Machine Learning in ProOrder ProRealTime

• 04/28/2020 at 9:17 AM #128557
  GraHal

  zilliq wrote:

  omeone who has understand can explain the different step

  If you check back on this Topic and find a post by JuanJ (the Heuristics original Author) then he gives an explanation.

  

  GraHal

  Participant

  Topics: 211

  Replies: 8760

  Been thanked: 2176 times
  04/28/2020 at 9:25 AM #128558
  zilliq

  Thanks Grahal,

  I read his explanations, but no offense, it’s not very clear / simple to understand …

   

  

  zilliq

  Participant

  Topics: 58

  Replies: 683

  Been thanked: 152 times
  04/28/2020 at 9:44 AM #128567
  GraHal

  I thought it was simple to understand, but not so simple to relate the explanation to the actual lines of code that perform the actions in the explanation … is that what you meant maybe?

  Although I did sit looking at the full code one night and by going back over and over it seeking out the bits of code that align with JuanJ explanation … then it did all start to gel together for me.

  I’m sure you could do that as you are loads better than me at coding.

  Maybe we need to establish we are talking about the same JuanJ explanation first? I’ll try and find it and post the link.

  

  GraHal

  Participant

  Topics: 211

  Replies: 8760

  Been thanked: 2176 times
  04/28/2020 at 9:49 AM #128568
  GraHal

  juanj wrote:

  Then to further elaborate on GraHal’s explanation: The algorithm also keeps track of the following:

  Yes here it is @Zilliq … click on the blue link to get to the full post with the explanation.

  

  GraHal

  Participant

  Topics: 211

  Replies: 8760

  Been thanked: 2176 times
  04/28/2020 at 9:57 AM #128569
  zilliq

  Thanks Grahal,

  I understand the code, i understand the explanations but I don’t understand why StratAvgB could be the Best Profit and WIncount B the best win rate because I know to code them and it’s more complicated.

  And why should we implement ValueX for example depends on these variables

  (WinCount < 5 or StratAvgA > StratAvgB) and Mode = 1 Then

    Periods = Periods + (Increment+2)

  Why if we implement the results will be better ?

  I will re-read all the posts to understand. Actually I add them to different strategy and the results are worst so I should have a mistake

  See U later

  

  zilliq

  Participant

  Topics: 58

  Replies: 683

  Been thanked: 152 times
  04/28/2020 at 10:08 AM #128574
  zilliq

  The algorithm also keeps track of the following:

  • Best average performance (Profit) measured over all the previous evaluation periods (StratAvgB)
  • Best win rate (Win %) measured over a all the previous evaluation periods (WinCountB)

  It then measures and saves the same metrics but only for the last evaluation period (StratAvgA and WinCountA) and compares them to the above.

  It then starts by systematically incrementing the target variables both up and down to see what appears to improve performance most

   

  I understand the code and explanations more and more but don’t understand why implement should improve the performance as Juanj says “To see what appears” is different from “To improve performance”

  Ok we implement but why would it improve performance ?

  If we implement and re-test for the previous trades and compare results from period A to period B or all barindex  I would understand but it is not

  See U

  

  zilliq

  Participant

  Topics: 58

  Replies: 683

  Been thanked: 152 times
  04/28/2020 at 10:09 AM #128576
  GraHal

  zilliq wrote:

  I will re-read all the posts to understand.

  Hey great!

  At the beginning of this Topic … JuanJ shared his Heuristics version hoping others might improve it … so maybe you are the man to do this  … as you are coming up with good ideas already! 🙂

  

  GraHal

  Participant

  Topics: 211

  Replies: 8760

  Been thanked: 2176 times
  04/28/2020 at 10:16 AM #128577
  GraHal

  zilliq wrote:

  “To see what appears” is different from “To improve performance”

  I take this as … try one value higher, does it improve profit – No, try one value lower, does it improve profit – Yes … Action … keep trying lower values until profit starts to decrease.

  zilliq wrote:

  Ok we implement but why would it improve performance ?

  Because the Heuristics are changing variable values on the fly / while the System is running in Forward Test.

  I probably misread what you are asking?

   

  

  GraHal

  Participant

  Topics: 211

  Replies: 8760

  Been thanked: 2176 times
  04/28/2020 at 10:17 AM #128578
  zilliq

  🙂 Thanks

  If someone could explain the interest to implementation ?

  I think it would be good to compare the situation/conditions on previous winning trades to reproduce them on actual trades but it will consume a lot of CPU and what compare…

  Still try to understand why to implement the Valuex (Juanj if you could explain 😉 ) Thanks

  

  zilliq

  Participant

  Topics: 58

  Replies: 683

  Been thanked: 152 times
  04/28/2020 at 10:44 AM #128584
  GraHal

  zilliq wrote:

  why to implement the Valuex (Juanj if you could explain

  JuanJ is a full time trader and seems hardly ever checks this Topic so you may be waiting a long time?
  Also you need to tag JuanJ as @JuanJ … then he may get an email due to being tagged and he then may visit the Topic?

  ValueX is being varied by the Heuristics and ValueX is substituted in a Strategy as a variable and so (hopefully) the Strategy performance improves.

  I am probably misunderstanding what you are asking, apologies?

   

  

  GraHal

  Participant

  Topics: 211

  Replies: 8760

  Been thanked: 2176 times
  04/28/2020 at 10:59 AM #128588
  zilliq

  No you’re right thanks Grahal

  I read and read and read the all posts and seems to be complicated for all 🙂

  I like your “(hopefully) the Strategy performance improves”

  I will continue to try to understand the correlation between past performance and the need to implement of the parameter (why change/implement the valuex if it works in the past ?)

  The concept of juanj is may be to change/implement the valuex if you have a losing trade to see if it will win this time with a different value ? But why would it be better, same proabibility to get worst, no ?

  I continue to search an explanation…

  See U

  1 user thanked author for this post.

  GraHal

  

  zilliq

  Participant

  Topics: 58

  Replies: 683

  Been thanked: 152 times
  04/28/2020 at 11:24 AM #128590
  Scooby

  zilliq wrote:

  The concept of juanj is may be to change/implement the valuex if you have a losing trade to see if it will win this time with a different value ? But why would it be better, same proabibility to get worst, no ?

  In my point of view on this algo, this can be better if the ML take care about enough trades.

  Obviously if you take only the 3 lasts trades, if there is a Trump who open his mouth during the last trade, it can completely change the algo.

  But, if you take 20/30/… trades, the algo will not take care about the last one.

  Obviously we cannot see the futur but the ML permit to manage the parameters with the last x days/week/month of the market then it’s probably better (maybe not on all the case?..)

  I’m currently trying this with some simple algo but it’s effectively not simple 😀

  

  Scooby

  Participant

  Topics: 9

  Replies: 52

  Been thanked: 26 times
  04/28/2020 at 12:44 PM #128611
  juanj

  Hi Everyone, apologies again for not being around too often. Just too many other things requiring my attention in order to regularly answer questions regarding code here. I have noticed a sudden influx of tagged posts in my inbox, there seems to have been many questions already but for now I will answer only the latest one from zilliq.

  Q: why change/implement the valuex if it works in the past
  A: I take it you are referring to the systematic incrementation of ValueX, so obviously with this being a heuristics algorithm I am using incrementation to constantly discover through trial and error the current best performing values. These newly discovered values then become the new benchmark for future discovery. What in my opinion makes this superior to the standard PRT backtest/optimization to find the historic best performing value is the opportunity to dynamically adapt the variable to new ever evolving market conditions. As we all know, what used to be the best in the past is not guaranteed to be the best in the future.

   

   

  

  juanj

  Participant

  Topics: 51

  Replies: 505

  Been thanked: 170 times
  04/28/2020 at 1:05 PM #128612
  zilliq

  Thanks @Juanj very nice to reply and thanks for your time and work 😉 Absolutely no offense for my replies :

  I understand the system implement or “deplement” if the system on the last trades is worst than all but changing the value can get better but worst too ?

  The only possibility to know is to test but to have a losing trade too and not sure (or with good probablity) to implement or deplement will ameliorate the probability to have a next winning trade

  Moreover generally the parameter where the best we obtain with a backtest and changing the valuex will result in choosing a worst combinaison of values

  And to conclude determining that a valuex of 12 the 4 april doesn’t mean it will be the best value for the 12 april for example as it depends on cycle as showing on the works of Goertzel notably

  What’s your point of vue ?

  Thanks

  

  zilliq

  Participant

  Topics: 58

  Replies: 683

  Been thanked: 152 times
  04/28/2020 at 1:18 PM #128618
  nonetheless

  Bard wrote:

  It would be greatly appreciated if you can get to the bottom of this and figure out what is happening!?

  Doubt if this gets to the bottom of anything but came across this comment of Vonasi’s in another thread. Might be part of your problem:

  “Unfortunately SET STOP PTRAILING does not use tick by tick data and so gives false results.”

  Have you tried with a hard coded trailing stop?

   

  1 user thanked author for this post.

  Bard

  

  nonetheless

  Participant

  Topics: 83

  Replies: 1237

  Been thanked: 978 times
• Author
  Posts