Question on neural network programming - page 2

New comment
 
grasn:
Mathemat:

Thanks, grasn. I too stopped dreaming after my amateur experiments (a couple of years ago), but apparently I haven't worked that cycle out yet - especially since I haven't even taken on qualifying NS...

Yeah, you're welcome. Don't mind my pessimism about NS. Everyone has to go their own way. By the way, long time researches of signal structure helped me a lot to develop the model which I developed on the materials of a friendly forum(https://www.mql5.com/ru/forum/50458). It so happened that the ideas stated by Vladislav and many other participants of discussion (I don't mean Alex) are very well laid on my own experience and understanding of processes.

PS: By the way, I recommend MineSet for research (if any pattern needs to be found), developed by SGI and sold here: http://www.purpleinsight.com/ when SGI collapsed. There is a necessary set of Data Mining tools, including classification, as well as excellent visualization capabilities (after all, SGI created it, and no one came up with a better than the eye).


The NS actually works very well. Your pessimism is understandable, I went through the same situation. Standard approaches do not work unambiguously, apparently you have not investigated related applications. Neuroshell2 is a very good and very clear program for studying NS. There are practically all types of NS. You can easily connect it to MT4 and immediately see the results.

And in general, any NS can be easily programmed directly in MQL4. The same way you can pick weights of NS by means of GA MT4 or your own. Pessimism is defined only by lack of imagination and fantasy. In principle, there are no limits...

 
Mathemat:
Yes, dob-zorge, that's what you need to input, not predict it.
That's what I'm wondering, is it possible for the NS to give an approximate Close.
Or should we be guided by the NS output signal.
In calculations NS uses formula of sigmoid function F(x), output equals -0.5 to 0.5.
 
klot:

The NS actually works very well. Your pessimism is understandable, I went through the same situation. Standard approaches don't work unambiguously, apparently you haven't explored related applications. Try it. A very good and very clear programme for studying NS is Neuroshell2. There are practically all types of NS. You can easily connect it to MT4 and immediately see the results.

And in general, any kind of NS can be easily programmed directly in MQL4. Also you can select the weights of the NS using the GA MT4 or your own one. Pessimism is defined only by lack of imagination and fantasy. Basically, there are no limits.

For example, the Fibonacci levels work not worse than the most sophisticated NC. And this is not pessimism at all, but just a conclusion based on common sense, research and comparative analysis of NS forecast results with other models and strategies. I used NeuroSolution for my work, also a very good tool (MineSet for finding patterns).

Just take my word for it, I have plenty of imagination and fantasy, and have tried a lot of things. I'm certainly not a categorical opponent of using NS, but I've drawn conclusions for myself (you're right, standard approaches don't work at all).

In any case, I wish you more successful predictions. :о)

 
grasn:
klot:

The NS actually works very well. Your pessimism is understandable, I went through the same situation. Standard approaches do not work unambiguously, apparently you have not investigated related applications. Neuroshell2 is a very good and very clear program for studying NS. There are practically all types of NS. You can easily connect it to MT4 and immediately see the results.

And in general, any kind of NS can be easily programmed directly in MQL4. You can also pick weights of NS using GA MT4 or your own. Pessimism is defined only by lack of imagination and fantasy. Basically, there are no limits...

For example, the Fibonacci levels work not worse than the most sophisticated NC. And this is not pessimism at all, but just a conclusion based on common sense, research and comparative analysis of NS forecast results with other models and strategies. I used NeuroSolution for my work, also a very good tool (MineSet for finding patterns).

Just take my word for it, I have plenty of imagination and fantasy, and have tried a lot of things. I'm certainly not a categorical opponent of using NS, but I've drawn conclusions for myself (you're right, standard approaches don't work at all).

In any case, I wish you more successful predictions. :о)

Who told you that neural networks are needed for predictions?
 
Hydrometeorological centres, fortune tellers and soothsayers do the forecasting.

Maybe everyone should mind their own business, and give neural networks the tasks they are good at!
For example: recognise the image of a gallowsman or a ghost?
 
VBAG:
Forecasts are made by the hydrometeorological centre, fortune tellers and soothsayers.
There are also stock market analysts:))
 
Integer:
VBAG:
The hydrometeocenter, fortune tellers and soothsayers do the forecasting.
Also stock market analysts:))
Yes, of course, and many more who are never prepared to take responsibility for their words.
Or their wording goes on the verge of wordplay, so that it can always be interpreted exactly the opposite. Like, "Well, we told you so....".
Yeah, that's not really the point.
I didn't mean to offend anyone,
I just hate the word prediction, and it's a shame about neural networks. ..,
and with them, for the country.
Regards to all the participants.
Vladimir
 

I will describe a little bit my representation of the data to the neural networks. For example, in my Expert Advisor on Champ, one of the nets takes ratios of averages as inputs:

       ema15 = iMA("EURUSD",PERIOD_M15, 199,0,1,0,1);
       ema13 = iMA("EURUSD",PERIOD_M15, 132,0,1,0,1);
       ema8 = iMA("EURUSD",PERIOD_M15, 66,0,1,0,1);
       ema7 = iMA("EURUSD",PERIOD_M15, 46,0,1,0,1);
       ema5 = iMA("EURUSD",PERIOD_M15, 6,0,1,0,1);       
       
       input[0] = (ema15 - Close[1])*15000;
       input[2] = (ema13 - Close[1])*10000;
       input[1] = (ema8 - Close[1])*12000;
       input[3] = (ema7 - Close[1])*15000;
       input[4] = (ema5 - Close[1])*10000;
       input[5] = (ema15 - ema8)*12000;
       input[6] = (ema13 - ema7)*15000;
       input[7] = (ema8 - ema5)*10000;

Multiplication by numbers, to get smoother inputs.

Then the grid was trained on randomly pulled signals from history (let's say 400 is very little => remembering any period simply eliminated). The outputs I have made are as follows:

1.0 - more than 70 pips up and less than 30 pips down during a day.

0.9 - 60 pips up 25 pips down

0.8 40 up 20 down

0.75 - flat

0.7 - 40 pips down 20 pips up

0.6 - 60 down 25 up

0.5 70 down 30 up

After a day of learning, good results are obtained when testing with other samples (not from the training)... Then the EA is developed, and if we lower the bars (say, not 70 points of Provit, but 20), the results are impressive.

Good luck to all.

 
There you go, plan, there's a classification after all, and that's encouraging. Thanks for the idea!

P.S. What's the grid, if it's not a secret? I still fiddled with Jordan/Elman.
 
Mathemat:
There you go, plan, there's a classification after all, and that's encouraging. Thanks for the idea!

P.S. What's the grid, if it's not a secret? I still fiddled with Jordan/Elman.



Actually, I wrote everything myself. In C# :) That way it's easier to understand and implement something of my own. For example, I have a modified mesh learning algorithm. The meshes are multilayer (for example, 8 - 60 - 20 - 1), combined into committees: each one implements a different idea.
1234567
New comment