 |
|
| Computers Forum Index » Computer Artificial Intelligence - Neural Nets » Building an ANN for market predictions... |
|
Page 1 of 1 |
|
| Author |
Message |
| Kjetil... |
 Posted: Tue Jun 23, 2009 12:42 am |
|
|
|
Guest
|
I've been looking for an interesting problem for learning
Clojure/functional programming, and after seeing
<http://www.fatvat.co.uk/2009/05/neural-networks-and-clojure.html>, this
idea bubbled to the surface.
I guess it's a bit lofty, and that even if I should manage to come up
with a network which produces perfect all output for all my training
data (not going to happen, I know), prediction is a different matter
entirely.
The plan is to rewrite the some object oriented/imperative, Python
and/or Java based ANN code I've found online in a functional style,
without looking at the source mentioned above again.
Will a simple feed-forward network with x+1 input nodes, y hidden layers
and z outputs do? Every node connected to the output from all nodes in
the preceding layer.
Do I need more than a single neuron type, using tanh as the activation
function?
For inputs, I'm thinking along these lines; probably have way too many:
- +1, bias node: as all the ANN code I've seen has one. This is an
exercise in cargo cult programming anyway, so. 
- % change in closing price over the last 1,2,3,4,5,14,50,150,365 days
for Dow Jones, Nasdaq and Nikkei. normalized to [-1, 1] (max % gives 1?)
- % change in volume for the same indexes/days, normalized to [-1, 1]?
- % change in closing price for the target market main index,
normalized to [-1, 1]
- Interest rates. 15% gives 1, anything above is chopped off? current,
1, 2, 6 and 12 months ago?
- Exchange rate vs. USD. Change in %? Same resolution as the indexes?
- Exchange rate vs. YEN.
- crude oil prices. change in %? Same resolution as the indexes?
- day of year, normalized to [0, 1]?
- day of week, normalized to [0, 1]?
Hidden layers and nodes per layer? Not the faintest idea what to go with
here. Any point in allowing a variable number of nodes per hidden layer?
Outputs: What kind of predictions should I be training for? Both
direction and magnitude in a single output is probably a bit optimistic?
Am I better off aiming for (separate) indicators on fall and rise of
more than X%?
In either case, I'd like to see some kind of output for the next day,
the day after that, a week ahead, and a month ahead. (yea, I know..)
I have daily data for the last 10 years. If I look harder for exchange
rates; about 30. I plan to train the network on samples in random order,
as the market is bound to change over time.
I know I'm aiming a bit high, and my main goal is just to write some
nice functional lisp, but if you have pointers or want to shoot me down
right away, any input is appreciated.
Thank you, |
|
|
| Back to top |
|
|
|
| tomhoo... |
| Posted: Mon Jun 29, 2009 8:15 pm |
|
|
|
Guest
|
A word to the wise:
Anything simple will not work.
This is because if that were true, everyone would be doing it and
everyone would be rich.
That is not to say it cannot be done.
There are so many esoteric issues with this modeling that it is
pointless to discuss them in a 30 second post.
However, I will tell you probably the biggest thing of importance:
Write your own code.
That way, months/years from now, you will not find yourself boxed into
a corner after investing countless hours
in someone else's platform. |
|
|
| Back to top |
|
|
|
| Stephen Wolstenholme... |
| Posted: Tue Jun 30, 2009 2:59 am |
|
|
|
Guest
|
On Mon, 29 Jun 2009 13:15:31 -0700 (PDT), tomhoo <tomhoo at (no spam) gmail.com>
wrote:
Quote: A word to the wise:
Anything simple will not work.
This is because if that were true, everyone would be doing it and
everyone would be rich.
The only person I know to have had any repeatable success with market
predictions uses a very simple method. The key to his method is
avoiding too much data!
Steve
--
Neural Planner Software Ltd www.NPSL1.com |
|
|
| Back to top |
|
|
|
| unixops... |
| Posted: Fri Jul 10, 2009 9:40 am |
|
|
|
Guest
|
Dear Kjetil,
Please use some careful reasoning. Please don't forecast data, an
understanding of economics and mathematical modeling is extremely
important. Consider actually studying economics. Understanding market
structure or a practical model is more important than your data. The
model is the data, this establishes a specific criteria to further
develop, improve, and most importantly validate what you are doing.
Sincerely yours,
Doug |
|
|
| Back to top |
|
|
|
| Tim Guetzkow... |
| Posted: Sun Aug 16, 2009 5:39 pm |
|
|
|
Guest
|
Am Mon, 22 Jun 2009 22:42:01 +0200 schrieb Kjetil:
Hi Kjetil,
you would be amazed how many guys had a similar idea. ;)
There is a lot of stuff to find through google.scholar: financial
forecasting, neural nets / genetic programming and so on.
BUT, i'm not going to shoot you down, because im also interested in this
topic. ;)
I answer you because:
A) I would like to get in touch with you for mindstorming etc.
B) I would LOVE to hear more about this
Quote: I have daily data for the last 10 years. If I look harder for exchange
rates; about 30. I plan to train the network on samples in random order,
as the market is bound to change over time.
It seems to me very difficult to get reliable data in this field. Just
stocks (lol)? Also currency? Please give me a hind how to get my hands
this data!
You are right, imho it is one of the major fields for NN away from
science. As said before, write your own code. There is lots of stuff
and .lib's in the net to learn from. I would like to share.
Please, if you are still following these thread (what i sadly don't
expect  ) show a sign of life.
--
sharkbyte
Yeah i know: bad english. I will improve... |
|
|
| Back to top |
|
|
|
| diana kay... |
| Posted: Fri Aug 21, 2009 12:57 am |
|
|
|
Guest
|
On Aug 16, 12:39 pm, Tim Guetzkow <i-hate-s... at (no spam) no-fool.org> wrote:
Quote: Am Mon, 22 Jun 2009 22:42:01 +0200 schrieb Kjetil:
Hi Kjetil,
you would be amazed how many guys had a similar idea. ;)
There is a lot of stuff to find through google.scholar: financial
forecasting, neural nets / genetic programming and so on.
BUT, i'm not going to shoot you down, because im also interested in this
topic. ;)
I answer you because:
A) I would like to get in touch with you for mindstorming etc.
B) I would LOVE to hear more about this
I have daily data for the last 10 years. If I look harder for exchange
rates; about 30. I plan to train the network on samples in random order,
as the market is bound to change over time.
It seems to me very difficult to get reliable data in this field. Just
stocks (lol)? Also currency? Please give me a hind how to get my hands
this data!
You are right, imho it is one of the major fields for NN away from
science. As said before, write your own code. There is lots of stuff
and .lib's in the net to learn from. I would like to share.
Please, if you are still following these thread (what i sadly don't
expect ) show a sign of life.
--
sharkbyte
Yeah i know: bad english. I will improve...
Hi,
I have developed an ann that is used for market predictions. The link
is:
http://download.cnet.com/Cause-And-Effect-Correlator/3000-2066_4-10734312.html |
|
|
| Back to top |
|
|
|
|
|
All times are GMT
The time now is Fri Dec 11, 2009 4:17 pm
|
|