Finding a set of indicators to feed into the neural network inputs. Discussion. A tool for evaluating the results. - page 3

New comment
 
ivandurak >> :

I too am very interested in finding a minimum set of indicators and evaluation of the results, but for my own purposes.

Only instead of a close price it is necessary to use the result of trade.

Cgm ... You forget that for maximum learning efficiency the network inputs must be statistically independent, there must be no correlation between the data fed to each input. All machines are corrected with each other, you can check. There is a quite handy and simple software - AtteStat, it is an add-on to Exel, but very handy.
 
rip >> :

That's exactly what I mean...


How do you form a vector which you then pass to JGap, is it just a vector of W values or are they encoded W values.

What is a target function? I can give you an example - if we take as a target f-function E[i](t) = D[i](t) - Y[i](t), where E is error, D is value expected on output, Y is value obtained by input of training sample X, i is neuron norm, t is epoch number. If we take E[i](t) = Sign(D[i](t) - Y[i](t))*(D[i](t) - Y[i](t))^2 on a number of tasks, the result is much better. Say, if we form a series reflecting attractors of classical dynamical systems (Lorenz, Henon, Rössler,...) we can even train the network to approximate such data, not deeply but still.


I haven't tried it with series formed from currency quotes :) because I don't think it will work :)

No. I only pass the value of the target function to the genetic algorithm, and the genetic algorithm outputs a vector of values for each gene, which I convert into a matrix of neural network weights.

 
IlyaA >> :
With a design like this, you can achieve a near vertical eviti with no slippage. Will you address the issue of retraining at the neuron?

The retraining issue is on the back burner for now... I'll take 2 months of M5 (it's 12*24*22*2=12,000+ values) and use them to train a neural network that has 150 -300 scales... I think it's a long way from retraining here

 
rip >> :

And overtraining may not happen ... If the author cites as a graph of error on a test sample, you can tell at a glance what happens to overtraining.

What error are we talking about? the target function is larger - so the gene is more appropriate...

 
IlyaA >> :


I agree. he is working with a black box. overtraining is very likely. Dear iliarr can you publish the training graph.

I'm logging the thread number, generation number ( to within 10), target function value... I don't think this information can tell you anything about retraining... I don't think there is any retraining because the training sample is much larger than the number of weights in the neural network

 
joo >> :

You shouldn't be using the waving arms. Or rather, you should not be using only moving averages. Try to experiment with a set of different types of indicators, preferably each indicator's algorithm should be radically different from the others. Then you will get more information for the network.

One more point.

You are using a reverse trading system based on NN signals. This is exactly the same as the standard muvingaverage expert. No better or worse.

Look for a way to determine the size of SL and TP with NN, and ways to accompany open positions. You can open at random as well.


GA is just an optimization tool (screwdriver for the machine). With minimal differences you can use it or any other optimization algorithm (screwdriver).

This was the main question for which I created the topic... What set of indicators to use? I don't know enough about indicators to be able to make a good choice, and I don't have enough resources to do a stupid search... If you have a complete set of indicators, i'd appreciate it.

When i've got the real time intel, i don't know what i've got and i don't know how to do it.

 
iliarr >> :

I log the thread number, generation number (to within 10), the value of the target function... I don't think this information can tell you anything about retraining... I don't think there is any retraining as the training sample greatly exceeds the number of weights in the neural network


The public needs to see a graphical dependence of the learning error on time (number of epochs).
 
12000 values :-D with so many weights it's a lot.
 
ivandurak >> :
What if we do it on the principle of the lucky monkey. For example let's take CCI and check it on all available history, we will choose profitable sectors that will not lose all the time. Then we take Momentum, Bollinger, Muvings and choose profitable areas. Trading is done virtually and a system that shows as good as the initial selection is admitted for real trading. If history repeats it should work. Also, the advantage of this approach is an approximate estimate of the duration of a good situation. What are your criteria for selecting profitable areas like number of trades, the average transaction, maximum drawdown, duration of a profitable area, I have a small idea, I'll tell you later.

Did you mean it? Or maybe you need some modifications?

a[0]=iCCI(Symbol(),0,12,PRICE_TYPICAL,0)
a[1]=iMomentum(NULL,0,12,PRICE_CLOSE,0)
a[2]=iBands(NULL,0,20,2,0,PRICE_LOW,MODE_LOWER,0)
a[3]=iMA(NULL,0,13,8,MODE_SMMA,PRICE_MEDIAN, i)
 
iliarr >>

This was the main question for which I created the topic... What set of indicators to use? I don't know enough about indicators to be able to make a good choice, and I don't have enough resources to do a stupid search... If you have a complete set of indicators, i'd appreciate it.

If you've got a good working knowledge of the detector, don't worry, I'll get it right away.

i've already got a good working knowledge of the indicators, i'll give them a good result.

and in 2 days you'll have your own opinion.

12345678910
New comment
Reason: