Open the MQL5 Market, search for gold, and half the top results say AI, neural network or machine learning. Some of them mean it. Some of them are a moving average with a new name. Before you pay for an AI gold trading EA, you want to know which one you’re looking at, and whether the AI is what makes it money.
I build an AI EA, Alpha Pulse AI, so I have a stake in this. That’s why every check below is one I’d want you to run on mine too. Not financial advice.
Two very different things are both called “AI”
Before any check, find out which kind it is:
- A model inside the EA. Someone trained a model (a neural network, gradient boosting, anything) on historical data and put it in the file. It runs offline, it can be backtested, and its weak point is the years it was trained on.
- An EA that calls a language model. It sends the market context to GPT, Claude, Gemini or similar over the internet and acts on the answer. It needs a WebRequest URL allowed in MetaTrader and an API key that you pay for. It can’t be backtested in the Strategy Tester, because WebRequest doesn’t run there.
If an EA says it uses ChatGPT but never asks you for an API key or a URL, it isn’t calling ChatGPT.
Check 1: does it really call the model it names?
For an EA that claims a language model, MetaTrader tells you. In Tools, Options, Expert Advisors, an EA that calls OpenAI, Anthropic or Google needs their address in the list of allowed URLs, and it needs your key. No key, no URL, no calls. I showed how to spot the fakes in the video below: every big model release brings a wave of them.
Check 2: does the AI change the result?
An AI that adds nothing is decoration. The question to ask: what does the same strategy make without it?
In my My AI Gold review I tested exactly that. Keeping its exits and its hours, random entries made a profit factor of 1.04 and a simple trend-and-pullback rule made 1.01. Its own picks made 2.22. That’s a strong sign the model is real and does the work. If a seller can’t show you something like that, you don’t know what you’re paying for.
Check 3: which years has the model seen?
A model trained on history knows those years. It will always look good on them. The years that tell you something are the ones it never saw.
Set the backtest start years before the seller’s chart begins. If it trades nothing, ask why. My AI Gold won’t open a trade before January 4, 2016. When I ran real 2012 to 2015 gold prices relabeled as later years, it was still profitable, but with a profit factor of 1.28 instead of 2.25. Same EA, very different picture.
Check 4: does the live account match the backtest?
Open the live signal, backtest the same months and compare them trade by trade. With My AI Gold, 89 trades matched with the same results, but the live account also took 25 trades the version on sale today doesn’t, and those lost money. Live profit factor: 1.84. The backtest of the same months: 2.69. A backtest that looks like the live account minus its losses is worth knowing about before you buy.
For an EA that calls a language model there’s no backtest to compare, so the only honest evidence is a forward test. Here’s what to record in an AI trading forward test.
Check 5: can you see what it decided, and who has the last word?
An AI you can’t inspect is a black box with a price tag. Ask two things:
- Do you see each decision and the reason for it? Not a vague “AI score”, but the action, the confidence and why.
- Which limits can the AI not break? Position size, daily loss, maximum trades, drawdown protection. The model should decide inside your limits, never around them.
And one practical question: who pays for the AI? A model inside the EA costs nothing to run. An EA that calls a language model costs API credits on every decision, paid by you.
AI gold trading EA: the five checks in one table
| Check | Model inside the EA | EA that calls a language model |
|---|---|---|
| 1. Real calls | Runs offline, no key needed | Needs an allowed URL and your API key |
| 2. With vs without AI | Ask for the same rules without the model | Ask what the EA does if the model says wait |
| 3. Years it hasn’t seen | Backtest years before its chart | No backtest: forward test on demo |
| 4. Live vs backtest | Same months, trade by trade | Live account vs your own demo run |
| 5. Decisions and limits | Ask what you can see | Verdict, reasoning and hard risk limits |
How Alpha Pulse AI answers the same five checks
Mine is the second kind: it calls the model you choose, with your own key. GPT-6.1 Sol, Claude Opus 5.5 and Fable 5.1, Gemini 3.8 Flash, Grok 4.6, DeepSeek V4, Qwen 3.7 and Xiaomi MiMo. So here’s how it does on the list:
- Real calls: it needs the provider’s URL allowed and your API key. The cost is usually around $50 a month, depending on the model.
- Decisions you can see: the model returns a verdict (buy, sell or wait) with its confidence, risk and reasoning, the panel shows the decision history, and every request is logged.
- Limits it can’t break: daily loss limits, position sizing, maximum trades and drawdown protection are yours. The model decides inside them.
- Years it hasn’t seen: it can’t be backtested, like any EA that calls a language model. That’s a real limit, and it’s why there’s a free trial: run it on your own demo and judge the forward test.
If you go for it, here’s the ten-minute setup with Claude Opus 5.5, and the list of what to check before you buy Alpha Pulse AI.
Watch out for fake AI bots: in GPT-5.5 Is Here. Fake AI Trading Bots Are Coming Too, I show how to spot an “AI bot” that never calls the model at all.
If you test gold EAs, the data matters. I use Axi‘s XAUUSD history for my tests. Affiliate link; I may earn a commission if you open an account through it. And if your plan includes allocated capital, keep that decision separate from the AI: Axi Select, for example, only allows a third-party EA for generating signals, with trades placed manually.
A good AI shows you the years it hasn’t seen. I test every new model and every AI EA I can get my hands on. Join the DoItTrading newsletter to get the next test.
Trading with AI EAs involves real risk to capital. API costs depend on your settings and usage. Past results of any model don’t guarantee future ones. This is not financial advice.
Frequently Asked Questions
How can I tell if an AI gold trading EA really uses AI?
First find out whether it has a model inside or calls a language model. If it calls one, it needs an allowed URL and your API key. Then ask what the same strategy makes without the AI, test the years before its backtest starts, and compare its live signal with the backtest trade by trade.
Can I backtest an EA that uses ChatGPT, Claude or Gemini?
No. Those EAs call the model through WebRequest, which doesn’t run in the MetaTrader Strategy Tester. The honest test is a forward test on a demo account.
Is My AI Gold real AI?
Probably. In my test, with its exits and hours, random entries made a profit factor of 1.04 and its own picks 2.22, and no simple rule reproduced its entries. But it won’t trade before 2016, and on 2012-2015 gold its profit factor dropped to 1.28.
Why does it matter which years an AI model has seen?
A model always looks good on the years it was trained on. The years it never saw are the closest thing to the future you can test, so a big drop there tells you what to expect live.
How much does an AI trading EA cost to run?
A model inside the EA costs nothing extra. An EA that calls a language model pays API credits on every decision; for Alpha Pulse AI that’s usually around $50 a month, depending on the model and settings.
Can I try Alpha Pulse AI before buying?
Yes. There’s a free 7-day trial for your own demo account, with your own API key. The MQL5 Market demo can’t run the AI because it only works in the Strategy Tester.
