I cannot share the model. Try re-training the Reshetov machine on just these four.

Vizard_:

You can get 55.3% on the OOS (test)... but of course these are just parrots...

Cool, it's not even parrots, it's a prediction of eurusd m5 bar gains over the last 7 weeks or so. But considering the spread I think trading will be at a loss.
I will try the same experiment with different lags, like open0-open1, open0-open2, etc.

[Mihail Marchukajtes]

The point of this endeavor was to determine the following.

Either my data are collected correctly and they will work with other algorithms and optimization systems.

Or it is all the same miracles of Optimizer Reshetov, which can make a worthwhile model out of any data.

[Dr. Trader]

Vizard_:

It would be better if you practiced on the image, because the point is not to process it, but...

I agree, I need quality noise reduction.
I kind of learned how to train a model with crossvalidation so that the accuracy does not drop too much on new data. Like in this case, if a couple of years ago I would get 100% on the training and 50% on the test, now it's just 50% there and there. But that's obviously not enough. When I learn how to squelch, I'll probably squeeze out a couple of percent.

[Maxim Dmitrievsky]

Yuriy Asaulenko:

I checked on stock instruments. Unlike a few years ago, all movement there is 15-20 minutes... and silence.

In Forex, yes, the minutes rather do not rule. If I try to use it I'll see how much more accurate I am.

Probability distribution is the same in any TF, i.e. the probability of exceeding 3 sigmas and catching a black swan, what are we talking about?) TF is just another representation of the same graph

[Aleksey Terentev]

Dear Sirs, before making market forecasts, you would have to familiarize yourself with the problematics of the task. Predicting time series is not the classification of plants and cancers by chips.

Mihail Marchukajtes:

Mikhail, there is only one question.

Why in the dataset of the timeseries is your data sorted?

What is there to predict?
At least how to divide the dataset into a training and a test one?