Targets on Machine Learning approach

Discussion in 'Automated Trading' started by HarrySquatter, Apr 24, 2018.

  1. Hello!
    I'm pplaying with different machine learning algorithms and stock data but I'm a bit confused. Can anyone tell me what is the target algorithm accuracy when it comes to predicting next movements (up/down) upon which one can build a strategy? Should it be 51% or 60, maybe sth different? I'm just trying to build my intuition when it comes recognizing the results that I'm getting from the models that I build. Any suggestions on this topic would be appreciated.
     
  2. fan27

    fan27

    I use ML in such away that each strategy candidate has to pass a filter of a win rate of 75% where the stop and profit targets are set at 2 x the Average True Range (14). The strategy candidates are tested on a small group of ETFs and all of those that pass the filter are then run on a larger group of ETFs. Those strategies that pass the filter on the second run are determined to have an edge. From there I run different order combinations to determine how to maximize the edge.
     
  3. truetype

    truetype

    Without knowing your strategy details, it's likely the results on various ETFs are highly correlated.
     
  4. fan27

    fan27

    Correct. Strategies performing well on one group of ETFs is likely to perform well on another group because most of the ETFs are based on stock performance which share similar characteristics. Making it a two step process though does separate training data from out of sample data and improves performance of the test run as over 1 million strategies might be tested on the small group of ETFs which I can keep loaded in memory. Of those strategies tested, there might only be a couple hundred or less that are then tested on a larger group of ETFs.
     
  5. I am going to taking machine learning next term for my MS in CScience since it seems to be latest buzzwords today in the job market. Can someone tell me what is the diff between ML/AI when it comes to trading v. the backtesting platforms in the market today such as Amibroker, Tradestation or Investor RT? Sounds kinda the same.
    Thanks.
     
  6. truetype

    truetype

    Essentially, If you can understand why your bot is buying or selling, it's not ML/AI.
     
    logis likes this.
  7. IOW, you treat it like some black box process where u "surrender" to the tendencies instead of saying what would performance be if I use Trigger 1 or 2,etc?..interesting...
     
  8. fan27

    fan27

    There are different flavors of ML/AI, mainly supervised and unsupervised learning. Unsupervised learning is typical of deep learning algos that determine patterns without need of user input. Where as in supervised learning, the user provides the features and the algo determines which feature values makes up a viable pattern. As a simple example, a supervised learning approach as it relates to stocks might have the following features:

    Todays Close
    Todays Close - 1
    Todays Close - 2
    Todays Close in relation to 200 ma

    So for each point in the stock time series data, the algo will determine the value of the features. For example:

    Todays Close is Up
    Todays Close - 1 is Down
    Todays Close - 2 is Down
    Todays Close < 200 ma

    The above strategy will be backtested, results recorded and then the algo will go to the next point in time and repeat.

    There are other implementations than what I listed above, but that is the basics of how I go about using machine learning.
     
  9. Thanks for the input. I hope that after my ML class, I can discern the differences bet ML learning v. backtesting thru a commercially available tool like Amibroker, Sierra. In the above example I can run a backtest like...( won't know till ML class ends where they diverge)
    Buy1 = Close > Close.1 'higher than yest
    Buy2 = Close.1 < Close.2 ' yest close lower than 2 days ago
    Buy3 = Close.2 < Close.3 ' close 2 days back lower than 3 days back
    Buy4 = Close < MA(200)
    Buy= Buy1 and buy2 and Buy3 and Buy4;
     
  10. fan27

    fan27

    Note that you had to manually code the rules for your strategy above. Machine learning can semi-automate the generation of those rules (supervised learning) or fully automate the generation of the rules as can be the case in unsupervised/deep learning.
     
    #10     Apr 24, 2018