Neural Network For Daytrading

Discussion in 'Trading Software' started by Lobster, Dec 7, 2002.

  1. Don't you think it's time for the little guys to get their own little black boxes?

    I am thinking of a neural network that receives its data from Medved Quote Tracker (there is an interface, right?) and gives real time output while learning to trade according to user specifications.

    Is there anything like this?

    If not, why don't we program it? What features would you guys look for in such a program?
  2. Wealth-Lab can pull data from quotetracker, and there's a few wealthscripts on their website utilizing neural nets. Don't think there's a turn-key solution for you there, but it might get you close.
  3. Thanks, let me check it out!
  4. Wealth-Lab has it for free (well considering you bought the software)...

    I think Tradestation has a third party software...

    I also recall a software like that already too...

    Also, Neural Net is not a black box... because you are pre-determining the specifications. It's more of a gray box and most of the neural nets are more like a constantly adjusted curve-fitting optimization algorithm.

    One interesting thing you should consider is to have the performance adjustment and/or learning based on Monte Carlo results. This should make the system more viable.
  5. GreenTea


    I don't think Neural nets work that well. They're just a fancy way to curvefit to datasets...
  6. dottom


    Not if you know what you are doing. Following the proper model building discipline to achieve generalization, you will find it is hard to curve fit data.

    Most people who first try out NN's don't understand key model building concepts. They simply throw data and indicators at some software package and give up shortly thereafter.
  7. GreenTea


    Hi Dottom,

    I think you may be right, personally I tried out 2 NN software packages, Brain maker and the Neuroshell in the past but wasn't able to find any good models which could be used profitably during walk-forward testing. What are some key model building concepts used to achieve generalization? thanks.
  8. Neural Nets can be a viable thing but they could also be used to reveal a painful truth about your trading style when back testing. I have a friend who developed a criteria model based upon his own trading logic. He then ran it back over a few years of data and understood why his portfolio looked the way it did. It actually just about mirrored when he would have, or would not have, traded exactly.

    Rather than put together some new model trading, we tweaked his parameters for entry and exit to optimize his performance and trade timing. Accordingly he has become profitable using his own style of trading. We have now addressed what we call his "bread & butter" trading. To augment that, now he is working on the "extras" to sweeten the pot. :)

  9. Normalization of inputs.
    normalization of inputs
    nrmlztion of npts
    nrmzshn fnpts

    Keys to the kingdom right there.

    Seriously though. Make the number of hidden neurons small. Select inputs that aren't correlated. Play with TDNN (i.e. inputs are price, lag(1) of price.... lag(n) of price). And get the smoothest possible inputs, with as little lag as possible (and be prepared to pay for the privelige..)

  10. I haven't followed this area very closely but I have seen ads for a couple of NN programs that tout intraday models that they claim are viable. Is this just typical vendor bs?

    I'm no expert in this area, but it seems to me the NN is just doing some pattern matching. The problem is some patterns are useful, others are random. I wonder if it would be feasible to attempt to isolate the patterns that appear to have some logical basis and incorporate them into a separate rule-based system?
    #10     May 29, 2003