Machine learning in trading: theory, models, practice and algo-trading - page 2283

New comment
 
mytarmailS:

I don't consume anything here, I just communicate on the forum, if you ban me it will only make me feel better... so...

And these 500 functions that you rewrote, they're dead in the sense that they're like a ruler with which you can measure something, and this "something" is just a third-party package which you shut off access to

You consume. And on this site, and in R and other places.

But at the same time and still demanding "give, and if you do not give - dead, rzheka, denied access.

Try to create and give. Feel some respect for other people's work based on your own.

 
Renat Fatkhullin:

Consume. Both on this site and in R and other places.

But at the same time you also demand "give, and if you do not give - dead, rjaka, closed access.

Try to create and give. Feel some respect for other people's work based on your own.

I guess you are right.

 
Rorschach:

Would it make sense to make native support for ZeroMQ?

There is no practical sense.

There are enough methods of communication and data transfer:

  • Pipes
  • files
  • webrequests
  • Raw sockets including TLS
  • DLL

We have a different specialization - read massive/diverse data and do most of the work inside the tester. Here, any going outside the boundaries kills everything.

At the moment, for us, integration with Python is a piecemeal solution to cover the ML developer ecosystem.

We are constantly upgrading MQL5:

  • got rid of the 32-bit legacy
  • replaced the profiler and debugger, we'll make it even better.
  • add fastest possible matrix/vector operations, so that you can perform heavy calculations without libraries
  • Launch MQL5 modules/packages based on C/C++ libraries.
  • Add more native integrations, like we did with OpenCL, DirectX, SQLite.


Creating an ML model on the side is not enough. It is necessary to load this model natively, safely and securely into the robot, so it can be

  • run it in the tester
  • run it in the terminal
  • sell it in the market without losing intellectual property.

That's why we are gradually increasing the capabilities of MQL5. We are planning to use WinML + ONNX to load models and the native execution within the platform without any third-party binding.

 
Renat Fatkhullin:

There is no practical point.

Methods of communication and data transfer are sufficient:

  • Pipes
  • files
  • webrequests
  • Raw sockets including TLS
  • DLL

We have a different specialization - read massive/diverse data and do most of the work inside the tester. Here, any going outside the boundaries kills everything.

At the moment, for us, integration with Python is a piecemeal solution to cover the ML developer ecosystem.

We are constantly upgrading MQL5:

  • got rid of the 32-bit legacy
  • replaced the profiler and debugger, we'll make it even better.
  • add fastest possible matrix/vector operations, so that you can perform heavy calculations without libraries
  • Launch MQL5 modules/packages based on C/C++ libraries.
  • Add more native integrations, like we did with OpenCL, DirectX, SQLite.


Creating an ML model on the side is not enough. It is necessary to load this model natively, safely and securely into the robot, so it can be

  • run it in the tester
  • run it in the terminal
  • sell it in the market without losing intellectual property.

That's why we are gradually increasing the capabilities of MQL5. We are planning to use WinML + ONNX to load the models and the native execution within the platform without any third-party binding.

Thanks for the detailed answer

 
Renat Fatkhullin:
Can you share the information:
1) Do you use the python library of MT5?
2) Do you use it outside or inside MT5?
3) What functions does the library lack? Access to indicators?

We are preparing an upgrade of MQL5 adding fast matrix operations. This will allow to perform massive calculations normally.

We will also develop connectors to analytical packages and implement the standard WinML integration.


It would be interesting to have access to the data in the order book.

 
Renat Fatkhullin:
Can you share the information:
1) Do you use the python library of MT5?
2) Do you use it outside or inside MT5?
3) What functions does the library lack? Access to indicators?

We are preparing an upgrade of MQL5 adding fast matrix operations. This will allow to perform massive calculations normally.

We will also develop connectors to analytical packages and implement the standard WinML integration.

Please, make the OHLC synchronization mode correct, so at least the standard indicators do not have bugs when requesting data from the upper TFs.

The main reason for the python to receive data from indicators, because training on all ticks is suicidal.

What is more annoying is slow speed of reading/writing files (csv/txt) in MT5.

 
Renat Fatkhullin:
Can you share the information?
1) Do you use the python library of MT5?
2) Do you use it outside or inside MT5?
3) What functions does the library lack? Access to indicators?

We are preparing an upgrade of MQL5 adding fast matrix operations. This will allow to perform massive calculations normally.

Further we will develop connectors to analytical packages and implement the standard WinML integration.
1) Yes
2) More outside
3) I had enough of everything...
 
Renat Fatkhullin:
It is not enough to create an ML model on the side. You have to upload this model natively, safely and securely into a robot to be able to use it:
  • run in the tester
  • run in the terminal
  • sell it in the marketplace without losing intellectual property.

The ML theme is getting away from you.
I use MQL5 to collect data and then to prepare the current data for neural network polling. Everything else is in python.
MQL in this chain only due to old inertia, because I started with it, otherwise all these issues are solved in python. Of course MQL are speed and clarity, but at the same time:
- crutches to get data from crypto exchanges
- inability to directly interact with api of crypto exchanges to trade
- inability to place in the Market your EA without opening the code (no automatic validation is possible if there is webrequest)
- total ignorance and inability to use MQL terminal (everybody use a browser).

 
Aleksey Vyazmikin:

Also annoying is the slow speed of reading/writing files (csv/txt) in MT5.


I try to store my data in binary form. But for external data I can't do without CSV.

 
Renat Fatkhullin:

You missed the moment when three years ago trading went to the masses. And it's not just crypto.
All of the plans you described are cool technically, but they are geeky stuff, they won't save you. If you want to jump into the last train, you need to urgently make a web version of the tradingview level, with all the functionality of the mql5 terminal.
Take this project as a basis and develop this direction, otherwise the train will pass, leaving you with one and a half geeks.

1...227622772278227922802281228222832284228522862287228822892290...3620
New comment