Neural Network Trading, Serious people only!

now i am aware that price does not depend on the inputs i listed above (for instance, COT contracts.) but my thinking was, what will central bankers base their decisions on? afterall, to know where forex is going, you'd need to get inside THEIR heads, right?

The Question) what makes a currency attractive to buy/sell? interest rates and economic reports are the best i can come up with.... and COT shows the big money, although i bet a lot of them are hedges (the contract could be a loser, but the business could make heavy profit within their industry.... in other words, the risk they hedged against)

notice that i am not saying previous price shouldn't be an input, but not the only input.

my software is by Sciengy Team and called Recognize, Predict, Forecast! it is free, and will create a back propagated network after training.

Been reading some papers lately on some benchmark tests of different neural network models and predictive ability on eurusd. All of them agreed Gaussian Mixture Models are the "best" yet.I thought some of you might be interested in this info.

one other application of a NN i'd like to try is the opening of the london session.... basically training my NN with 1 minute bars......

i'd train it with only one hour of data/day (60 1 minute bars that occur during the session open), and the idea is to see if the price will continue or reverse for the remainder of the session.

thoughts?

Train it recognize 'the myth'? Interesting concept.

. sorry about this..pressed the wrong buttons disregard the post

Charlinks,

What happened? Did you hit it big or give up?

I am looking for some positive feedback from someone that has had some success with NN. I am currently working with MatLab to creat a NN. Tried lots of things. I have already implemented the dll that works with MetaTrader and all I lack now is the decent net. Anyone having any luck? Not much out there on this...I am assuming that once you get it you are gone!

Brent

I hit it big but NN are only a secondary thing to the system. NN by themself won't do you any good.

It's like expecting a 5 year old (that has a fresh clean not cluttered with crap neural network in their brain) to learn how to trade and be successful without any outside education or knowledge.

They may start sucking less at it... but most likely won't become successful.

Good luck to everybody.

My prediction was pretty good... Almost prophetic!

http://www.forexfactory.com/showpost...ostcount=11338

I've been testing with FANN and a C++ testing framework on and off for the last year. I've also been testing with GP and GA. Genetic algorithms are just fast optimizers, Genetic programs, produce incomprensible results. Forward testing has failed for all the above.

I belive FANN can't be trained to use high, low, open and close or any combination of raw indicators, etc.

On my todo list is use reinforcement learning with a neural network (FANN).

http://leenissen.dk/fann/forum/viewtopic.php?t=269

Also, since ANNs are good at pattern recognition, has anyone tried training with a series of candlestick patterns? I found this nice little categorization in TA lib that categorizes the following information about a candlestick and was thinking about running it through FANN.

#define TA_REALBODY(IDX) ( fabs( CLOSE_[IDX] - OPEN_[IDX] ) )
#define TA_UPPERSHADOW(IDX) ( HIGH_[IDX] - ( CLOSE_[IDX] >= OPEN_[IDX] ? CLOSE_[IDX] : OPEN_[IDX] ) )
#define TA_LOWERSHADOW(IDX) ( ( CLOSE_[IDX] >= OPEN_[IDX] ? OPEN_[IDX] : CLOSE_[IDX] ) - LOW_[IDX] )
#define TA_HIGHLOWRANGE(IDX) ( HIGH_[IDX] - LOW_[IDX] )
#define TA_CANDLECOLOR(IDX) ( CLOSE_[IDX] >= OPEN_[IDX] ? true : false)
#define TA_REALBODYGAPUP(IDX2,IDX1) ( min(OPEN_[IDX2],CLOSE_[IDX2]) > max(OPEN_[IDX1],CLOSE_[IDX1]) )
#define TA_REALBODYGAPDOWN(IDX2,IDX1) ( max(OPEN_[IDX2],CLOSE_[IDX2]) < min(OPEN_[IDX1],CLOSE_[IDX1]) )
#define TA_CANDLEGAPUP(IDX2,IDX1) ( LOW_[IDX2] > HIGH_[IDX1] )
#define TA_CANDLEGAPDOWN(IDX2,IDX1) ( HIGH_[IDX2] < LOW_[IDX1] )

Dan

redpine,

Can you elaborate on what it is about FANN, or what is it lacking to allow training of High, Low, Open, and Close data?

Also, have you ever worked with Cortex? Just google "cortex neural"

If so, do you suppose it has the ability to train high, low, open, close data?
I'd appreciate any thoughts you may have on its capabilities versus that offered by FANN.

I originally tried, high, low, open, close and never got it to give good results. ANN's are mainly used for pattern recognition. What is the pattern that can be recognized? High is higher than low? A pattern like with Japanese candlestick patterns is a series of this data characterizations over time. A single point of time is a very simple pattern. Right now I do all my ANN training with 12 candle sticks and characterize them as follows.

1. Real Body size
2. Upper shadow size
3. Lower shadow size
4. High low range
5. Candle color
6. Body gap up
7. Body gap down
8. Candle Gap up
9. Candle Gap down

I then take the min/max values over the next 4 hours as the output and train the ANN. The approach looks promising, but I haven't been working on this a lot lattely. Any suggestions would be appreciated.

I've never used cortex. I did demo test Pelatron. It seems to come to a solution very good, but I never hooked it into my framework to see if it would be a good predictor.

I've tested, ANN's, GA and GP they all are really good curve fitting approachs, but not good predictors. How to not curve fit has me perplexed.

Dan

Great concept but I think the failure will be due to "static" limited scope of a single ccy pair. If you combine the wicks of 4 pairs I think you'll come up with a much more reliable trade. I've done this myself just looking at EURUSD GBPUSD USDCHF USDJPY. If all have bullish USD wicks then start buying dollars. This works most of the time. I would like to keep in touch with you on this front. I have been looking for coders to do precisely this from getacoder.

Not true. ANNs can be used for both function approximation and pattern recognition as well.And i'd say they're used mainly for function approximation, not for pattern recognition because the most common type of ANN used is the MLP.

I've actually been working on a system for the past months that involves ANNs as well as other things from the more general field of DSP and I do function approximation with them, not pattern recognition. Unfortunately I can't tell you exactly what it is that I do As a suggestion though, if you're interested in using NNs and GA or GP, read as much as you can on DSP-related technique and try to combine them.

You could start learning R language and use the R statistical package for all of this since they have a bunch of free add-ons for almost anything you could need, or if you're a student you could try buying a Matlab student license which should be fairly cheap...or a Wolfram Mathematica license.

If you want my opinion, it's not really worth continuing on implementing an ANN-based pattern recognition tool for candlesticks.Mostly because even if you could manage to get it to correctly recognize patterns, a pattern does not have a 100% guarantee that price will react the way it predicts it will...sounds to me like a whole lot of trouble for not much.

Hi Charlink ?

very very great...

How do you do that ?
Can you tell us here... please ?

regards

hi,
that is a good idea and your method seems to be a form of Cluster (community) analysis, please refer to the link,
Mr IN10TION has made a good news reader and some good CPairs Cluster (community) analysis and graphical and numerical monitorings...

http://www.forex-tsd.com/news-signal...sreader-1.html

hi to all ,
these are good resources for NN library in C and C++:

1-FANN library by Steffen Nissen (Fast Artificial Neural Network Library)
"http://leenissen.dk/fann/index.php"

2-LWNN library by Peter van Rossum (Light Weight Neural Network Library)
"http://sourceforge.net/projects/lwneuralnet/"

3-A good hardware acceleration for large computational application:
by using GPU multi processing threads (NVIDIA CUDA , AMD ATI...)
Nvidia is provided a good C library for using GPU power in your application
a comparision:
(Intel Core 2 Quad Extreme QX9770 : less than 100 GFLOPS (single 32bit float) or (double 64bit float) )
(NVIDIA GTX260 , GTX280: 933 GFLOPS(single 32bit float) and 100 GFLOPS (double 64bit float))
"http://www.nvidia.com/object/cuda_home.html#"

Hi everyone,

This is one of the best threads on nn's that I have seen sofar.
I am using the nn indicator from Cortex, I took it off the web and put it into a metatrader indicator script. I have to say it works very well if you know how to use it.
I wanted to create an EA using 2 nn signals and trade based upon that.

Very simple. If you train an nn it will recognize patterns and predict the future price based upon that. The patterns within the market are well known, there are the fibonacci patterns and the elliott waves, and they are highly correlated.
What you need then for a successfull EA is a good money management system. Those are available as well, there is a very good one here on ff forums that can be used.

My (simple) cortex nn indicator indicates the tops, reversals and bottoms very well.
I am sure that with a little help we can make a very profitable EA.
What I am looking for is not scalping 10-20 pips here and there, that you don't need an nn for. No, a system that can predict the next move up or down. Set clean trades with tight stops. I believe that can be done with nn.

Anyway, I am going to read all these posts and check back in. This is exciting! Thank you for all your efforts!

Wow:nerd:that was a lot of reading.

Anyway, I have an example of how to use NN with Elliott trading. And it is here that I think NN's can be very very useful.
I took the picture that was posted earlier and put some elliott numbers on them.
Elliott works great, and in relation to fibo numbers you can, in certain patterns, come to almost 100% certainty.
For instance, coming out of a wave3, a 4 wave will retrace 23.6, 38.2 or 61.8% in a normal zigzag formation. It will do so in 3 waves as it is a correction. Once it reached a fibo number in 3 waves, you can put your trade with a very short stop. Works (almost) everytime.
Now there are some problems. If you mislabeled or are in one of those nasty "correction" patterns, it is very hard to make a good prediction of where it is going.
Neuronetworks can help. They are 'trained' to recognize patterns, and will recognize wave and fibonacci patterns.
For this reason, I like the NN to be working on actual price action, not some lag indicator. With a lag indicator you will predict what you are seeing right now, where based upon price action you will predict what is coming.
I have installed the Cortex Indicator. I don't use the EA because it is not accurate enough. However, it does show you where the big moves are. It is very precise in predicting the 3 waves, which are the biggest waves.

So when in doubt about the next move, a NN indicator would be perfect and the savvy trader can then forecast almost exactly where the next move will go to. Creating a profitable EA for that is then not far away.
I haven't seen anything like this yet, and I guarantee you, that it would be a hit.

hi
can we pick this thread up again? Is anyone here using NNs successfully?
Also has anyone here had any experience with Joone?

G

...which is why I too was waiting to hear more from folks such as tdion and CodeBreaker.

My feeling is that some sort of hedged (basket) mixture model (Dow, oil, etc) makes the most sense. This may be done with less effort using other methods though, which is similar to the way that I currently trade.

I found this to be a very interesting thread and would love to hear more. Unfortunately I'm not qualified to answer any deeper questions - my background is in AI as opposed to finance.

Best, Christopher.