Log in

Current EA development thoughts and ideas dropbox thread

Started by JohnLast 2624 days ago Replies (9)

Before going to vacation for spinal implant i compared the ergodic, the pfe and the entropy as inputs: the results were quiet similar. Yes there is a room for optimization but the equity curves were the same.

What will matters to me is that an upgrade will have to include a money management module. 


 

    JohnLast 2556 days ago

    Here I drop a shot how ideally the performace of the Spinal Implant looks like.

     

    image
    От EA tests
    JohnLast 2556 days ago

    And here I add a shot on another mod that is under consideration. Here I do something untraditionnal.

    I use the percetron function in order to give a different wejght of the iVAR with different windows as parameters to estimate the Hurst exponent.

    As this is a common problem for the use with the fractal dimension indicators. In fact if you use a different number of bars you will get a different estimate for the fractal dimension.

    So here I combine several different windows together. Well we can argue how precise this would be theoretically. But still this may be working.

    On the example here I optimized for one week of data and applaying the results for several months I got this.

     

    Here I do what I call reverse back testing. In fact most of the people make their tests using the normal time that goes from the past towards the future.

    But why to do so. We can choose a period and make reverse back testing towards the past not towards the future.

     

    For example I want to optimize using a present data and see if that works for the past instead using a very past set of data and see if that works for the current data.

    I hope you see my point.

    I was a little bit surprised to see such an equity curve for reverse back testing. Here the EA is very, very selectivve in the trades. 

     

    image
    От EA tests
    JohnLast 2548 days ago

    As for the EA tests I mentionned to Jaguer is that I really think that here the approach may be somewhat different than into the other forums and social networks.

    I mean that in the public forums there is not a serious evaluation and testing of the EA (with some important exceptions).

    Nevertheless some excellent code and EA have been created and shared. 

    So if we consider a cycle:

    trading idea - developpment - evaluation - forward testing - actual trading - monitoring - corrective action

    On most of the forums the developments stops somewhere, for example at the monitoring phase. 

    So there are many, many dead projects. On the other hand there are many EAs that do appear.

    If you want to make some serious work (I just can't evaluate and test everything). As it is a hobby for me and not a job. 

    Sooner or later most of the guys admit that. The work involved becomes quickly overhelming. 

    So if I do not analyse everything thorougly it is because I just can't. 

     

     

      JohnLast 2548 days ago

      I just want to add and maybe that merits a blog post or even a special page. That the methods developped can be implemented into other EA codes. 

      For example the percetron functions can be integrated into other excellent EA, and probably boost their performance. 

      Those methods can't (I think so) make a loosing EA into a profitable EA, but if correctly applied can make some extra pips over the original.

      For example the ASCtrend expert with kernel function outperformed out of sample the normal ASCtrend EA in all possible risk settings (is sample).

       

    JohnLast 2545 days ago

    Here I would add a link to the blog post about perceptron indicators.

    It is important because this site was designed mainly as project management tool. The system Elgg is used primarlily as intra-entreprise platform, and is used even by NASA.

    So it is important to keep track on all the ideas and thought related with our projects. 

     

    JohnLast 2543 days ago

    I was playing with one EA system from the mql database. I used the chaos kernels, and it really boosted its performance. Quite amazing. I will post that tomorrow.

     

    JohnLast 2542 days ago

    I like the performance of the Spinal Implant with the double chaos kernel.

     

    image
    От EA tests

     

    We can compare it with the other EA that uses a kernel for the iVAR too. This one enters into the market only when it is really convinced that there is a good entry.

     

    image
    От EA tests

     

    JohnLast 2520 days ago

    Here I would drop another idea.

    Recently in a blog post I asked a question about high probability trend following system using a PNN approach.

    The questions were formulated as follows:

    What about a trend following system with high number of profitable trades! Is that possible?

    No if we use the common waves to do so moving averages and common indicators. 

    Maybe, if we use a different approach?!

    The idea was to use the Entropy math with higher numbers of bars into the calculation of the indicator. The approach is really interesting because the concepts behind the indicator EntropyMath are really compealing. 

    Anyway the idea to use a neural system in order to predict an indicator using a lot of bars in the calculation is not new. I remember the Proteus 5 system that was using long period MACD as input.

     

    The logically I made some tests with the PNN. I used the PNN using the bud and sell with the PNN with the same parameters but only for short.

    The results are interesting:

    The PNN with sell short only outperform the PNN with buy and sell.

    PNN sell short:

    Total profit of 687

    Profit factor 2.88

    Winners 57%

    PNN Buy & Sell

    Total profit of 426

    Profit factor 1.45

    Winners 46.5%

     

    So the idea may be to use two PNNs.

    One would serve as identification of the current trend using higher frames (or the current frame with different parameters):

    The other will open trades only in the directions of the  trend identifications PNN.

     

    image
    От EA tests

     

    image
    От EA tests

     

     Hower For the spinal implant there is no difference during the first tests. Even more the performance deteriorated in july, but I think O have predicted that because of the summer range.

    Please compare the beginning of july 2012 with the end to see that it had more ranged characteristics (if you look at it and separate from the overall perspective for the EURUSD). I talk mainly for EURUSD because all the tests were done on this pair, and that is because it had the most predictable characteristics according to the tests I performed.

     

     

      JohnLast 2520 days ago

      The idea is not paricularly original after all. However it would limit the risks that originate from the long term trending system.

      A long term trending system would need bigger stops than a short term system.

      Anyway according to the classical technical analysis school that is the best thing a technician may do: the triple screen of Elder is a salient example.