Machine Designed Trading Strategies

Discussion in 'Strategy Development' started by JWT, Apr 27, 2010.

  1. JWT


    There are several proposals on the Internet from different companies. for 30k euro for $60k +$20k every year after 2 years period. for "only" $2k
    and so on....

    All of these programs create a portfolio of automatic systems, which in theory are efficient in the future.
    Can't say that I am ready to pay even for the cheapest just to try it.
    I wonder if anyone have any positive or negative experience using this approach.

  2. MGJ


    If you're a programmer, or have a programmer who works for you, it's not all that difficult to create your own program that "designs" trading strategies. If you don't want to invest X dollars of money, buying someone else's program, you have another option: invest Y hours of time, writing your own.

    When you've written and debugged your program, go ahead and let it run a LOOOONG time ... CPU cycles are cheap compared to losing trades. Generate a large number of candidate systems and apply some pretty stringent filtering rules to winnow them down to a small pool of finalists, for acceptance / rejection by the human trader.

    I found that about 0.005% of the machine designed systems were decent enough for me to seriously consider them for live trading, i.e., a "hit rate" of about one in 20,000. Your results might vary of course.

    I also found that it was far easier to machine-design trading systems that trade long-only or short-only, than to machine-design trading systems that trade both long and short. A quick read of David Harris's books seems to indicate his APS software has the same limitations, so it only spits out long-only or short-only systems. Write your own program, tinker with it for a few weeks, and see whether your software shares this weakness or not.
  3. This is called curve-fitting.
  4. One the first programs in this area, if not the first, is APS from this company. I have this program and it does an excellent job in finding systems based on micro-patterns but...

    you really have to know what you are doing with these programs. The most important thing is establishing a sound procedure for testing the significance of the systems they produce. Most systems they output do not have an edge unless they are backed by substantial out_of_sample testing. Also systems that incorporate indicators result mostly in curve fits.

    Unless you know how to use these programs the correct way - and that it is not trivial - I do not recommend spending your money getting them. Such tools can do more harm than good in the hands of people who do not have extensive system testing and evaluation experience.
  5. JWT


    I'll review the APS.At least they give a demo.

    P.S. Think I should note.I develop and trade systems since 2000.