Is AI-assisted EA development actually improving Gold (XAUUSD) strategies or just making overfitting easier? - page 3

New comment
 
Ryan L Johnson #:

Don't forget to list prompters who are only good at prompting!😂

I know, I know, kids... "vibe coding."😎

Indeed, they’ve actually made a career out of it ; prompt engineering... But that doesn’t change the fact that you still need to convey the business logic to the LLM, and in this specific case, that business logic is trading
 
I’ve been using AI for over half a year now, and I’ve noticed big differences between tools from different developers.
 
First off, some AI come with preset biased ideas built in. Once I asked it to create an MT5 EA that doubles returns monthly, and it straight-up said it was impossible. It even kept warning me about online investment scams. I really didn’t know whether to laugh or cry!
 
Another common issue is poor memory. If you switch between different tabs or windows during a conversation, the AI can’t share previous data at all. It loses all context and repeats the same mistakes. Even on the same page, if you upload large files like CSVs, it may cut off its memory and forget everything we talked about earlier.
 
Third, most of its knowledge is pulled from online sources. It tends to use any logically sound content it finds without second-guessing. You have to be extra careful here. For indicators like MA and Bollinger Bands, the outputs aren’t linearly proportional, so they simply won’t work for certain uses. I’ve wasted quite a lot of time figuring this out.
 
Last but not least, it can make vague and risky judgments. When trading starts running at a loss, it’ll just keep adding more filters blindly. It makes everything seem fine on the surface, but leads you further down the wrong path — and that can eventually result in a total account wipeout.
(Remarks: Thanks AI to translate Chinese to English)
 
I think it's a good idea to use AI. But you need to be careful not to make mistakes.
 
AI can help with scaffolding and idea generation, but it can also make overfitting easier if the validation step is weak. The guardrails I would use are walk-forward validation, explicit symbol and timeframe checks, and a small live or demo micro test before trusting any result. It also helps to separate signal logic from execution logic so you can test each layer on its own. Are you running into trouble more on the idea side, or when turning the idea into stable code?
 
Hesham Ahmed Kamal Barakat #:
Are you running into trouble more on the idea side, or when turning the idea into stable code?

That is the predominant question in this thread, precisely!

IMHO, the MT5 Strategy Tester is exponentially more valuable than any AI engine for the purpose of developing a profitable strategy. AI can't analyze tick data, nor even OHLC data, so I really don't see how AI could affect overfitting. If you found a way to hook AI into the Strategy Tester, now that would be interesting.

If you test enough combinations of standard MT5 indicators, on enough standard timeframes and custom charts, you will eventually come up with a profitable strategy. I have done this manually in the past. Of course, I had to ultimately refine the strategy with time, day, and month filters, and risk-based lot sizing. Still, just imagine if I could have made AI do the first part for me.🤔

 
Daniel Andrew Clay:

Hey everyone,

After spending months building and testing EAs on XAUUSD using AI tools (ChatGPT, Deepseek, Copilot), I've noticed something interesting — AI makes it incredibly fast to generate complex strategies, but it also makes it dangerously easy to curve-fit a beautiful backtest that dies on a live account.

I wanted to open a real discussion here:

1. Have you used AI tools to build or improve your XAUUSD EA? Did it actually help in live conditions?

2. What's your process to verify a Gold EA is NOT overfitted — walk-forward? Out-of-sample? Monte Carlo?

3. Do you think AI-generated strategies can ever replace experienced discretionary logic for a volatile pair like Gold?

I'll share my own results: I built many EAs using AI assistance on M1, M5, M15, M30, H1 XAUUSD. Three failed within 2 weeks live. The third is still running after 6 weeks with modest drawdown But I honestly can't tell if it's working because the strategy is solid or if its just got luck with market conditions.

Curious what the community thinks. No sales pitches please just real experience.

Depends. When I use AI for assistance I always review every line of code and logic. Being in depth and reviewing what the AI blurted out if it's quality or not makes the difference.
 
Gio Rendel Masagca Rivadillo #:
Depends. When I use AI for assistance I always review every line of code and logic. Being in depth and reviewing what the AI blurted out if it's quality or not makes the difference.

Wow! Power to you in that regard.

Personally, I avoid that because I type like a demon and follow my own formatting. I guess that a certain level of MQL5 fluency is required to avoid AI as well though.

123
New comment