NPP builds a Emini system

Discussion in 'Journals' started by no_pm_please, Dec 8, 2006.

  1. It's just basic probability. All I'm doing is modeling what happens if we bet 1% of a fixed amount of money on every parameter.
    I'm assuming the parameter for this model is independent. Any one outcome is as likely as another. I'm also assuming it is bounded from 10 - 99 (maybe a faulty assumption). With that information you run tests for all values from 10 through 99. Using profit factor as a proxy for payoff and number of trades for trials we can compute the expected outcome for each 1% bet as ((profit factor - 1) * number of trades)/number of tests (90). You do this for each test and then sum all the test outcomes to come up with a expected return for the entire test.

    I've included a simpler example in the gif where there are only 5 tests. Each test has a different profit factor and number of trades.
    The payoff (profit factor - 1) per 1% risked is in column I. The total
    profit per test is in column J. Add all the column J's and you get the expected return for each 1% risked. In this case the expected return is 56.676%.

    If there were no edge, the expected return would be negative or at least close to 0. If I make a change to the model I can re-run all the tests and see if the expected return increases or decreases. If it increases, then I've improved the overall model. I don't know if system traders do this test to see if what they've got has an edge, but without it, I think you're just curve fitting.
     
    #41     Dec 15, 2006
  2. Maverick1

    Maverick1

    Hi NPP,

    Don't mean to digress, but was curious to see what happened to your divergence method? Did you give up on it?

    Thx
    Maverick
     
    #42     Dec 15, 2006
  3. Nope, I've been using it all this year. Been having great success with it. I've only been trading it from the buy side since September since the market is trending so much. People gave me so much crap about not being able to backtest my methods that I decided to not post my current production stuff and instead thought I'd see if the systems traders would come out to play.
     
    #43     Dec 15, 2006
  4. bidask

    bidask

    let me see if i understand. you got 42.67% in 2004 by doing the following:

    1) get individual expected returns by varying the parameter from 10-99.
    2) sum the individual expected returns and you get 42.67%

    correct?

    then do the same for every other year. sum the expected return for each year and you get 35.71%.

    what do you do to get the 25.98% number?

     
    #44     Dec 16, 2006
  5. If you do the optimization for 10-99 based on all the data from 2003 - 2006 and sum the numbers you get 25.98%. That's why I said it was a poor way to come up with the overall parameter. Just even weighting of each parameter each year beats it by 10%.
     
    #45     Dec 16, 2006
  6. Attached is the ELD for TradeStation 8.1.


    Code:
    
    input:
        per(84);
    
    {entry/reverse}
    {enter on dip strategy}
    
    if adx(14) < 1.01*(adx(14)[1]) then begin
        if rsi(close,per) >= 50.5 and rsi(close,per)[1] < 50.5 then buy("NEP_LE") next bar on open;
        if rsi(close,per) <= 49.5 and rsi(close,per)[1] > 49.5 then sellshort("NEP_SE") next bar on open;
    end;
    
    {exit with profit}
    if marketposition = 1  and rsi(close,per) <= 52 and rsi(close,per)[1] > 52 then Sell("NEP_LX") next bar at open;
    if marketposition = -1 and rsi(close,per) >= 48 and rsi(close,per)[1] < 48 then buy to cover("NEP_SX") next bar at open;
    
    {exit with large profit}
    if marketposition = 1  and rsi(close,per) <= 58 and rsi(close,per)[1] > 58 then Sell("NEP_LLX") next bar at open;
    if marketposition = -1 and rsi(close,per) >= 42 and rsi(close,per)[1] < 42 then buy to cover("NEP_LSX") next bar at open;
    
    
     
    #46     Dec 16, 2006
  7. bidask

    bidask

    npp: the latest analysis seems to establish two things.

    1) using one parameter for all the years is not good

    2) the strategy has an edge.

    for #1, didn't you already determine this when you put together the first spreadsheet?

    for #2, i'm not sure if i agree with you that the analysis shows the strategy has an edge. it seems like you're saying that the strategy has an edge because it has positive expectancy. but a curve fit strategy will also have positive expectancy. how would your analysis look different if it was curve fit?
     
    #47     Dec 18, 2006
  8. bidask

    bidask

    here are the results if you take it back a few more years. the numbers in the earlier years are significantly poorer. i know that my data doesn't not match yours exactly, but these results seem to suggest curve fit.
     
    • rsi.gif
      File size:
      5 KB
      Views:
      194
    #48     Jan 13, 2007
  9. Yes, that's true. We've been in a lower volatility environment for years than from the 90's. It's also consistent with the use of a single variable for all periods. The single variable model is not very useful when we've seen a huge change in market volatility over the years. We all know as a market becomes more volatile you have to shorten up the time horizon.

    I've been working on this on and off for the past few weeks. I broke volatility up into 5 periods (as measured by vix). By applying a different number for each volatility period I was able to get some decent results going back to the mid 90's. To do the test I had to switch to a SP contract since the ES didn't exist way back then. ES contract returns would be approx. 1/5 of the posted results for 1 SP contract.
     
    • et96.gif
      File size:
      14.8 KB
      Views:
      217
    #49     Jan 28, 2007
  10. I'm going to check out other options such as overlapping multiple models that might have a even better payoff.
     
    #50     Jan 28, 2007