neural systems

Discussion in 'Trading Software' started by bobpit, May 3, 2008.

  1. bobpit

    bobpit

    From what I understand, software that uses neural networks technology, it searches for patterns and creates a strategy to the user. The user will have no idea what the patterns are and cannot modify it. Is that it?
     
  2. HoCk

    HoCk

    do you think someone would be dumb enough to sell you something like that if it actually worked...of course not..the neural nets that do work are being used and kept secret... :)
     
  3. bobpit

    bobpit

    ok. So when a software advertises "neural nets technology". What is the benefit to the user?
     
  4. Baywolf

    Baywolf

    My neural nets, its better than yours, I can teach you, but I have to charge...
     
  5. Murray Ruggiero

    Murray Ruggiero ET Sponsor

    There three different classes of products we are talking about. The first are neural network based indicators which you can't backtest. This are worthless, even though they sell for 1000's.
    If they will not let you backtest it easy, then run away!!!.
    Next Trading Systems based on neural nets. What important is the out of sample results on data not used in training otherwise you will see super results on the training set and much worse results in out of sample data. This can be hard to judge if you don't know where training ended. If you don't know where out of sample data start and training ends you might be kidding yourself.

    The final class which have value are tools , like TradersStudio+NeuralStudio and things like NeuralShell Trader. This tools let you build your own neural network based trading systems. The key in building neural network systems that work is to start with a trading concept which is profitable and use the neural network to improve it.

    In terms of not understanding what the neural network is doing that is true but you can do analysis like sensitivity analysis and mapping of the hidden node output to try to learn how the network makes it decisions.

    I have written many articles on this topic for Futures Magazine. All of these articles are in my book Technology and Trading for the New Millenium

    http://www.tradersstudio.com/Produc...cmd/CatalogItemDetails/psmid/657/Default.aspx
     
  6. HoCk

    HoCk

    it's a fancy word to sell software :)
    and do you really think they need you to press the "on" button on a neural network based system which is finding it's own patterns and making it's own trade decisions...of course not.


    I'd have to look at the specifics of the system you are referring to in order to give an honest opinion though.
     
  7. You can try Trading System Lab, which is based not on neural networks but on genetic programming:

    http://www.tradingsystemlab.com/home.aspx

    I find it fascinating but I cannot afford it. It sells for 60K + 20 K yearly after the first two years.

    This is a white box. It finds patterns and outputs the code of complete trading systems.

    Another program that is a white box is APS:

    http://www.tradingpatterns.com/

    This one finds patterns using data mining. It is also a white box and outputs the codes of the patterns. Much cheaper than the first one but not as fancy.

    These are the only two white boxes in the market I was able to find.

    Do not purchase a neural net system unless it tells you exactly the code of the system so you can backtest it yourself.

    Bill
     
  8. Think of a neural net as a black box, that when trained (most are supervised) with given patterns, returns an approximate function for the black box. It tries to minimize the error by adjusting the internal coefficients of your approximate function each time you train it Once the error has been minimized, it settles upon a function that gives the least errors and determines a minimal error function approximator.

    Keep in mind that approximation is a fancy schmancy curve fitter for all intents and purposes. Give it similar data to the past and it returns similar outputs to the past as it uses the function you found by training. Give it completely different data (i.e. random) and it will do nothing useful since by it's nature it looks for previous patterns. It is also useful for classifying and distinguishing data patterns into groups based on how it is trained.
     
  9. Murray Ruggiero

    Murray Ruggiero ET Sponsor

    I have not tried either of these products. In terms of neural networks I have worked more with Kernal regression over the past few years because it deterministic , produces the same answers each time and works as well as many of the neural network algorithms. Kernal regression is included in NeuralStudio.
    I have also used Rough Sets to induce rules from data but you need to preprocess the data like a neural network.

    Here is my stand on the issue. I like domain expertise to be embeded in the model. The search should be guided. This can be done by the preprocessing or postprocessing of the results.

    My favorite example of why is based on one of my clients who had a high power PHD developing a Nasdaq trading system for him in late 1999. The system made him very rich until 2002. Then it fell apart and he lost about 1/2 of his profits in about 6 months. He hired me to analyze why. The system was based on a black box method. I found that the predicted output was correlated .80+ with today minus four days ago. which if you look at a nasdaq chart during this period was very powerful. He figured he had the holy grail because it also worked well in the bear market. It got killed in sideways markets. If he understood this model he would not of lost much money.

    Black box systems are not all bad. You just can not put them in a place were their failure will cause the system to fail. You need to put them where they are used to improve the system.

    Let's look at an example. You develop a system which buys and sells on a limit 1 tick below yesterdays low with other filters. The system is very profitable. I would use a black box model to predict todays high and low and some amount inside that range based on the distribution of errors of the model.

    This way if the model degrades if most likely will not be much worse than using yesterdays low.
     
    #10     May 13, 2008