Machine learning in trading: theory, models, practice and algo-trading - page 988

[Alexander Ivanov]

Good afternoon 😂

Without a powerful genetic algorithm you can't create a neural network.

First create a code-genome.

Then connect it to the NS.

👍

[Grigori.S.B]

Aleksey Vyazmikin:

Machine learning is based on features (patterns/features) that will highlight an event. Accordingly, you need to specify what you need to look at, and the MO algorithm will try to find any patterns in what is shown and work out the rules of behavior.

And how to formalize these signs and rules? As applied to the head/shoulders pattern. Or maybe there is another example? I am not asking for a grail, I am interested in the methodology of formalization.

Aleksey Vyazmikin:

And accordingly, the more observations, the more accurate the rules will be on a longer period of history.

And what about the training sample? I thought that it was just the substitute for formal rules that are not always easy to describe analytically.

[Aleksey Vyazmikin]

Grigori.S.B:

And how can these signs and rules be formalized? As applied to the head/shoulders pattern. Or maybe there is some other example? I'm not asking for a grail, I'm interested in the formalization methodology itself.

This is a creative process - there may be different solutions. Probably the simplest would be to give the network information on ZZ in the form of its structure - the ratio of segments to each other.

Grigori.S.B.:

And what about the training sample? Thought that just she replaces the formal rules that are not always easy to describe analytically.

If we're talking about NS, there's a search for functions that describe the rules, if about the decision tree or forest, there are more formal rules, but there, of course, they are obtained during the learning period, so I said that the more observations, the better, in my opinion.

[Ivan Negreshniy]

Grigori.S.B.:

Questions from a newbie. Please advise on methods of applying machine learning. For example, a trader has found a certain pattern in the market. Suppose it is a GP (head-shoulders) pattern. Variants:

1. Traded hands and has a history of profitable and losing trades.
2. Found this pattern on the history on the charts and can mark the entry/exit points.

Can I use this history/statistics for machine learning in variants 1 and 2? How can this be done? Approximately how many trades are needed for training (minimum/maximum)? Will the algorithm recognize patterns only on the TF it is trained on? Will the MO algorithm "understand" that the trader's deals were made on the GP pattern, and if it "understands" it how? How many bars of history to the opening of a position will the MO analyze?

Both of your options are suitable for machine learning and there are tools that automatically generate ready-made EAs or indicators based on marked charts or trade reports.

In this case, what the trained model "will understand" and what will be its performance will depend on the quality and quantity of data, the type of model and the initial parameters of training, the MO will analyze the depth of history, which you give it.

To build your own model you can use the many available libraries, examples from articles, etc., and if you want a simple test to assess whether it is worth it, I can help you with your example.

To do this you will need to put your signals on the chart and save it as a template in your MT4 terminal and also specify the number of bars by which you identify your pattern and I can teach and generate a test, neural network or random forest model for you using these data.

[Igor Makanu]

Maxim Dmitrievsky:

the developers are machinlerners for Christ's sake.

hmmm, imho the most sensible statement in all the pages of the topic that I have been able to study! ))))

I know it's not realistic to read 1000 pages, but I would like to learn machine learning

Due to the fact that in this period of my studying the markets I decided to "take off the rose-colored glasses" and try to find a machine learning tool that works somehow....

i have already drawn up the indicator ind_Weierstrass.mql4 - it builds the Weierstrass function - i can visually select the necessary chart parameters, the indicator normalizes the obtained values, and i displayed the normalization results in the Expert Advisor log

here is the script_WeierstrassHST - it creates a history file with the symbol name in the settings (NZDUSD by default), the chart period TimeFrame, the number of historical bars History, the number of digits of the symbol Digit, tick volumes volume = 5

Basically, it is enough to create history data

then using a standard PeriodConverter from MT we convert all the timeframes (5,15,30,60,240,1440,10080,43200)

here the historical data is ready - with default settings, the daily and above TF did not work :( - I didn't want to pick parameters input double Weierstrass_A = 0.33; input double Weierstrass_B = 1.5; input int Weierstrass_N = 10

disconnect Internet in the terminal (logout) and you can work as with the normal chart

So, as far as I suggest, the Weierstrass function should return the result of machine learning >90? (I mean forward testing)

Who can show me such a machine learning? - I am interested in a specific example

Thanks in advance!

[Maxim Dmitrievsky]

Igor Makanu:

So, as far as I'm concerned, on the Weierstrass function the machine learning result should be >90? (meaning forward testing)

Who can show me such a machine learning? - I am interested in a specific example

thanks in advance!

Yes, the cycles there can easily be predicted even by eye, but it has nothing to do with the market since they are non-periodic.

If something can be predicted by eye, then the MO can do it too. The market, in the general case, is not predicted by eye

I don't know how to compare cycles with B-M fx and the market, except through correlation, and through corr. bullshit.

[forexman77]

Maxim Dmitrievsky:

Yes, the cycles there can easily be predicted even by eye, but it has nothing to do with the market because they are non-periodic

If something can be predicted by eye, then the MO can handle it. The market, in the general case, is not predicted by eye

I don't know how to compare cycles with B-M fx and the market except through correlation, and through corr. bullshit.

I guess it is possible to do it through correlation. But, the number of bars must be specified and the cycles will often have different number of bars, so the example from the past will not always be fully covered.

I have not yet found a way around it. After all, the correlation should have the same size arrays...?

[forexman77]

Igor Makanu:

hmm, imho, the most sensible statement in all the pages of the topic that I have been able to study! ))))

Why create the HST file, you can write the indicator data to a file and send it to the network? Or is it done for something else?

If you want to explain the gist of what this function does "for dummies" (I've read it on the web, I failed to understand anything, because the description there is in scientific language).

There are no questions about understanding of the code, everything seems to be clear.

[Igor Makanu]

forexman77:

Why create an HST file, you can write the indicator data to a file and feed it into the network? Or is it done for something else.

2.Weierstrass function, if it is not difficult, can you describe the essence of what the function does "for dummies" (I read it on the web, I can not understand anything, because there is a scientific language description).

There are no questions about understanding the code, I think everything is clear.

1.I deliberately made hst that if someone is already configured in MT, then immediately would give the result

2. wiki to your aid, the function is periodic - so far it's not important, what matters is the hardware, which can work correctly

Maxim Dmitrievsky:

Yes, the cycles can easily be predicted even by eye, but it has nothing to do with the market since they are non-periodic.

If something can be predicted by eye, the MO will do the same. The market, in the general case, is not predicted by eye

I don't know how to compare cycles with B-M fx and the market except through correlation, and through corr. bullshit.

I now want to go from the opposite direction - there is non-random data, it means that the hardware should give an excellent forecast and then it may become more complicated

I.e. I suggest not to pull the mapper to the data, but to give the data that the mapper should process with a 100% guarantee

[Maxim Dmitrievsky]

forexman77:

Through correlation it is probably possible. But it is necessary to specify the number of bars, and the cycles will often have a different number of bars, so the example from the past will not always be fully covered.

I have not yet figured out how to get around that. After all, the correlation should have the same size arrays...?

Yes, well, you can fill them with some intermediate values

The only thing I use in fractals is zechal symmetry and sometimes I get nice patterns with good inputs on charts. But by hand, not automated in any way

like this one from the last one.

sometimes multifractals work too - formations repeat one after the other but with different periods, exactly like in the B-M