New EA DeepGold LLM EA – when gold listens to artificial intelligence (well, almost)
- Thread starter NPNavigator
- Start date
- Credits
- 831
Error 4014 in the status of your dashboard is a classic of the "too smart robot bumped into the security service" genre. Imagine: your advisor is trying to reach the AI brain on DeepSeek servers, and the terminal responds to him: "It's not allowed! Where is the pass?"
Technically, the code crashes when calling WebRequest(). By default, MetaTrader 5 prohibits any Internet requests until you personally give the go-ahead to specific URLs. Without this, EA sees only 4014 – Function is not allowed.
How to fix it:
1. Open the terminal → Tools → Settings (or Ctrl+O).
2. Go to the Advisors tab.
3. Check the box "Allow WebRequest for specified URLs.
" 4. Click the "Add" button and type: https://api.deepseek.com
5. Click OK and restart the Advisor.
That's it, the doorman has now been issued — AI analytics will work, the status will change from "Error: 4014" to something life-affirming like "OK (verified)" or just "OK".
I specifically added a connection check at startup (CheckAPIConnection()) to immediately understand if there is access to the API. I wrote the code honestly: "I'm testing the connection..." → received a turn from the gate → showed an error in the panel. The humor is that the robot is fully functional, it just wasn't given an Internet fishing rod.
Thank you very much! Works.
- Credits
- 831
Hello, colleagues in the analysis of rakes on the topic of "neural networks in trading"!
I remind you: this Expert Advisor is a pure experiment. Here we are checking whether a large language model (in my case, DeepSeek) is capable of really giving useful signals for gold, or whether all this LLM hype in algotrading is just an expensive way to merge a deposit with beautiful logs.
Since the first version, I have made a couple of important improvements that, hopefully, will bring the answer to this question closer.:
1. I replaced the self-written JSON crafts with the normal JAson.mqh library. I attached the file below, if anyone does not know, this file needs to be added to the Include folder. Previously, the parsing of the neural network's response could break on any tricky symbol — and we silently missed the signal. Now everything is handled strictly and without nerves.
2. Added saving the prediction history to the file. Previously, after the restart, the ADVISOR "forgot" all its feedback, and each time the LLM started from scratch. Now the accumulated statistics (where she guessed and where she pointed her finger at the sky) are not lost and are sent to the model with each request. In theory, this should help the neural network to gradually calibrate precisely for our gold on the M15. Or she can just prove convincingly that she doesn't know how to study, which is also a result!
What is the result? The EA has become more stable, loses less signals, and — most importantly — gives LLM a chance to gain real-world experience. Time will tell if this will help her become a "golden prophet" or if we will just get a beautiful collection of her mistakes. Join the experiment and share your observations. Hype with hype, and statistics will judge.
I remind you: this Expert Advisor is a pure experiment. Here we are checking whether a large language model (in my case, DeepSeek) is capable of really giving useful signals for gold, or whether all this LLM hype in algotrading is just an expensive way to merge a deposit with beautiful logs.
Since the first version, I have made a couple of important improvements that, hopefully, will bring the answer to this question closer.:
1. I replaced the self-written JSON crafts with the normal JAson.mqh library. I attached the file below, if anyone does not know, this file needs to be added to the Include folder. Previously, the parsing of the neural network's response could break on any tricky symbol — and we silently missed the signal. Now everything is handled strictly and without nerves.
2. Added saving the prediction history to the file. Previously, after the restart, the ADVISOR "forgot" all its feedback, and each time the LLM started from scratch. Now the accumulated statistics (where she guessed and where she pointed her finger at the sky) are not lost and are sent to the model with each request. In theory, this should help the neural network to gradually calibrate precisely for our gold on the M15. Or she can just prove convincingly that she doesn't know how to study, which is also a result!
What is the result? The EA has become more stable, loses less signals, and — most importantly — gives LLM a chance to gain real-world experience. Time will tell if this will help her become a "golden prophet" or if we will just get a beautiful collection of her mistakes. Join the experiment and share your observations. Hype with hype, and statistics will judge.
This EA primarily uses DeepSeek AI to predict market direction (up/down). Since the trading win rate is likely highly dependent on the prompt engineering based on historical predictions, I suggest adding a customizable prompt input box to the parameter panel, so that users can experiment and optimize the prompts themselves.