Return Styles: Pseud0ch, Terminal, Valhalla, NES, Geocities, Blue Moon.

Pages: 1-

Applying ANNs to the stock market

Name: Anonymous 2010-02-15 23:04

I was wondering if anyone else has bothered doing this.  I have a few PDFs sitting around that cover neural nets pretty well (representations & learning), and a few that are from 90s that illustrate (understandable) networks for projecting stock price.  Last I recall, one of those papers from early 2000 generally did better than random picks.

I want to try it, knowing that I won't likely make any money, thanks to the superior algorithms already playing in the market on sub-millisecond trades.  Sort of wanting to get shot down for infeasibility or have my anus haxed by someone better-versed than I am in this field.

Name: 2 GET 2010-02-15 23:17

ANNs are probably too small to do submillisecond trades, but beyond that, there's a definite possibility of success if you can find a way to deliver the *massive* amount of information needed to describe the stock market to it

Name: Anonymous 2010-02-15 23:17

>>2
s/small/slow

Name: Anonymous 2010-02-15 23:41

sed: >>3: "s/small/slow": unterminated substitute in regular expression

Name: Anonymous 2010-02-15 23:51

Yeah, if I go for this seriously, I would take the ANN approach with the typical data inputs (index open/close, day of the week, etc.), then optimize on the neural network configuration through genetic algorithms (simply picking a good number of nodes and layers).  Data can be pulled off of yahoo/google finance for most of this.  Since that's really not that much index data, you would definitely want to be creative in picking new data inputs (I've heard of things like picking "key words" out of a company's quarterly report, and so on).  News items would be great to integrate.  Of course, it's a delicate balance; it is probably easy to overwhelm the network with useless input.

The first step is to probably work on an ANN network optimizer (usually a genetic algorithm that closes in on a seemingly optimal number of nodes/layers), and perhaps design it to be distributed (this would be really cool, and is on my giant todo list for when I get free time).  Then, you take that general system and start crapping data sets into it.  Computing power is cheap, so you could test a lot of data, and the goal would be to optimize ANNs automatically (beside network configuration, you could also optimize inputs).  Still, screw aorund with it too much, and you start destroying the ability of the ANN to generalize in order to make predictions.

It's all still pretty magical to me...

Name: Anonymous 2010-02-16 0:37

Neural nets are only good for playing tic-tac-toe and enlightening Sussmans.

Name: ​​​​​​​​​​ 2010-09-08 22:23

Name: Anonymous 2011-02-03 1:08

Name: Anonymous 2011-02-17 20:08

that's cool and all, but check my doubles over there

Don't change these.
Name: Email:
Entire Thread Thread List