How to build a NN-EA in MT4 - page 17

 

Neural Networks

Can someone tell me what exactly is a Neural Network and how we apply it to Forex please? I know it has something to do with predicting trends instead of following trends and learning from itself ... but that's all i know. Any help is appriciated. Thanks in advance

 

we have that thread, should be about forecasting system

you should never trust those NN, until you try it yourself on demo -- as market is full of chaos and unpredictability, but NN is more useful in predicting consumer behaviour

 

Neural Network Training Problem

Hello,

Im working on an indicator currenctly that implements a FeedForward Neural Network. It trains itself using Backpropogation. At the Moment i have the indicator set up to train itself on 80% of bars in the chart and test out what its been trained on on the last 20%. I have all inputs scaled from 1 to -1 and all the activation functions are sigmoid & symmetric. I trained the output using a zigzag for entry and exit points. This is where i believe my problem is ... because the indicator shows where a new zig could be formed in realtime. This is good but sometimes it shows me little movements for like 10 pips at a time and then resumes the other trend. I think i have another problem where i have the indicator trained to show only buy(1) or sell(-1) signals, and no in between or sideways action. On that note i added a TP to the indicator so that if the trade was below lets say 25 pips ... it would register as 0 in the neural network. This didnt work either.

So My question comes to this: which data or conditions should i use to actually train the network. If i can train it im almost 100% sure i can get this thing to predict with a decent amount of certainty.

any help is appriciated ... Thankz

 

?

jasondavis87:
Hello,

Im working on an indicator currenctly that implements a FeedForward Neural Network. It trains itself using Backpropogation. At the Moment i have the indicator set up to train itself on 80% of bars in the chart and test out what its been trained on on the last 20%. I have all inputs scaled from 1 to -1 and all the activation functions are sigmoid & symmetric. I trained the output using a zigzag for entry and exit points. This is where i believe my problem is ... because the indicator shows where a new zig could be formed in realtime. This is good but sometimes it shows me little movements for like 10 pips at a time and then resumes the other trend. I think i have another problem where i have the indicator trained to show only buy(1) or sell(-1) signals, and no in between or sideways action. On that note i added a TP to the indicator so that if the trade was below lets say 25 pips ... it would register as 0 in the neural network. This didnt work either.

So My question comes to this: which data or conditions should i use to actually train the network. If i can train it im almost 100% sure i can get this thing to predict with a decent amount of certainty.

any help is appriciated ... Thankz

Have you tried to feed the # of bars between highs and low's to come up with the number before next high/low will occur?

 
Pava:
Have you tried to feed the # of bars between highs and low's to come up with the number before next high/low will occur?

no i havent ... i'll try right now thankz

 

You also could try this...

Have let's say 3 independent NNs (they call it committee of networks).

Always have 3 possible values at your output (as you already figured out). SELL, BUY and DO NOTHING.

Leave that ZigZag as your input for the first network, have let's say 2 MAs crosses as input for 2nd network and Stochastic in overbought and oversold zones as input for the third network.

Now, if you have at least two same signals together (from two networks) then act, otherwise stay aside.

Hope, this helps.

jasondavis87:
no i havent ... i'll try right now thankz
 

OREST

have you got the link to the NN indicator

is it accurate , say abouve 80%

do we install it , just like normal indicator, can we use it on 5M

 

I guess what OREST said,

the presentation will be just like this indicator

-- if you got a good proven method and know a little bit , how to do the coding , at least, program it to the stage to DISPLAY the result of your methodology, you may PM me to get the CODE of this indicator -- MQ4 code

see whether you like this way !

Files:
 

xx3xxx,

No I haven't seen reliable NN indicator yet. NN learning curve is very steep and one has to be proficient in both Math and trading in a way to succeed.

Yes, your indicator could be used as input for one the NNs I mentioned.

xx3xxx:
I guess what OREST said,

the presentation will be just like this indicator

-- if you got a good proven method and know a little bit , how to do the coding , at least, program it to the stage to DISPLAY the result of your methodology, you may PM me to get the CODE of this indicator -- MQ4 code

see whether you like this way !
 
Orest:
You also could try this...

Have let's say 3 independent NNs (they call it committee of networks).

Always have 3 possible values at your output (as you already figured out). SELL, BUY and DO NOTHING.

Leave that ZigZag as your input for the first network, have let's say 2 MAs crosses as input for 2nd network and Stochastic in overbought and oversold zones as input for the third network.

Now, if you have at least two same signals together (from two networks) then act, otherwise stay aside.

Hope, this helps.

Thankz thas an awesome idea, i'll try that right now ... unfortunately the periods since last zig method didn't exactly work