Adding more parameters to your model vs. curve fitting?

Discussion in 'Strategy Development' started by mizhael, Apr 21, 2010.

  1. Hi all,

    Please shed some lights on me about adding parameters to your model while avoiding curve fitting...

    Here is my question: I have a model that runs reasonably well, now I found that if I throw in one more parameter, I can improve the Sharpe Ratio by 4%. It's a stop-loss. Do you think I should include this parameter in my model?

    Any thoughts?

    Thanks a lot!
     
  2. Visitors

    Visitors

    run it and see how it performs on long term large data backtest
     
  3. So you're currently running it with no stop loss in place?
     
  4. How many input parameters you add is a matter of your personal tastes. You must define your own beliefs about the degrees of freedom that your strategy will use. I say this because there are fierce opinions on this issue as to what is one input parameter too many. Meaning you produce only curve fitted responses as parameters are added.

    I have traded profitable strategies with up to 6 input parameters in use. I normally use only 5; which I have been told is high for degrees of freedom. So to compensate for extra parameters I thoroughly back and forward test my strategy with filters that screens out the settings that will perform poorly.
     
  5. 4% is close to the statistical error. Better check the change in profit factor.

    A stop-loss that does not change the timing order of the entry signals is not a curve-fit, it is just a risk management method. If the stop-loss method changes the timing order of entry signals then it is not a matter of curve fitting, you have a new model all together. Curve-fititng will come into play if you optimize the value of the parameters to minimize (maximize) some objective function.
     
  6. Mizhael, seriously, give it up. Find work in some other field.

    Don't worry scrote, plenty of 'tards leading really kick-ass lives. My first wife was 'tarded. Now she's a pilot.
     
  7. Is it a date-mining if you "screens out the settings that will perform poorly."?

    How would you do it without data-mining?
     
  8. When I add the stop-loss of course it has a threshold parameter, and I have to optimize a bit with respect to that threshold parameter, right?

    So after adding the stop-loss logic and even after optimization, the Sharpe ratio is only 4% higher. Shall I conclude that it is not worthwhile adding this stop-loss mechanism and parameter?
     
  9. For this particular one, I don't have a stop-loss... at this moment. That's why I am thinking of adding one...
     
  10. In a nutshell... (1) PRICE ONLY. (2) If "your comfort" requires that you MUST use some technical indicator(s)... USE ONE! (Doesn't matter which one.)

    Y'all should send me (or my favorite charity... The Max Fund)... $5K for wising you up!
     
    #10     Apr 22, 2010