Commenti | Neuro Start | Scarica gratuitamente il Robot di trading (Expert Advisor) per MetaTrader 5

Itharius 2021.05.15 15:25 #1

How to optimise it?

Dmytryi Voitukhov 2021.05.15 15:47 #2

Itharius:
How to optimise it?

For optimization, just turn it on in a circle.

For the first few eras, profits fluctuate. Since with each next pass, the entry point may not coincide with the previous training and this will be a new picture for the program. I teach for short periods - 2-3 weeks. Then the next ones. And over 15-20 epochs, the testing schedule becomes strongly growing.

Dmytryi Voitukhov 2021.05.15 16:24 #3

Dmytryi Voitukhov:
For optimization, just turn it on in a circle.

For the first few eras, profits fluctuate. Since with each next pass, the entry point may not coincide with the previous training and this will be a new picture for the program. I teach for short periods - 2-3 weeks. Then the next ones. And over 15-20 epochs, the testing schedule becomes strongly growing.

At the first launch, the script creates the necessary files by itself.

Set TypeOfWork to Learn.

Data files are larger than 50 mb and this is a lot of the limit here. Publish the download link or send me an email.

Itharius 2021.05.15 19:21 #4

Dmytryi Voitukhov:

At the first launch, the script creates the necessary files by itself.

Set TypeOfWork to Learn.

Data files are larger than 50 mb and this is a lot of the limit here. Publish the download link or send me an email.

well...it takes time...a lot of time:)

Itharius 2021.05.15 22:39 #5

Just one question....do u use this in live trading?

Or this is a new idea...?

Dmytryi Voitukhov 2021.05.16 00:29 #6

Itharius:

Just one question....do u use this in live trading?

Or this is a new idea...?

that's what good training is for

Itharius 2021.05.16 17:15 #7

Ran training process over 400 epochs for a month of April (04.04.2021 to 02.05.2021) which is 4 weeks.

|Got a very good results of around 100% profit for that month...

Ran for one week in may (02.05.2021 - 09.05.2021) in TypeOfWork - work and noticed a big loss.

Does this suppose to happen?

Dmytryi Voitukhov 2021.05.16 20:22 #8

Itharius:

Ran training process over 400 epochs for a month of April (04.04.2021 to 02.05.2021) which is 4 weeks.

|Got a very good results of around 100% profit for that month...

Ran for one week in may (02.05.2021 - 09.05.2021) in TypeOfWork - work and noticed a big loss.

Does this suppose to happen?

I don't know how many months of training it takes to work confidently. It is optimal for the solution to include a period unique enough for the answer to be correct and general enough to recognize similar figures in the future. If the Depth of Analysis is set to 1 minute, a graph with a length of 1 * 50 = 50 minutes is placed in the solution. If the Depth of Analysis is set to 2 minutes, a graph with a length of 2 * 50 = 1.5 hours is placed in the solution. If the Depth of Analysis is set to 5 minutes, a graph with a length of 5 * 50 = 4.1 hours is placed in the solution. The vertical scaling of the input data is 1:20. Which makes the input fairly general and normalized.

Dmytryi Voitukhov 2021.05.17 06:54 #9

If the data depth is too short, the new data spoils the previous ones. If it's too long, there will never be a similar one. I will add - the use of more than 1 (in agents) processor during optimization does not make sense. One instance will constantly overwrite the results of the second.

Itharius 2021.05.17 11:58 #10

Dmytryi Voitukhov:
If the data depth is too short, the new data spoils the previous ones. If it's too long, there will never be a similar one. I will add - the use of more than 1 (in agents) processor during optimization does not make sense. One instance will constantly overwrite the results of the second.

This what I thought...ran it with disabled agents. Only one working, on 30 min depth of data.

Dmytryi Voitukhov 2021.05.17 21:07 #11

Itharius:
This what I thought...ran it with disabled agents. Only one working, on 30 min depth of data.

30 minutes = 30 * 50/60 = 25 hours - a very unique picture. It is impossible to find the same combination of figures in the future.

Itharius 2021.05.31 22:55 #12

What StepProf and Pass parameters are for?How do u getting on with this EA?

Dmytryi Voitukhov 2021.06.02 00:38 #13

Itharius:
What StepProf and Pass parameters are for?How do u getting on with this EA?

StepProf - from other strategy. Work like this - if many orders exist (+ & -) but summary greater then last balance with StepProf - close all.

Pass - for separate sale. Sets the duration of the work.

My actual state in pics after 3 epoch.

But with a smaller timeframe, you need to set TP and StartTrail less, respectively.

To 50 for example.

Dmytryi Voitukhov 2021.06.03 15:21 #14

I suppose it is better to remove the Day of the month, Day of the week, Hour of the day from the input data. Your opinion?

Itharius 2021.06.04 00:15 #15

Dmytryi Voitukhov:

StepProf - from other strategy. Work like this - if many orders exist (+ & -) but summary greater then last balance with StepProf - close all.

Pass - for separate sale. Sets the duration of the work.

My actual state in pics after 3 epoch.

But with a smaller timeframe, you need to set TP and StartTrail less, respectively.

To 50 for example.

Ur input data section looks different then mine. Can't see tresholdout data. Ur tresholdh1 is set at 0.2..different than setfile with EA.... dunno tried to learn EA over a few weeks..got good results on bactest, but failed to make profit in live trading...maybe needs more time..

Dmytryi Voitukhov 2021.06.04 11:51 #16

Itharius:
Ur input data section looks different then mine. Can't see tresholdout data. Ur tresholdh1 is set at 0.2..different than setfile with EA.... dunno tried to learn EA over a few weeks..got good results on bactest, but failed to make profit in live trading...maybe needs more time..

tresholdh1 is fixed at 0.2 and therefore hidden from changes. while testing can be stopped. I should think.

Thanks for participating.

Itharius 2021.06.04 23:41 #17

Dmytryi Voitukhov:
tresholdh1 is fixed at 0.2 and therefore hidden from changes. while testing can be stopped. I should think.

Thanks for participating.

EA delivered to the market has different settings....I have set the tresholdh1 to 0.2...ran the bot over a few epochs...and start to like it😀

..does the bot learns while working, or need to relearnef every few weeks....?

Dmytryi Voitukhov 2021.06.05 11:45 #18

Itharius:
EA delivered to the market has different settings....I have set the tresholdh1 to 0.2...ran the bot over a few epochs...and start to like it😀

..does the bot learns while working, or need to relearnef every few weeks....?

In order for the advisor to train in the process, it must be launched somewhere in the training mode on a copy of the terminal. It is important that FILE_COMMON is the same for both terminals. This can be seen in the log in the line with "Find data ...". Periodic loading of new data in Work mode is carried out in the attached version. In any case, I strongly recommend not to work on a real account yet !!! You can enable or disable the processing of the day of the month, day of the week, hour of the day in neuro input.

File:

Neuro.ex5 76 kb

Itharius 2021.06.05 14:24 #19

Dmytryi Voitukhov:
You can enable or disable the processing of the day of the month, day of the week, hour of the day in neuro input.

What is that for...? Is it period of loading up a new data for EA?

Dmytryi Voitukhov 2021.06.05 15:13 #20

Itharius:
What is that for...? Is it period of loading up a new data for EA?

When training in real time, the advisor saves the training results every 1 hour. If the instance running in combat mode is not reloaded (for example, change the timeframe), it does not load new data. But now it has been done - the data will be refreshed every 3 hours. Input data for training: Day of the month - if enabled - there must be training on all days of the month. So it is for the Day of the week. And the hour of the day. Should I turn off combat mode so you don't lose money? :)

File:

Neuro.ex5 76 kb