Adaptive Trading Strategies

Discussion in 'Strategy Building' started by jalsck, Sep 4, 2010.

  1. What do you mean uglier? You presented results that were below dismal and I presented results that were much superior. Now you come back to claim without evidence that my results were uglier than what you got.

    The fixed 5/30 system that trades on next open has an equity curve that is almost a straight line up.

    Please present results and not vague claims. Otherwise you can be misleading people here.
     
    #61     Nov 11, 2010
  2. jalsck

    jalsck

    There was no intention to mislead anyone or to offend. I was talking about my fixed 5/30 system not yours. I have attached a spreadsheet that lists on what days the 5/30 system is short and long. I compared this to the trade dates in the Dakota report and they were in agreement. At this point I don't know why we are getting different results.
     
    #62     Nov 11, 2010
  3. jalsck

    jalsck

    This spreadsheet calculates and graphs the equity curve based on my understanding of the system.
     
    #63     Nov 11, 2010
  4. rickty

    rickty

    jalsck,

    My results agree with you. The equity curve is attached.
    (Code is shown below).

    Richard


     
    #64     Nov 11, 2010
  5. rickty

    rickty

    ... the code ....
     
    #65     Nov 11, 2010
  6. jalsck

    jalsck

    Thanks for letting us know rickty.

    Nitro, the penny dropped today - I didn't really do what you suggested by feeding the S&P 500 ARM03 system the detrended data. I should have calculated the performance on the detrended data as well. I will need to detrend the data series in a spreadsheet to do this. Can you suggest a good method for detrending the data? I was thinking of taking the natural log and then subtracting a 6 year SMA or EMA.
     
    #66     Nov 11, 2010
  7. nitro

    nitro

    http://www.dsptutor.freeuk.com/dfilt1.htm

    You have to be careful not to pass future data to the filter. I say that because I have done it accidentally. Some people just take a simple low pass filter and use the difference of the actual data and the lowpass filter (bandpass) as the data to feed to the model. Timothy Masters in his book discusses the pitfalls with all approaches in his preprocessing techniques section:

    http://www.amazon.com/Neural-Novel-...=sr_1_2?ie=UTF8&s=books&qid=1289487547&sr=8-2
     
    #67     Nov 11, 2010
  8. I apologize for this but I forgot a filter active in my backtester which (to my amazement) turned the fixed 5/30 system to a highly profitable one! So my results without the filter for the fixed 5/30 system and trade at the next open are:

    Win rate: 36%
    Net profit: 6.15
    97 trades, 48 long, 49 short
    long wins 24, short wins 11

    Equity curve is more or less like yours.

    So what are the system backtest results after applying Dakota?
     
    #68     Nov 11, 2010
  9. jalsck

    jalsck

    Thanks for the info Nitro. I have the book and have read a couple of other books on digital filters. There is a filter I have coded up that I can use that has the same lag as an SMA but is much smoother.

    Bill thanks for explaining why we were getting different results. I will need to do a run using QQQQ data from Pinnacle to produce trading stats over the same period because I need 4 years of data prior to the first signal being produced. I'll do that now.

    Best Regards,

    James
     
    #69     Nov 13, 2010
  10. jalsck

    jalsck

    The following are the results for my implementation of the fixed 5/30 system. These figures differ slightly from Bills due to data or slight implementation differences. These are the numbers to compare against the adapted version.

    Trade Count: 97
    Winning Trades: 32 (33%)
    Total Equity: 10.98
    Win/Loss Point Ratio: 1.048

    The adapted version allows the shorter SMA period to vary from 2 to 8 and the longer SMA period to vary from 20 to 40.

    Trade Count: 103
    Winning Trades: 32 (31%)
    Total Equity: 10.98
    Win/Loss Point Ratio: 1.034

    The equity curve generated by the adaptive version was actually worse than the fixed version over the 4 years leading up to the point where the adaptive version on the supplied QQQQ data file started generating signals (from my initial run).

    To summarize, the fixed 5/30 version did slightly better than the adaptive version overall.

    Regards,

    James
     
    #70     Nov 14, 2010