The "Better" neural network

In my opinion, Neural Networks is an overrated methodology, especially when it comes to applying it to the beast that is the currencies market. When you remove all the bells and whistles, NN is actually linear in it's implementation.

Bottom line, NN is still a fledgling technology. I honestly don't know how much NN contributed towards the success of the "Better EA".

Better was a 100% NN implementation of 3 different NNs, generating 1300% in a 2 month period.

Gosh, I didn't know it was THAT profitable.

Begs the question why did he even place it in the competition?

He has a fund.... the MQL Championship was good exposure to potential clients.

I agree except you don't know how the 3 NN are configured. From my experience i have had good results from NN but the money man is just as important. NN only give a slight edge and its how you exploit that edge to make a system. In fact 3 NN's suggest he is using majority voting. The 1st NN is configure to look for buy signals. 2nd to sell and the 3rd NN is do nothing! Then you take an EA controlling 5 separate trades. The ist lot is fire off as a taster. The more the market goes to the direction you think the more you add to the trades. As the other NN's begin to suggest a Do Nothing or go short from long the position is unwound. The mistake of a newby system is to use one NN and either go long or short with your entire lots allocated.

However, NN is a hill climbing system. Its NOT predictive it waits till something happens before it can signal something. Genetic Programming is much better at looking for the global search area. In these case its better to use a classification model of majority voting against probabilities of an event occurring.

Regarding the inputs for an NN then use GA to select the inputs and check the outputs. This enables thousands of imputs dials to be tested rapidly compared to the output which is basically how the backtester works. Again in my experience oscillators work best and of the standard range CCI wins hands down showing the best "predictive" quality.

Hi tdion,
Your thoughts on multiple ma's is a good one so do not be put off by negative cooments. One thing I would steer you towards is avoiding general ma's. All timeframe charts have one primary ma that follows the price, it has a guide ma that gives the overall direction. For example in the 4hr chart it is 8ema and 80 ema as the guide whereas on the 1hr chart it is 6 ema and 120 ema.
So you should use these tops and bottoms as your parameters rather than a general tool of a succession of ma's.
Very good luck to you you are on the right track.
Final comment is to find the exact stop so you can assess the risk of the trade. CCI is very good for this.
George

Thanks for the input.

I am training a NN as we speak! I am thinking weekly COT data (and whatever fundamentals I can supply) would be good inputs.... Unemployment, interest rates, etc....

What would be really cool is if your NN could predict COT data based on moving averages....

In other words, you train it with COT and MA over several years, and then see if the NN can guess the COT changes based on MAs.

That would let you know the correlation is strong between price and fundamentals.

Tdion, I have extensive knowledge about NN and trading, I worked on that for many months some year ago. If you know how to use them NNs will give you astonishing results.

This post centers very well the right technique to use:
http://www.forexfactory.com/showpost...7&postcount=45

This other the same:
http://www.forexfactory.com/showpost...11&postcount=8

charlinks says:
He is right. NN can learn to predict a lot of things in a lot of fields but in trading it is way better to train them on trading, not on predicting.

Moreover pay attention to free tools for NNs, they very often are full of bugs.
Here you have only three options:
1) Use free tools and never be sure why they produce bad NNs (it's your fault? the toolkit fault? the bad inputs?)
2) Use expensive commercial products (but I have no experience with them, even if I would like)
3) Learn the math and code everything yourself.

I got astonishing results only when I went with point 3).

Also: you cannot think that you give to a NN some MAs and that's all. You cannot know which data it wants to work properly, the only way to go here is what Bolt said:

I also suggest you use GA even later during training, I made an algo with RPROP and GA combined. In training GA is useful to build the topology of the NN with a cascade-correlation approach.

So choosing inputs and topology must be made with GA, you cannot do it manually, you will never get good results (not in a life, unless of course you already have so much experience that you already know).

You do all that for more NNs and combine them in a committee and trade only when the vast majority of the committee agree. This way you almost get no losing trades, what you get at a certain point is that the committee never agree anymore, so you don't trade and don't lose money. That's when you have to restart all the process and rebuild a new committee.
To know why I suspended my work with NN read this:
http://www.forexfactory.com/showpost...26&postcount=1

But as soon as I can have more CPU power (more PCs) I will resume the whole project. People that says NN are not useful for trading says so because they don't know anything at all on how to use them.

Well, seeing EURUSD to 2.00 on your avatar lost all credibility with me. I don't think we'll touch over 1.50....

Any additional information would be obliged. I think flashing level 2 data every minute to the NN, combined with price and COT, might have some traction.

I am starting to think that exploring super computing aids can be a very viable option...

I am pondering few ideas:

1. Link a bunch of PlayStation 3 and use their parallel CPU power. I have not installed Linux on any but it might be worth trying.
2. Use powerful graphics cards and use the GPU power
3. Use a library for parallel processing that will allow algorithm development in a platform independent fashion (RapidMind comes to mind but that may cost money to license)

EDIT:
Note that CELL blades are about $10K while PS3 is about $400 and only half the power. Big banks do run a lot of analysis on huge blades to analyze how would their market manipulation affect the market... They run these things on about 1000 machines (That's like $10M in blade servers alone)

Oh, well, that's just a joke, but also a hope... I have to buy something in the U.S.A.

Maybe yes, maybe no. NN are black boxes, you cannot guess what will work and what will not with a NN unless you try, and try hard and long. So far my experiences with them tell me that they are not that good at financial markets predictions, but can learn how to trade.

Having written the whole toolkit myself it's not hard at all for me to modify it so that it works distributed on more PCs in parallel (I already did something similar with PCs in my house but the process was too manual, but with little modifications...).
Then the best solution would be to a have a network of traders that would give some CPU of their computers via Internet, as project like seti@home and the like already do.

Ubeee,

I see that you have some experience with NNs and wanted to pick your brain...

I am currently using MatLab to develop my NNs and have created a dll that talks with MetaTrader. I just need some assistance on the network side. What are some good inputs; what is better, prediction or classification. What can you tell me to help in finding the right model.

Your input would be greatly appreciated!


Brent