A model will happily tell you that gold has support at a specific price, that the market is pricing a certain probability of a rate cut, and that a pair looks bullish on the four hour chart. Three statements, three different failure modes. The first is invented. The second was true on some date you cannot see. The third is a description of a chart the model never opened.
None of that makes the tool useless. It makes it a tool for a narrower job than the marketing suggests, and the narrow job is worth having.
The three tasks that genuinely work
Reading long documents is the first. Central bank statements, quarterly reports, platform documentation, a prop firm rulebook, a broker's terms. Paste the text in and ask what changed against the previous version, or what a specific clause means in plain language. The source is in front of the model, so the answer is checkable, and checking takes seconds.
Writing and fixing code is the second. Indicator logic, a Pine Script error, an alert condition, a spreadsheet formula for position sizing. Code has a hard pass or fail, which is the whole reason this is safe. If it does not compile or the plot looks wrong on the chart, you know immediately.
Reviewing your own history is the third, and it is the one most traders skip. Export the trade list, hand it over, and ask specific questions. What is my average result by day of the week. How many trades did I take within thirty minutes of a losing trade. What happened to my average holding time in the worst month. This is the analysis a trading journal is supposed to make possible, and it takes a model about a minute.
Why the market calls are worthless
Two reasons, and they are different.
The first is data. A model without a live connection is answering from training data with a cutoff, so anything about current prices, rates, spreads or this week's calendar is stale by construction. A model with a browsing connection is better but is only as reliable as what it fetched, and it may fetch a blog post rather than the actual release. Numbers that matter belong to the price feed and the economic calendar, full stop.
The second is more subtle. A language model produces the most plausible continuation of your prompt. Ask whether a setup looks good and you have already told it what you want to hear. Ask the same question twice with the bias reversed and you will often get two confident, contradictory answers. There is no internal position being defended, so agreement costs the model nothing.
Nothing a language model outputs about a market is financial advice, and it should not be treated as a signal. Trading, particularly with leverage, carries a high risk of loss, and an idea generated by software is no safer than one generated by a stranger.
Prompts that produce something usable
The pattern that works is to supply the facts and ask for processing, rather than asking for facts. Compare two requests. "Is EURUSD bullish?" gives you fiction dressed as analysis. "Here are my last 180 trades as CSV. Group them by instrument and session, show win rate, average win, average loss and expectancy for each group, and list any group with fewer than fifteen trades separately" gives you arithmetic you can verify.
Three habits improve output sharply. Give the model the raw material instead of asking it to recall. Ask it to show the intermediate numbers so you can check one row by hand. And ask it explicitly what would make its answer wrong, which surfaces assumptions it would otherwise leave buried.
Turning output into a rule
A pattern found in your history is a hypothesis, not a strategy. If a model reports that your Friday afternoon trades are your worst group, that could be a real behavioural effect or thirty trades of noise. The next step is the same one you would apply to any idea: define the rule precisely, test it against history with proper backtesting, then forward test it on small size before it changes anything real. Models are good at generating candidate rules and have no ability whatsoever to tell you which ones survive.
The same scepticism applies to anything a model says about third party services. Ask it to evaluate a signal provider and it will produce a confident assessment of a service it knows nothing about. The checks that actually matter are in evaluating signal services, and every one of them requires looking at real records rather than asking a model for an opinion.
Where the line sits
I keep one rule and it has never needed adjusting: the model can read, count and write, and it cannot decide. Analysis of what already happened, summarising, code, structuring a plan, drafting a checklist, explaining a rule I do not understand. All fine. Direction, entry, stop, size and whether to take the trade at all come from the chart and the written plan, which is the point of having a trading plan in the first place.
The broader question of where these tools belong in a trading workflow, including what firms should and should not deploy to clients, is covered in AI assistants in trading. The short version is that the guardrails matter more than the model, and the most important guardrail is that nothing generated by software gets to place an order.
"Ask it what happened and it is often excellent. Ask it what will happen and you are reading a very well written guess with your money behind it."
— Alex Onta, Executive Director, SINGUARD
Key Takeaways
- Language models are strong at reading long documents, writing indicator code and analysing your own trade history, because all three outputs can be checked.
- Current prices, rates and calendar events must come from the platform and the calendar, never from a chat window with a training cutoff.
- Supply the raw material and ask for processing. Asking for facts or opinions produces confident text with nothing behind it.
- A pattern the model finds in your history is a hypothesis. It goes through backtesting and forward testing before it changes any rule you trade.
Frequently Asked Questions
Can a language model give me trading signals?
No, and treating its output as a signal is the main way traders get hurt with these tools. It has no live price data and it will agree with whatever bias your question carries. Trading carries a high risk of loss regardless of the source of the idea.
What is the single most useful thing to ask one?
Hand it your own trade export and ask for grouped statistics: results by instrument, session and day, average win against average loss, and behaviour after a losing trade. It is arithmetic you can verify and most traders never do it by hand.
Is it safe to paste broker or account details into a chat?
No. Account numbers, credentials, API keys and client identity documents should never go into a third party chat. Firms handling client data need models running under their own keys with logging and retention they control.
About the Author
Alex Onta is an Executive Director at SINGUARD. He built eTrader, the terminal, the mobile apps, eTrader Broker, Copytrading, Business and Community, along with the worldwide clustered-server infrastructure it all runs on, with his brother Roman Onta helping on the design, and he leads that division today. Together with Roman he builds the Prop Firm CRM, the Broker CRM, Scalegram and CopySignals, and the two of them carry worldwide compliance, payment processing and international business structuring side by side. He lives and works in Dubai for most of the year. Meet the executive duo leading Singuard's five divisions.