Coding trading bots with AI (not being a programmer)

New comment
 

Hello everyone,

This topic is aimed at those who, like me, are not professional programmers but still want to build effective trading bots. My programming background is minimal—I studied a bit of BASIC and COBOL back in 1985, when those languages were still widely used. Despite that, I’ve been exploring algorithmic trading for just over a year now.

During this time, I’ve experimented with several AI assistants. I began with ChatGPT, then moved to Copilot, briefly tested Gemini, and eventually tried Claude—which, in my experience, has been the most consistent at producing error‑free code on the first attempt. These tools are evolving rapidly. When I first started, Copilot wasn’t particularly strong, and ChatGPT seemed to be the most reliable. Later, Claude became my go‑to option for clean code generation. More recently, however, I’ve returned to Copilot, which has improved significantly over the past twelve months.

Interestingly, it was Gemini that managed to identify a major error in one of my recent scripts—an issue that none of the other platforms detected. That was a surprising and valuable reminder that each tool has its strengths.

At the moment, I use a combination of these “helpers,” and this hybrid approach has produced some encouraging results.

I’m curious to hear about your experiences with these platforms. How have they performed for you, and what workflow have you found most effective? Grateful for any input you may have.

 
smspano@hotmail.com:

Hello everyone,

This topic is aimed at those who, like me, are not professional programmers but still want to build effective trading bots. My programming background is minimal—I studied a bit of BASIC and COBOL back in 1985, when those languages were still widely used. Despite that, I’ve been exploring algorithmic trading for just over a year now.

During this time, I’ve experimented with several AI assistants. I began with ChatGPT, then moved to Copilot, briefly tested Gemini, and eventually tried Claude—which, in my experience, has been the most consistent at producing error‑free code on the first attempt. These tools are evolving rapidly. When I first started, Copilot wasn’t particularly strong, and ChatGPT seemed to be the most reliable. Later, Claude became my go‑to option for clean code generation. More recently, however, I’ve returned to Copilot, which has improved significantly over the past twelve months.

Interestingly, it was Gemini that managed to identify a major error in one of my recent scripts—an issue that none of the other platforms detected. That was a surprising and valuable reminder that each tool has its strengths.

At the moment, I use a combination of these “helpers,” and this hybrid approach has produced some encouraging results.

I’m curious to hear about your experiences with these platforms. How have they performed for you, and what workflow have you found most effective? Grateful for any input you may have.

Although I’m a programmer, I believe this approach works for everyone. For complex projects, especially trading EAs, I’ve found the best results come from breaking the strategy into very small, clear steps, almost like writing pseudocode.

Instead of asking an AI to build the entire system at once, I prompt it to handle one specific task at a time. For example, if the EA is based on trend lines, I start by asking it to identify all swing lows on the chart. Once that is implemented, tested, and confirmed to work correctly, I move to the next step, such as identifying only the most recent two swing lows if that’s all the logic requires.

Each step is coded, tested, and validated before moving on to the next. This incremental approach consistently produces more accurate and reliable results than generating everything in one prompt.
 
smspano@hotmail.com:

Hello everyone,

This topic is aimed at those who, like me, are not professional programmers but still want to build effective trading bots. My programming background is minimal—I studied a bit of BASIC and COBOL back in 1985, when those languages were still widely used. Despite that, I’ve been exploring algorithmic trading for just over a year now.

During this time, I’ve experimented with several AI assistants. I began with ChatGPT, then moved to Copilot, briefly tested Gemini, and eventually tried Claude—which, in my experience, has been the most consistent at producing error‑free code on the first attempt. These tools are evolving rapidly. When I first started, Copilot wasn’t particularly strong, and ChatGPT seemed to be the most reliable. Later, Claude became my go‑to option for clean code generation. More recently, however, I’ve returned to Copilot, which has improved significantly over the past twelve months.

Interestingly, it was Gemini that managed to identify a major error in one of my recent scripts—an issue that none of the other platforms detected. That was a surprising and valuable reminder that each tool has its strengths.

At the moment, I use a combination of these “helpers,” and this hybrid approach has produced some encouraging results.

I’m curious to hear about your experiences with these platforms. How have they performed for you, and what workflow have you found most effective? Grateful for any input you may have.

If you don't understand programming, even when coding with AI, even using multiple AI assistants, you'll still likely encounter errors. This is because, while fixing errors, the AI, trying to please the user, may slightly or significantly alter the logic of an advisor or indicator.
 
AI can help
to code, to learn if you know how to ask the right questions
if you know how to read the provided code

Because, on the one hand, it performs well for simple tasks, but as soon as you increase the complexity, it will fail.

So for simple strategies, why not... (it won't produce profitable EAs)

But if you want to find the holy grail, you'll have to learn to structure your ideas, your codes
New comment