Sunday Jun 04, 2006
Another Attempt to Predict Stock's Price by Neural Network
This is the 2nd version of my very primary neural network to predict the Price of stock market on BlogTrader Platform. 800 sample data were used to train the network, and generalization was applied forward and backward.
Here's the result:
Click on the picture to enlarge it
Posted at 01:17PM Jun 04, 2006 by dcaoyuan in AIOTrade | Comments[10]
I have a hard time believing this will work. The extrapolate effectively, there has to be information in the training data that relates to long term trends.
Or are there secrets hidden in the stock trends that can even predict decisions by people within the company? I don't believe so (at least nothing super reliable), and I have to imagine these kinds of things have been tried many times before.
Still, sound like a fun experiment. Could be fun to see more results.
Posted by Tom on June 05, 2006 at 10:31 AM PDT #
Hi,
I'm working on something similar as well, and I hope it'll work. Though, as you said trining the NN is tricky and the NN may not be enough by itself. I actually already found your site quite a few times in the past :), and I thought it would be nice to integrate my solution (if it turns out to perform well enough).
Tom, I'm not very into the markets (and that is the very reason I'm trying to use some soft computing methods :) ). But I do believe that there is some information hidden in the price graphs. It may be tricky to find the right training data (and it may turn out that the best is not to use the price info just before the period you want to predict). But as there are some 'rules' that the analysts and trades use, it must be in the price charts. Think of all the (more or less primitive) technical indicators, the Elliot cycles(!), etc. There may (must) be undiscovered rules. Of course a NN won't ever be 100% effective but it's enough (for me at least :) ) if it's as effective as a good trader. Of course a prediction or trading method is effective only as log as it is not used by too many traders...
Posted by atleta on June 05, 2006 at 11:11 AM PDT #
Caoyuan,
The word "Tring" does not exist in English. I think what you are trying to say is "attempt", "test", or "experiment" ...
Posted by Gert on June 06, 2006 at 02:50 AM PDT #
Gert,
You are right. Thanks
Posted by Caoyuan on June 06, 2006 at 07:26 AM PDT #
Hi atleta,
As you said:"I do believe that there is some information hidden in the price graphs. It may be tricky to find the right training data (and it may turn out that the best is not to use the price info just before the period you want to predict). But as there are some 'rules' that the analysts and trades use, it must be in the price charts."
That's what I also experienced. The selection of sample data will truely affect the result, and I'm not sure what's the rule hidden in it.
Usally, I give out a guess, and then try it. Again and again.
Posted by Caoyuan on June 08, 2006 at 05:46 PM PDT #
Hi Caoyuan
yes, input data matters a lot. I have checked a few articles, and that is the most interesting conclusion. You have to experiment a lot (even in automated way - think GA...), use some statistical method to find the correlating sets.
The interesting and sad thing about markets and trading is that you can't really publish what you do because it makes your results (if you have any :) ) unusable in the future. I researched the web on this topic a lot (unfortunately I do it quite lot instead of doing the atual development :( ) and it seems that a lot of people and companies use neural networks. So useing NNs is not a competitive advantage anymore, you have to be more tricky. BTW I don't think that a simple MLP will produce any good results
P.s.: Blog comments are the stupidest way of communication ever inveted :). How long are you gonna check back for comments? And how long am I?
Posted by atleta on June 11, 2006 at 01:09 PM PDT #
nn and ga are no secret and it's not the theory (if you can read and ever heard of the invention of books) but the application that's tricky. hedge funds use methodologies you probably will never hear of, so yes, keep your simple nn secret - they will never guess what you're at unless you distort the markets with the big profits you
're gonna make :-) just kidding. ta works best in predicting the markets but how do you program it? nn and ga and ats and backtesting are really fun though: good to play around and try and test new ideas. i thought about implementing a framework myself or buying a cheap commercial software. unfortunately there's not much open source software around. this project is the only exception so far as everybody seems to enjoy reinventing the wheel for himself :-) i hope the progress keeps progressing to a workable environment before it collapses. and i hope that a community picks up working on it. it would be a pity to see/let it die.
Posted by what does humai mean? on June 13, 2006 at 07:37 AM PDT #
The UI part of nn will need more works to be done. Anyway, although I reinvent:-) the wheel for myself intertesting, I'll feed back to this project progressively.
Posted by Caoyuan on June 13, 2006 at 02:34 PM PDT #
Caoyuan,
The more i follow your ideas, the more I come to believe that your NN work would lead to success in the forex (foreign currency exchange) market: It is a 24 hour market, can not be manipulated as easily as stocks, has a much higher volume (ca. $2 trillion !!! per day) and is strongly influenced by news events - especially in the USA - on a regular basis.
As I posted in another thread, I currently use my own system based on MySQL and PHP, tracking and processing forex data with quite excellent results. Your own strategy might easily lend itself to similarly good results.
Think it over, I'll gladly give you more info, if you wish.
Gert
Posted by Gert on June 15, 2006 at 02:18 PM PDT #
student
Posted by 196.1.217.237 on May 01, 2008 at 06:56 AM PDT #