A quick and free library for MT4, much to the delight of neuralnetworkers - page 20

New comment
 

Demon_eJ, I think I'm beginning to understand the reason for your "success". (Let's keep it on a first name basis. OK?)
Look, optimization of a two year period (2007-2008) was completed using the first variant of the Expert Advisor. You have sorted the optimization by profit and selected the pass with the highest profit value. You clicked it twice and set new values of external variables. For example, SL is set to 164.
Then you change the testing period to 2009.01.01 - 2009.01.31 and perform a single test.
Have I described everything correctly?
What did you get as a result of this OOS test? Did you get good results? After that you continue to do single tests with this SL=164 at this test period? (steadily improving results)
If the answer to the latter is YES then you are fine-tuning the network already with the January 2009 data, which you cannot do in real life.

 
YES. Almost like that. But I first found all the SL parameters for the whole year 2009, without doing any tests at all, just optimising. Then, when the optimisation was over, I started running tests. In January I have tested 10 times and recorded all profits and then each month in the same way with a new SL. So the network was trained during the test and that was the reason for the good results?
There is no room for improvement with the new variant. One can optimize a perfect straight line, and in the "future" the same value is either a loss or about 0.
 
Demon_eJ писал(а) >>
There is no room for improvement with the new variant. You can optimise a perfect straight line, but in the 'future' the same value is either plum or about 0

There you go. And you were in a hurry for the real. There's no need to rush.
Also it is necessary to understand, that if we have made, suppose, 1000 passes during optimization, and then have chosen the best, in our opinion, pass № 857 with SL=90 for the test, the network will initialize not with those weights which were at the beginning of 857 pass (therefore the result of this pass can not be repeated), but with weights which have appeared at the moment of last pass with SL=90 from 1000 passes of this optimization.
I've got somewhere a variant of this Expert Advisor where all nets (the whole committee) for each optimization pass are written into files. Then any pass can be "fixed", repeated and analyzed.

 
Dear lasso, or maybe someone else will respond...
Could you please change the code so that this EA can trade several copies of one instrument. Currently, when a position is open on a currency, the EA will wait until it closes and only after it closes will it open with the next bar. Perhaps a magic code should be added for this purpose. The purpose of the task - to hang three EAs on one and the same instrument, for example, with different values after optimization.

Thanks in advance!

I am attaching the source FANN-EA
Files:
fann-ea.mq4  9 kb
 
I don't know what the problem is and where to go for help, so I am writing here. After optimizing the Expert Advisor I am running a test using the most profitable variant (profitability of about 2) and the result is not just no profitability of about 2, but it is losing half of it, as if there was no optimization at all. Please tell me what the problem may be. Thanks in advance for the answer.
 
fru1t >>:
Не знаю в чем дело и куда обратиться за помощью, поэтому пишу сюда. После оптимизации советника прогоняю его тестом по самому прибыльному варианту (прибыльность порядка 2), в результате не то что нет прибыльности порядка 2, а он сливает половину, как будто оптимизации и не было вовсе. Подскажите в чем может быть проблема. Заранее спасибо за ответ.

There can be many reasons for this phenomenon:

1. so-called "overtraining".

2. An "inadequate" teacher.

3. Fixed feet.

4. Insufficient number of neurons.

5. Overabundance of neurons.

6....

7...

You can go on for a long time.

Experiment. Notice mistakes (your own).

 
fru1t писал(а) >>
I don't know what's wrong or where to go for help, that's why I am writing here. After optimizing the Expert Advisor I am running a test using the most profitable option (about 2 profitability) and as a result it is not just no profitability of about 2, but I see that it is losing half of it, as if there was no optimization at all. Please tell me what the problem may be. Thanks in advance for the answer.


If you are referring to FANN-EA, the main reason for this "inadequate" behaviour is explained by the three posts above.

 
Can you tell me please, how many inputs and optimization period (for FANN-EA 30 inputs and one year optimization for one hour) are selected manually experimentally or can they be somehow estimated depending on what is at the input and output? I cannot find these parameters to give the right signals for the new data (that were not involved in the optimization) for at least some period of time. Thank you.

And I am also interested in the line
f2M_set_act_function_output (ann, FANN_SIGMOID_SYMMETRIC_STEPWISE);
in the ann_load function of the FANN-EA advisor. If this is normalising the output, why can't the inputs be normalised in the same way?
 
Is everyone here dead? Or doesn't anyone know what to answer...?
 

It's all dead here because (as has already been said many times) the particular EA is of no value in itself, except as an example of how to use the library.

The activation function is the curve of the output of a given function versus its input. It should be chosen based on which parts of the input range are weighted for the analysis.

Normalization is the reduction of input values to a range of -1...1 or 0...1. This is a prerequisite for the normal functioning of the neural network.

1...131415161718192021222324252627...57
New comment