Will this continue to work?

Discussion in 'Strategy Building' started by Bowgett, Jul 14, 2005.

  1. Linear automated or mechanical systems can work great for periods of time like 1 to 3 years---------then they usually hit a breakdown. Non-linear systems can last indefinitely as many institutions are trading methods that have not really changed in 30 to 40 years---------they have just become more and more efficient over time as the technology has changed for trading {lower trading costs per transaction and better entry/exit efficiency due to the better platforms and trading products}.
    #11     Jul 14, 2005
  2. Suggest you read Stridsman's book:

    Trading Systems That Work

    before you trade this or any other system with real money. Did you perform walkforward testing? I've produced dozens of IWM/R2K systems that look much better than one you show but would not trade them as I know they are the result of curve fitting.

    Also, there are many threads in ET on back testing and reading them is also a good idea.

    And IMO, a very important but little discussed topic is when to stop trading a system that may have worked well at one time but is showing signs of poor current performance. Do you have a monitoring plan to stop trading it if its performance goes south?

    #12     Jul 14, 2005
  3. Bowgett


    I perform walkforward testing with real trading. Because I trade more than 4 strategies at the same time I can afford loosing one or two. But this kind of testing is expensive.

    I have better ones too but drop them from my radar too because they look unrealistic. How do you know it is curve fitting? Right now I am just looking at it and saying to myself "Nah, it is too good to be true".

    I did the search but didn't find anything in particular.

    I am giving system one-two months to get to breakeven if it starts losing money and drop it if it cann't. I monitor its performance relative to its expected performance and if it deviates too much I reconsider future trading with it.
    #13     Jul 14, 2005
  4. Bowgett


    I suspect so too. Do you have any pointers on some studies to back that up? I think number of parameters can increase as you increase number of trades. If your system generated 30 trades then 3 parameters is too much but if you verified it on 500 trades 3 might be ok.

    I stopped using NN (Neural Networks) long time ago :D
    #14     Jul 14, 2005

  5. Are you saying it is cheaper to test a system by trading it as opposed to performing walkforward testing????????????

    I sort the curve fitted systems out by walkforward testing.

    There is a lot of good info on backtestin in ET but it takes more than a few minutes of skimming to find the good stuff

    #15     Jul 14, 2005
  6. Divide your historical data into 3 sets, one to use to optimize your parameters/rules, once you have done your optimization use the second set to validate your results and if that looks good then run it against the last data set. Do not optimize/tweak on anything but the first data set.
    #16     Jul 14, 2005
  7. Bowgett


    No, I am not saying this. This particular system generates a lot of trades. During backtesting you assume that you can enter at any time at any size and you can exit at any time at any size but in real life this is not how it works. I don't get fills almost every day when I get signals. Try to simulate this. I can add slippage and commissions but realtime trading is soo much different.
    #17     Jul 14, 2005
  8. Bowgett


    I don’t believe this works. Just imagine I used one third of data and found strategy that I posted then I used another third and I verified that it still works and then I used last third and it still works but this tells me nothing.
    #18     Jul 14, 2005
  9. Bowgett


    I was thinking about grouping similar strategies and assiging some kind of score to the group. If group of similar strategies performs well that means it is likely that one of strategies from this group will perform well in the future.
    #19     Jul 14, 2005
  10. It tells you much more than taking the data and optimizing across all of it and then declaring those results as meaningful (which is what I understood you did).

    If you curve fit to one set of data and then backtest against a second independent set and still get similar results then you can be more confident that it was not over-optimized to the first data set.
    #20     Jul 14, 2005