Canvas is cool! - page 102

New comment
 
Nikolai Semko #:
Yes, very interesting and useful. Thank you. I won't be able to go into more detail until a month from now. So far, time is in trouble.
 
fxsaber #:

https://claude.ai/share/a99469f8-79d9-4f3f-8aaa-55f951560093
It's a mess. My brains will soon be too lazy to work.

 
Nikolai Semko #:

https://claude.ai/share/a99469f8-79d9-4f3f-8aaa-55f951560093
It's a mess. My brains will soon be too lazy to work.

Thank you. I read it carefully. As one immersed in the topic, somewhat disappointed.


I noticed this statement. I don't know if it's true or not.

// Check the sign without if:
int sign = 1 - ((delta >> 31) << 1);  // faster than if(delta < 0)


My final advice

Implement in this order:

  1. LUT + Fast Path - 2-3 hours of work, +15% immediately
  2. RLE for zeros - 1 hour, +10% compression
  3. Bit flags - 2 hours, +5% all together
  4. Adaptive block - if you have time left, +3-5%

That would give you ~65 million ticks/sec at 4.8x compression - a great balance.

I can't even come close to imagining such a high debugging speed for myself in this area.


I guess on this topic it only makes sense to ask for links to advanced work done by people. Thanks again.

 
fxsaber #:

Thank you. Read it carefully. As one immersed in the topic, somewhat disappointed.


Noticed this statement. I don't know if it's true or not.


I can't even imagine such a high debugging speed in this area.


I guess on this topic it only makes sense to ask for links to advanced work done by people. Thanks again.

Yes, there can be a lot of mishaps. Still you have to remember what LLM is. LLM can't think, it just guesses the next word based on huge input data for training.
Agent before giving you the code runs it, tests and checks it, and if necessary, corrects it and runs it again. But it cannot do this for MQL5, but it can test the algorithm and logic, for example, in Python.
Here is the continuation of my conversation with him:
https://claude.ai/share/80a7e530-5f72-4bc7-b24d-35a7aef70df0
and I attach the files he generated and tested in Python.

Files:
files (31).zip  8 kb
 
Nikolai Semko #:

Yeah, there's a lot of mishaps that can happen. Still, you have to remember what LLM is. LLM can't think, it just guesses the next word based on giant input data for training.
The agent runs the code, tests and checks it before giving it to you, and if necessary, corrects it and runs it again. But it cannot do this for MQL5, but it can test the algorithm and logic, for example, in Python.
Here is the continuation of my conversation with him:
https://claude.ai/share/80a7e530-5f72-4bc7-b24d-35a7aef70df0
and I attach the files he generated and tested in Python.

Here is another continuation, but already with MQL5 code correction:
https://claude.ai/share/5b977bc9-e902-4db3-aba2-ebc8b52b9979
Files:
files (32).zip  13 kb
 
Nikolai Semko #:
here is another continuation, but already with MQL5 code correction:
https://claude.ai/share/5b977bc9-e902-4db3-aba2-ebc8b52b9979
Unfortunately, the MQL5 code there does not just need to be corrected, but rewritten.
 
Nikolai Semko #:
Before providing you with the code, the agent runs it, tests and checks it, and, if necessary, corrects it and runs it again. But it cannot do this for MQL5
You can agree with him to put the code into pieces and test it in isolated conditions. Or run the code yourself and give him the compilation and execution logs. This way you can deal with it step by step (searching for errors in someone else's/forgotten large code) quite successfully.
 
fxsaber #:
The MQL5 code there does not just need to be corrected, but rewritten, unfortunately.
Networks can only do what they write about on every website. When I asked how to synchronise two time series, the they gave me some unrelated answers.
 
Rorschach #:
Networks can only do what every site writes about.
Would know sites with discussion of the topic.

Forum on trading, automated trading systems and testing trading strategies

Canvas is cool!

fxsaber, 2025.11.18:40 pm

I guess on this topic it only makes sense to ask for links to advanced work done by people.


It seems that on some topics the strongest results are created in MQL5. For example, the compressor posted today.

1...9596979899100101102
New comment