ChatGPT-Style AI Trading EAs: An Honest Guide to What Works, What Fails, and What to Ask Before You Buy
AI-powered Expert Advisors are the fastest-growing shelf in the Market right now, and they arrive under several different identities. Some are sold as autonomous analysts: the AI reads the market and opens the trades. Some hand you the controls of the model itself: describe how you want to trade in plain English, plug in an API key, and the model executes your words. And some keep a conventional strategy at the core and use the model as a gatekeeper — the last yes or no before a signal becomes an order.
Different sales pitches, same underlying question. I build and sell trading systems for a living, and I use AI every single day — more than most sellers in this Market, I suspect. So this is not a post against AI. It is a post about which seat the model sits in. Because underneath every one of those pitches the technical question is identical: is a probabilistic model part of the decision path for your money — and if it is, can you still test the system, measure its risk, and verify what it does?
Let me give you the honest, technical answer — advantages, disadvantages, what genuinely works, what cannot work, and a checklist you can use on any "AI EA" listing you meet.
Why this category is exploding
Three real reasons, and none of them are stupid.
First, accessibility. Most traders have a strategy in their head but cannot code. A prompt feels like programming in plain English — and honestly, that desire is legitimate.
Second, personalization. A prompt-driven robot promises to trade your way: your risk tolerance, your sessions, your style. Off-the-shelf EAs cannot promise that.
Third, the models really did get good. Modern LLMs summarize a news feed, explain a chart, or critique a trading plan better than most humans. If you have used one seriously, you know the capability is not fake.
So the demand is rational. The problem is not the demand — it is where the current products put the model.
What a language model actually is (the one paragraph that matters)
An LLM is a probabilistic text predictor. Given the same input, it does not always produce the same output — by design. It has no memory of your account unless you feed the state in every time. It does not "know" prices; it reads whatever snapshot the EA sends and writes a plausible continuation. It can be brilliant, and it can be confidently wrong, and it cannot tell you which one just happened.
Keep that paragraph in mind. Every structural problem below follows from it.
One distinction before the problems, because it matters. Some "AI" EAs never call an external model at all. They carry a small trained network inside the file itself (MetaTrader supports ONNX models natively). These are deterministic programs: same input, same output, fully testable in the Strategy Tester. The five problems below do not apply to them.
They have their own trap instead: the look-ahead moves from the code into the weights. A network learns from its training data. If it was trained on the same years the backtest covers, the backtest proves nothing. The model already knows where every crash and every losing period is — it saw them during training. It does not avoid them with skill. It avoids them with memory.
This also means the backtest can be made perfect again with every new version. A drawdown happens in live trading. The seller retrains the model on data that now includes that drawdown and ships an update. In the new version's backtest, that losing period is gone — as if it never happened. Live losses in, clean history out, version after version. I am not describing a theory. Years ago, before I built systems myself, I paid for one of these — a popular one at the time — and watched my live loss disappear from the next version's backtest. I am fairly sure I am not the only one here with that memory. So judge an embedded-model EA only on data after its training cutoff, and ask the seller for that date. If the past keeps getting better with every update, that is your answer.
Everything below concerns the other kind: systems where a large language model, reached through an API, sits somewhere in the live decision path. "Somewhere" includes the gatekeeper setups, where a fixed strategy produces the signal and the model gives the final yes or no. If a random judge can veto or approve each trade, the results are just as unrepeatable as if the model traded by itself. A coin flip at the end of a fixed pipeline still makes the output a coin flip.
The five structural problems when an LLM sits in the live decision path
1. You cannot backtest it. This is the big one, and there is no engineering workaround. A backtest requires that the decision logic be reproducible on historical data. An LLM's decisions are not reproducible even on live data — same candles, same prompt, different answer tomorrow, and a different answer again when the provider updates the model. Any "backtest" of a prompt-driven system is therefore a hand simulation: a human imagining what the model probably would have said. The Strategy Tester — the one honest referee this platform gives you — cannot referee this product class at all.
2. Non-determinism is a feature of the model and a bug in your risk plan. Sampling temperature, provider-side load balancing, silent model revisions — all of these mean the same market situation can produce different trades on different days. Your risk numbers assume a distribution of outcomes from one strategy. A stochastic decision-maker is not one strategy; it is a drifting family of strategies wearing one name.
3. Silent model drift. Automatic model upgrades are usually framed as a benefit — the same prompt keeps working while the intelligence underneath improves on its own. Viewed from a risk manager's chair, the same feature is open to a very different reading: the behavior of a live trading system changes, without a changelog, on a third party's schedule. In every other corner of algo trading that is called an incident. In this category it tends to appear as a selling point.
4. Latency and cost reality. A chat API round-trip is measured in hundreds of milliseconds to seconds — on a good day, before provider queueing. "AI news scalping" that reacts to NFP within seconds through a chat API, against news-time spreads, is a fantasy genre. And every decision costs money: API fees scale with how often the robot thinks, which quietly rewrites your cost-per-trade in a way no backtest ever showed.
5. Unverifiable risk bounds. For a deterministic portfolio you can run Monte Carlo stress tests and publish a drawdown range: "this preset lives in roughly this band." For an LLM-driven system, no such band can exist, because the decision process itself is not fixed. The only honest drawdown statement for this product class is: unknown, and unknowable in advance. You will never see that sentence on a sales page, which tells you something.
What AI is genuinely good at in trading
Now the other half, because AI absolutely earns a seat at the table — the advisor's seat, not the executioner's.
Research and synthesis. Summarizing a central bank statement, comparing broker specifications, digesting a 40-page paper into the three things that matter. LLMs are world-class here, today.
Code review and development. Auditing an EA's logic, hunting for look-ahead bias, writing test harnesses, porting a strategy between platforms and proving the port is faithful. This is where I personally use AI most heavily — as a tireless, pedantic engineering reviewer. It has caught things I would have missed.
Journaling and post-trade review. Feed it your closed trades and it will find the patterns you are blind to: revenge entries after losses, size creep on winning streaks, the pair you always mismanage. As a mirror, it is honest and cheap.
A risk officer that reads, not trades. An assistant that receives your structured account state — open positions, exposure per currency, current drawdown — and writes you a plain-language morning brief with rule-violation flags. It never sends an order. It makes you a more disciplined trader without ever being able to hurt you.
Translating plain English into fixed rules — once. This, I think, is the future the prompt-EA category is groping toward without admitting it. The honest version of "the prompt is the strategy" is: describe your strategy in English, let the model translate it into an explicit, inspectable rule set — indicators, thresholds, exits — and then freeze it. From that moment a deterministic engine executes the frozen rules, and suddenly everything works again: the Strategy Tester works, Monte Carlo works, walk-forward works, live-vs-backtest comparison works. AI wrote the specification; the specification trades. The intelligence is in the design room, not on the trigger.
If AI must touch execution: the only defensible architecture
Some builders will put models on the trigger anyway. Fine — then there is exactly one architecture that deserves your money: the LLM proposes, hard-coded rules dispose.
That means a deterministic risk layer written in ordinary MQL5, sitting between the model and the order queue: maximum risk per position, maximum concurrent exposure, a news blackout window, a daily stop — enforced in compiled code that no prompt can override. Hard stop-losses on the server, never "mental stops" managed by the model. Some configurations circulating in this category lean on trading without hard stops, adding to winners as they run, or raising risk after a good streak. None of that is an AI innovation — those are the oldest account-killers in this business, wearing a new vocabulary.
A buyer's checklist for any "AI EA" listing
Seven questions. Ask them in the comments section before you spend a cent.
- Can it run in the Strategy Tester with the same decision logic it uses live? If not, every performance claim on the page is unverifiable by construction.
- What exactly happens when the API is down, slow, or rate-limited? "It waits" is an answer; make them say it.
- Who pays for the model's thinking, and how much per month at your trade frequency? Get a number.
- What changes when the provider retires or upgrades the model? If the answer is "nothing, it just gets smarter," the seller has told you the strategy mutates silently.
- Are the stop-losses hard orders on the server, or "managed" by the AI? Only one of these survives a disconnection.
- Can the seller show you the same decision twice? Same inputs, same output, on demand. Determinism is demonstrable in five minutes — or it is absent.
- Is there a live, verified signal, and how long has it run? Screenshots of one good week are not a track record for a system that changes with every model update.
A listing that survives all seven is rare. That rarity is the point.
Where I stand
Everything published under my products is deterministic end to end: same data in, same trade out, reproducible in the Strategy Tester by anyone, with drawdown ranges measured by Monte Carlo stress testing rather than estimated by feel, and a live signal doing the verifying in public. Not because AI is beneath the work — the opposite. AI sits in my development pipeline every day: reviewing engine code, auditing statistical tests, stress-testing assumptions, drafting documentation. It has made the products meaningfully better.
It will still not be given the trigger. The trigger belongs to rules that can be tested, and to a risk budget that can be written down before the first trade — because "start with the loss, not the profit" only means something if the loss can actually be computed.
To AI or not to AI was never the real question. Put the model where it is brilliant — research, review, explanation, design — and keep execution where it is honest: deterministic, testable, and boring. Boring is what survives.
If you want to see what the deterministic side of that philosophy looks like in practice, the portfolio systems on my profile are built exactly this way — fixed rules, published risk bands, live verification. https://www.mql5.com/en/users/seckinerkut/seller
