Nooby McNoob becomes a quant

Discussion in 'Journals' started by nooby_mcnoob, Mar 24, 2017.

  1. I think its desirable to have realistic expectations of SR going forward for a number of reasons (getting the right leverage, not paying too much for trading costs, not prematurely giving up on or meddling with a system whose performance doesn't meet your expectations.... to name a few). So you could still have a system which isn't overfitted (in the sense of being over optimised to the past) but on which the expected SR when trading would still be less than the backtested SR... which was what the original discussion was about.


    [Just when I'd just about given up on ET up comes an interesting thread with civil discussion... :D]
    #71     Mar 27, 2017
  2. Zzzz1


    Agree with most of your points. I generally count rebustness and stability as core constraints in the optimization function and I then optimize risk adjusted returns. Such optimization function prevents most always overfitting to past performance and market patterns. I employ other constraints such as correlation constraints (in order to not be overly correlated with certain assets or in relation to broad based markets benchmarks).

    Last edited: Mar 27, 2017
    #72     Mar 27, 2017
    Van_der_Voort_4 likes this.
  3. In this context, I think overfitting means to choose parameters that will help your back-testing succeed. Like "buy APPL in 199x". I don't know how you avoid this bias, however.
    #73     Mar 27, 2017
  4. algofy


    Yeah it's very tough with a stock that you know has skyrocketed not to curve fit the backtest because of knowledge of the past.
    #74     Mar 27, 2017
    Buy1Sell2 likes this.
  5. sle


    I think we are talking about different things. In vol space you see a lot of overfitting via introducing spurious parameters (sales desk "strategies" are notorious for that). The worst one I've seen was introduction of a 4-parameter Kalman filter into a vol selling strategy that would have had 4 significant draw downs over known historical data when done naively. That kind of over fitting is hard to avoid because you just don't have enough data across various market regimes and you end up accepting it.

    My approach has always been to have an apriori hypothesis on a statistical relationship rather then trying to go with the "find an edge and keep improving your Sharpe" way. YMMV.

    #75     Mar 27, 2017
  6. Zzzz1


    Hmm, isn't that exactly what I have been saying? In the end, however, you still need to optimize for risk adjusted returns, else why are you in the market in the first place. Hence the real challenge is not defining your goal seek but your constraints and boundary conditions.

    #76     Mar 27, 2017
  7. sle


    I am trying to say that in an ideal world, you want to approach strategy development from first principles. E.g. if you are adding a constraint, the reason for it should be extrenious to the strategy as well as your decision to keep/remove it. Otherwise it's a way to add bias to the strategy.

    #77     Mar 27, 2017
  8. Zzzz1


    Absolutely agree.

    #78     Mar 27, 2017
  9. Buy1Sell2


    This is why forward testing is really the best way to learn trading.
    #79     Mar 27, 2017
    digitalnomad and algofy like this.
  10. truetype


    As versus those people who prefer fragile, unstable optimizations?
    #80     Mar 27, 2017
    nooby_mcnoob likes this.