Machine learning in trading: theory, models, practice and algo-trading - page 73

[Alexey Burnakov]

Mihail Marchukajtes:

What???

Although there is some truth in these words. ....

You need to read just what classification means. Any normal site on machine learning has it in a nutshell.

[Alexey Burnakov]

Alexey Burnakov:

Validation data:https://drive.google.com/file/d/0B_Au3ANgcG7COGpJb24wbkxoaTg/view?usp=sharing

There are 5.5 years for 5 pairs. The dates are strictly after the training set. In the far right column, the value: price increment in 181 minutes (what was coded by the categories in the train).

Try to get an MO on validation greater than 0.0001. I got about 0.00013, which corresponds to the average spread. That is, zero.

At the same time I do not trade every observation, but only there where the signal is strong (about 5-10% of observations).

Thank you,

Alexey

And this is a useful picture for making a test sample (crossvalidation). You need to divide the set into 5 equal parts and the tail of each 5th part is future data.

[Mihail Marchukajtes]

Yury Reshetov:
https://bitbucket.org/jprediction/jprediction/downloads

I downloaded it, everything is good, but I have not managed to insert it into MQL yet, and for some reason the optimization results in the saved file are different from those displayed by the predictor itself. The predictor itself has a 90% data generalization level, while the extracted model has only 47%. Well, I haven't managed to run it in MQL yet....

[Yury Reshetov]

Mihail Marchukajtes:
I downloaded it, everything is ok, but I haven't got an opportunity to put it into MQL ... I also could not run it in MQL yet....

Because the java file has the Math.signum() function, and the mql file does not.

Mihail Marchukajtes:
I want to check the correctness of predictor's results and, for some reason, the saved file shows optimization results that are different from the ones displayed by the predictor. The predictor itself has a 90% data generalization level, while the output model has only 47%.

This is a property of committees in machine learning algorithms, which suggests that models combined in a committee will give better results than taken individually. Otherwise, what's the point of committees? Therefore, in jPrediction, individual binary classifiers have worse generalizability than ternary classifiers.

We should also consider such parameter as bias. It is desirable that it be less than 50%. It is even better to have it zero. The smaller its value, the more adequate the ternary classifier is.

[Yury Reshetov]

Mihail Marchukajtes:
I downloaded it, everything is Ok, but I haven't managed to insert it in MQL yet. Well, I haven't managed to run it in MQL yet....

You can also do this. In MetaEditor press Ctrl+H, then make autocomplete:

Then add the signum() function to the code:

double signum(double x) {
   if (x == 0.0) return(0.0);
   if (x > 0.0) return(1.0);
   return(-1.0);
}

[Mihail Marchukajtes]

Yury Reshetov:

You can also do this. In MetaEditor press Ctrl+H, then make autocomplete:

Then add the signum() function to the code:

Well, I'll try it soon, as I'm fed up with construction at home. I'm trading from scratch. I have no time to optimize it, as soon as I do it I will tell you about it at once....

[Mihail Marchukajtes]

Yury Reshetov:

You can also do this. In MetaEditor press Ctrl+H, then make autocomplete:

Then add the signum() function to the code:

I don't understand what variable 1d is, and where did it come from?

[Mihail Marchukajtes]

Mihail Marchukajtes:
I don't get it, it's swearing at variable 1d, what kind of variable is it and where did it come from?

All figured it out, changed it to d1 and everything worked. Especially in the code did not get into, but I understand that this committee and it works on 5 + + +. Thank you, Yuri, for your work. "Now we'll make twice as much hay for our cow" (from Matroskin.....)

[Mihail Marchukajtes]

Now a little about committees, back in 2007 for the first time I heard this term, and we built a committee of three networks, So voot committee knocks out the signal and equalizes the balance only when 2 of the three models work properly, if the committee 2 of 3 give a false signal, then the committee fails. like this ....

[Mihail Marchukajtes]

Yury Reshetov:

You can also do this. In MetaEditor press Ctrl+H, then make autocomplete:

Then add the signum() function to the code:

Yuri, I decided to take a swing at the holy of holies and make a grail, but I got the message that predicted values cannot exceed 10. Is this a deliberate limitation or a limit algorithm???? Because more than 10, it is extremely relevant so....