A proposal for testing black-box systems

Discussion in 'Automated Trading' started by botpro, Jan 6, 2016.

  1. Alex27

    Alex27

    #11     Feb 19, 2016
  2. Sergio77

    Sergio77

    Most black boxes are over-fitted systems.
     
    #12     Feb 21, 2016
  3. botpro

    botpro

    Sorry, but IMO that's a silly statement, because the system can be judged only after feeding it with new (ie. previously unknown) data.
    And if the system accepts only single datapoints then you can be sure that it works correct, as it has no chance to know the data in advance.
    It has to work on single datapoints, ie. bar by bar (or tick by tick) and give its trading decisions immediately.
    Then you have no reason for not to believe its result.
     
    #13     Feb 21, 2016
  4. gkishot

    gkishot

    No black box system will work ( or is expected to work ) on any random set of data.
     
    #14     Feb 21, 2016
  5. botpro

    botpro

    It is not wild random data one uses, it's either real market data or mathematically modelled data that is similar to real market data.
    Code:
    Random data would be this for example:               100, 60,  10, 90, 140, 190, 275,  15,  90, 150
    Normal stock data would look like this for example:  100, 102, 99, 96,  98, 101, 104, 104, 102, 103
    
    You just should analyse how much most of the stocks vary in their price daily, it is not much, for example just up to 2.5% up or down on a daily basis.
    One can compute that even exactly: it's called the "historical volatility", ie. the statistical standard deviation from the mean price.
    Currencies and Indices have an annual volatility of about 10% to 20%, stocks have about 20% to 40%. Of course there are also some outliers.
    Usually stocks of big companies are less volatile than stocks of startups and small cap companies.
    The higher the volatillity the more riskier it is, but higher risk can also mean higher reward if you can correctly predict the right direction.
     
    Last edited: Feb 21, 2016
    #15     Feb 21, 2016
  6. gkishot

    gkishot

    Are you saying that you expect it to work equally excellent on any random stock? Whether it's a winner or a loser?
     
    #16     Feb 21, 2016
  7. gkishot

    gkishot

    The difference is only in standard deviation.
     
    #17     Feb 21, 2016
  8. botpro

    botpro

    And what do you mean by that statement?
     
    #18     Feb 21, 2016
  9. botpro

    botpro

    Of course not. I don't mean random data, I rather mean realistic data paths.
    Read my updated posting again.
    How can you believe I would mean random data? That is nonsense. Why should anybody mean random data in such a context? This is far from and beyond reality.
     
    #19     Feb 21, 2016
  10. botpro

    botpro

    FYI: stock prices consist of 2 parts: the mean price plus a random component. The random component is the part from the annual volatility, for example just 1.5% up or down per day.
     
    Last edited: Feb 21, 2016
    #20     Feb 21, 2016