This EA is based on Deep Neural Network.

To get a copy of the EA contact '[email protected]'.

In our past many years of research, we can give some conclusions regarding forex trading:
1.If your algorithm or trading system doesn't use machine learning or regular repetitive optimisation , then you are simply wasting your time. Because forex is probably the only industry where machine learning has been used for decades even when there was not much computational power available.
2.Already most of the banks, liquidity providers and hedge funds are using highly sophisticated machine learning algos and hence, if you are trying to make money out of the forex market using traditional trading systems, then you are simply trying to make a fool out of yourself and wasting your time. So think about it carefully before even putting any money to any system which doesn't use machine learning.

So now we are mainly working on 'Deep Neural Network' and 'Machine learning' algorithms. This is an example of test results of 'Deep Neural Network'.

If anyone is interested in these algorithms can share his feedback and comments to improve the algo. In near future, we may publish more accounts using machine learning and neural networks.

Or if you are interested to subscribe to the signals of the system, then also you can comment here.

Neural Network trading systems are usually considered as black box trading systems which imitate the behaviour of human brain and it makes trading decisions by identifying certain patterns which are very difficult to identify for a normal human being.

But we will try to explain in details about our system in all of our future posts for those who are interested in Neural network and machine learning algorithms.

Features of our System:
*******************
1. Our current implementation of the Neural network uses optimised weights for specific indicator values used inside the code which are used for making trading decisions by the network.
2.Currently, we are using a 5 layer NN out of which one is the input layer, 3 are hidden layer and 1 is the output layer for making trading decisions. We may add more layers to the system if required.
3.In future versions, we may exclude weights out of the code so that the weights can be optimised by the EA user as well as it might be possible to apply the weights directly to the EA after the optimisation is complete.

Usually NN and Machine learning algos are so complex that there can be an endless discussion on it and still most part of the output produced by the systems are not completely understood by the developers. But as we progress further, we will try to understand more and also, we will update here for others who want to learn Machine learning and neural networks in Forex trading.

Results so far look very impressive! Congratulations on your achievements & getting this EA to the market.

I would be interested in subscribing to a signal here via mql5.com - anyone else interested? Please indicate your support in this thread.

bonco_gordo posted:
Results so far look very impressive! Congratulations on your achievements & getting this EA to the market.

I would be interested in subscribing to a signal here via mql5.com - anyone else interested? Please indicate your support in this thread.

Thank you for your interested and support.

We will update here once we start offering any form of signals.

And, yes, if anyone else is interested in getting the signals from the system, then he can mention here in our thread.

If we get sufficient number of requests within a short period of time, then we may make it available for signals.

Features of the System:
--------------------------------
1.The system uses multiple indicator values as input to the multi-layer neural network.
2.After the inputs are fed to the first layer, it goes through multiple hidden layers to produce output equivalents of the indicator values.
3.Finally, the outputs are used for making trading decisions for buy or sell or trade close signal.
4.The above process runs continuously on every couple of hours.

Though the above process is just a brief summary of what is going on inside the network, but it is very difficult to know exactly how the entries and exits are decided by the network.

Multi-layer neural network works like a human brain where each neuron holds a certain value usually between 0 to 1 and based on the input value a specific neuron is fired up which triggers the next neuron and so on.

In our EA, currently we are using stochastic and RSI indicators and few other indicators as input values to the neural network and the values are further processed in each layer until it reaches the final layer which gives a buy or sell signal.

We are trying to iterate the whole process of training the EA, but it has not been implemented yet. So it uses fixed weights which are optimised through training. We will explain what are weights and further regarding the improvement in our future posts.

New features added to the EA:
------------------------------------------
1.A time filter was added to the EA to restrict the trading of the EA during specific hours of the day
2.Also, Spread filter is added to check the widening of the spread during new events and stop trading during those times
3.Additional hidden layers were added to the network as well as the activation function for individual neurons was modified

In this post, I will further expand on neural networks and about our system improvements so far.

The main advantage of the neural networks is that the it can identify many hidden patterns in the market which usually a normal human being even an experienced trader can't identify.

So the main challenge is to feed the neural network the right kind of input parameters and sufficient number of parameters for classification of BUY and SELL signals at the output for entering trades.

But it again leads to another problem if the number of input parameters become very large, then it takes a lot of processing power or resources to calculate the output value. Also, it takes a lot of time for optimising the network with many combination of inputs.

Another important criteria is choosing the right kind of neuron activation function suitable for the specific algorithm for trade entry and exit. We are partially changing the neuron activation functions in the code to see which is best suitable for our trading system.

Please try it on a real account, demo accounts are just dreams which may or may not become reality

ovisun posted:
Please try it on a real account, demo accounts are just dreams which may or may not become reality

Yes. You are somewhat correct.

But it is not always the the fault of the system which fails in real account, but 99% of the times it is the way the real price feeds are manipulated by the broker individually for individual accounts in all most all so called as regulated ECN brokers who claim to be and we have enough experience in this to figure out such things.

So we don't do any such mistakes in running any of our EA directly in a LIVE account, but most of the times we use a trade copier to copy from demo account to live account which we will do for this EA as well in very near future.

But this system is more sensitive to any changes of any type and hence, the project is still under supervision and development and once we achieve a relatively high pip expectancy of 8 to 10 pips per trade, then we will launch a live account for this system.

OK, thanks for your explanation.

In this post, we will discuss the usage of right stoploss and how we handle stoploss in our system.

In any trading system using the right value of stoploss often determines the success of the trading system. For example, using no stoploss or very large stoploss can give very good results in short period of time, but ultimately creates the possibility of wiping out the whole account balance at anytime.

But what is the right value of stoploss to use is a difficult question to answer and there in no general answer for it. The stoploss depends mainly on 3 things:
1.The time frame in which the trading system is run
2.The type of trading system it is and the number of trades it places per day on average
3.Amount of risk it takes per trade

In our system, we have opened up the possibility to use the system in any timeframe and one additional higher timeframe as filter to the trading signals. So based on which timeframe we use accordingly we set the stoploss.

For a H1 timeframe it is recommended to use a stoploss between 30 pips to 40 pips on average and for a H4 timeframe it should be anywhere between 50 pips to 60 pips.

This looks interesting, but why are you hiding balance / trade history ? I'd be curious to see how this is trading :)

bmigette posted:
This looks interesting, but why are you hiding balance / trade history ? I'd be curious to see how this is trading :)

Well, we never shared any such details of any of our previous trading systems to anyone so far and same is applicable for this one also. That is like our first principle as not to display anything to others other than the performance of the system.

Regarding how it trades, it is impossible to figure it out even while watching live trades on chart, because it uses neural network. So it is useless to analyse such things from trade history unless if someone is trying to create a fake or duplicate version of our EA for resell which obviously we don't allow to happen.

This week we have tested the system with different timeframes to know how it performs. So far it seems to work best in H1 timeframe with a stoploss of around 30 Pips.

Soon we may start testing the system in a small live account. We tried to increase the pips expectancy in this account with current settings, but it doesn't seem to increase and hence, either we will directly run the system in a small live account or we may try to copy trades from a different demo account to a live account.

In case if we will use it directly in a live account, then we have to use machine learning along with the neural network, because simple neural networks are very sensitive to slight price changes which happen in live accounts and sometimes done intentionally by forex brokers.

Now it seems we have perfectly fine tuned the settings for the H1 timeframe and ready to move to a live account. We have done some minor changes to the code to improve the visual representation of EA settings displayed on MT4 chart.

Mainly this week we were tweaking the settings to best fit with the H1 timeframe and these are applied to the current account and the results seem to improve.

Last we week we have done the major changes in adding multiple activation functions to each individual layer of neural network so that a particular layer will use a separate activation function to filter out the BUY or SELL signals. So in the final layer it will completely erase the market noise and produce a perfect bur or sell signal as if a professional human trader will do.

Not only that the key secret of the system is the perfect timing of the exit. This will be available only for the EA users to fine tune the exits even better.

Congrats on the great profit results and very low drawdown. Be good if you could open up any of the stats info for us, to give us some more insights.

After you start it on a live account, be great if you could offer it as a signal.

I would also like to see you offering this as a signal.