Fit and Overfit

Discussion in 'Trading' started by man, Jan 24, 2003.

  1. man


    In trend following communities there are two different opinions whether to trade each market in exactly the same way or allow more or less substantial variations. Usually people with background in any kind of statistic or playing games prefer to to everything in the same way in order boost validity and therefore stabilize expectation value. People with strong trading background claim that there are substantial differences between markets due to many reason. A corn contract differs from the t-bond, they claim, therefore it is valid to treat them differently. I personally prefer the first way of thinking but see the arguments of the second.

    I would like to know how many of you found day trading strategies that work on very many different stocks in exactly the same way. My current feeling is that you find rather much diversification within the behavior of let's say the top liquid fifty US names. No big surprise, but I am always very concerned about overfitting ...

  2. DT-waw


    I personally would like to talk about how strategy deals with market's non-linearity, instead of the question "is strategy over-fitted to history?"

    In day- or swing trading U.S. and European stocks or futures, b/c the session lengths are different (6.5 hours vs 11 hours) you have to use different parameters, rules for each market.
  3. man


    What do you mean by market non-linearity? the market as such is non-linear per se for my understanding. if you mean that markets change over time I agree and would propose strategy embedded relearning, the more often the better. (One fo the reasons why you need to program on your own I guess).

    I did not mean things like different lengths in trading sessions. It is obvious that you must adjust for a longer session or a break in the session as you find it in asian stocks.

    What I mean more specific is that certain stocks look like having a life on their own. You can find a strategy for such a stock, but it won't be profitable at the next. I do not like that too much since it contains an element of overfit. And I am not sure about the rational behind. Its clear that you have a more competitive day trader community in QLGC than you have in stock XYZ, but does that justify different trading strategies?

    Actually I can give the answer on my own: yes, if it lasts after the test was done. That's in essence the only thing that counts. My question is: how many found systematic approaches that work on many stocks in the same way?