EAs WITH BRAINS - Aritificial Neural Network Trading.

Hi all,

 

I am working on building an EA based on the neural network. The biggest problem I think is processing power. The time taken to complete calculations on one instance is huge. I am looking for contributions of ideas on how to improve calculations performance. You all are invited to share your success and problems faced while coding a neural network trading system.

 

Regards to all,

 

Amit

i am a EA programmer may be i can help you... contact me at [email protected]....

You may be able to pre-calculate some of your calculations and then pivot off of tme with modification calculations.

For example....you want to go from San Francisco to Narita, Japan. With North being 0 or 360 the trip can start off going West (270 degrees).....this would be your initial calculation...go West.

Then periodically you recalculate your direction to intersect with Narita. Lets say the first calculation comes back with a Plus 17....so now you recalculate your initial direction of 270 and add 17.....now you are headed at 285 degrees. Hopefully a lot of the calculations were in the original and less on the additional periodic updates.

Slice the objectives up and cycle through them.
Lets say your goal is to determine the probability of hitting any of the following:

> 25 pips over the current price
20 - 25 pips above the current price
15 - 20
10 - 15
......
10 - 15 below
15 - 20 below
20 - 25 below
> 25 below

now cycle through these and get a value for the probability for each point starting from the current price and then step up and down one step. If you hit a very low value then you can focus on the other side assuming that your system is telling you that the probability is 80% probability of hitting + 10 or more pips within a time bound area. (like within the next hour)

To get more computing power you may want to use several computers and have them each optimize on one question you are asking your NN. For example: What is the probability of a move greater than 20 pips that is in the Long direction? One computer can just keep working on this question....and when it has it....it starts over so that it is continuous.

Consider making the NN act like a video game. In some video games you add life...for our purposes this will be profit. Consider if you make $100 you can spawn a new LIFE. If you lose some amount, say $50 you lose a life....maybe you need to borrow from more successful trades. Now you train the NN to add more life and protect against death....much like a colony of ants.

Best wishes,

Steve AKA Forex5x = Skype

>>The biggest problem I think is processing power.

i would be interested to learn more about your design specifics regarding how you have implemented this neural network in a currency trading application.

thank you.

I have implemented in my 'AM' expert elements of neural strategies to 'learn market' with short period. It is recycle learn period every week. Now, I'm concidering to make some patent on it - if its possible.

I do not under estimate your coding skills, but having an articial brain for EA we need NASA team.

Thats why I don't have implemented full artificial brain , but its elements and thats why I have a cycle learn because of high data collection.

Beowulf cluster?

Neural Networks dont work easily. Interfacing them with MT4 is almost impossible. Those that have done so, use the same configeration and therefore are not any better than the previous NN EAs that have be built. Your better off working on and converting indicators to become a auto EA trader and have % and ratios under control. DLL and other external controls dont work well unless your doing some human interaction on the side. then its not a normall NN or DLL. Good luck. Breaking new ground with interfacing NN is a great goal, but will be hard to manage. Many of those that claim NN, are not real. Even my own home built EAs, seems to be thinking, but, they are just collecting data and or counting history bars. Comparing RSI, MA, ATR and others isnt NN. But in my book, its close as we can get.

Garry
aka
Professor53

I wonder how many different inputs you use for your NN training? Looks like you are just curve fitting your strategy to recent historical data.
So far most of NN trading systems have failed, so have realistic expectations. The ones that have so-so results use very few inputs for learning and more or less guess very short term price direction.

Professor53 posted:
But in my book, its close as we can get.
Garry
aka
Professor53

ah.. now he is selling books instead of super-duper VPS's?

amitjamuda posted:
Hi all,

 

I am working on building an EA based on the neural network. The biggest problem I think is processing power. The time taken to complete calculations on one instance is huge. I am looking for contributions of ideas on how to improve calculations performance. You all are invited to share your success and problems faced while coding a neural network trading system.

 

Regards to all,

 

Amit

Look in the MLQ4 forum, you'll find implementations of NN.
You can get what you need by studying them.

thanks for so many suggestions! I came to an understanding that a true NN is very difficult to conceptualize and design.
It also requires a huge amount of computing power to be of some good in the forex trading. Also, we are not sure as to how it will behave....

I came to a conclusion that it is good to for study, however to code one and test one requires quite a lot of resources.

amitjamuda posted:
thanks for so many suggestions! I came to an understanding that a true NN is very difficult to conceptualize and design.
It also requires a huge amount of computing power to be of some good in the forex trading. Also, we are not sure as to how it will behave....

I came to a conclusion that it is good to for study, however to code one and test one requires quite a lot of resources.

Just what have you learnt that led you the the conclusions you have arrived at?

We had a college working on a system with neural network. For some reason he couldnt win on the long term however he had computers calculating all the time...

skaboy posted:

amitjamuda posted:
thanks for so many suggestions! I came to an understanding that a true NN is very difficult to conceptualize and design.
It also requires a huge amount of computing power to be of some good in the forex trading. Also, we are not sure as to how it will behave....

I came to a conclusion that it is good to for study, however to code one and test one requires quite a lot of resources.

Just what have you learnt that led you the the conclusions you have arrived at?

There are lot of things learnt. You may contact me at [email protected]

'EAs WITH BRAINS' ....name of the topic made my day! 😁

pls contact me i hav something for you

You can checkout my blog (https://mechanicalforex.com/2014/01/neural-networks-in-trading-building-a-powerful-trading-brain-with-several-nn.html), there are many other posts as well on the subject (you can use the search bar to look for other articles on my blog). I have built many NN and other machine learning implementations using constant retraining, etc. I generally use the FANN and Shark libraries for the building of machine learning strategies. With our programming framework it's very easy to build complex, constantly retraining neural networks and live trade them on MT4/MT5/JForex/Oanda. I hope this helps :)