YouTube has discovered another machine for printing money.
You type a few sentences into ChatGPT, ask it to create a trading bot, connect it to your brokerage account, and fifteen minutes later you can watch artificial intelligence earn money on your behalf. Meanwhile, you are free to relax, pursue your passions, or watch another video about becoming a millionaire without capital, knowledge, or any unpleasant contact with reality.
The video titles themselves are masterpieces of modern economic literature: “Making Money With a ChatGPT Trading Bot in Minutes,” “$1,386 in 12 Minutes Using My ChatGPT Trading Bot,” “8% Monthly Return During Testing,” or a bot producing “insane results.”
The problem is that building a trading bot is now easy.
Building one that remains profitable over time is still brutally difficult.
It is the difference between building a car that can move and building one that can win Formula One. The YouTube video usually ends when the vehicle leaves the garage. Nobody shows what happens at the first corner.
LLMs Predict Words, Not Prices
A language model can create the impression that it understands the market. It can discuss interest rates, analyze a central bank announcement, and explain why a company’s share price increased after its earnings report.
Usually after it has already happened.
The basic task of an LLM is to predict the next fragment of text. It is not inherently designed to forecast future price distributions, manage financial risk, or maximize returns after transaction costs.
Worse, language models can still generate answers that sound confident and convincing while being completely false. OpenAI itself describes hallucinations as plausible but incorrect statements produced with unjustified confidence and acknowledges that the problem remains a fundamental limitation of language models.
In an ordinary conversation, this might mean an incorrect date or an invented quotation.
In trading, it can mean a buy order.
The market does not award points for elegant reasoning. It does not care that the model provided five logical reasons why a price should rise.
Money can be lost in beautifully written paragraphs.
Backtesting, or Time Travel in a Spreadsheet
Most miraculous trading bots presented online have one primary piece of evidence supporting them: a backtest.
The strategy was tested on historical data and made money. The equity curve goes up. The indicators glow green. Somewhere in the background, a course registration form is probably waiting.
The trouble is that if you test enough strategies and parameters, you can eventually find something profitable even in random data.
Research on backtest overfitting shows that repeatedly testing variations of a strategy allows developers to select a configuration that looks excellent in the past but collapses when exposed to new data.
With LLMs, there is an additional and rather creative problem: the model may have encountered information about the testing period during its training.
It may therefore be “analyzing” 2022 while indirectly knowing what happened afterward. Researchers have described this effect as a profit mirage. Impressive backtest results can disappear once the model reaches data beyond its knowledge cutoff.
The bot did not predict the future.
It remembered the past while the tester pretended that the future had not happened yet.
Working Code Does Not Mean a Working Strategy
In many videos, the “AI trading bot” is actually a conventional strategy based on RSI, moving averages, MACD, or a few conditions written in Pine Script or Python.
The LLM does not discover a market advantage.
It writes code for an idea supplied by a human.
That can still be useful. A model can prepare an integration with a broker’s API, build a backtesting module, create a logging system, or help find bugs in the code.
But the fact that a program can correctly submit orders does not mean that it should submit them.
Automation does not repair a bad strategy.
It merely allows the strategy to lose money faster, around the clock, without pausing for reflection.
The Costs Missing from the Thumbnail
A strategy may look profitable before accounting for:
- commissions,
- bid-ask spreads,
- slippage,
- liquidity problems,
- delayed market data,
- partially filled orders,
- taxes,
- API and infrastructure costs.
Even LiveTradeBench, an experiment that tested 21 models during 50 days of trading on live data, acknowledges that excluding commissions, spreads, liquidity, and slippage can inflate performance, especially when models frequently change their positions.
The study also found that strong performance on popular language benchmarks did not meaningfully translate into better trading results.
A model may be excellent at writing code, passing exams, and explaining quantum physics while still trading like someone who discovered the “Buy” button five minutes ago.
The US Commodity Futures Trading Commission has explicitly warned that AI does not turn trading bots into money-making machines. It cannot predict the future or sudden market changes.
The regulator also highlights commissions, spreads, and subscription fees, all of which are remarkably effective at consuming theoretical profits.
A Strategy Works Until the Market Changes Its Mind
Markets are not stable systems.
A strategy may perform well during a calm bull market and then stop working when interest rates change, a crisis begins, a war breaks out, a pandemic appears, or market participants simply start behaving differently.
A large study called FINSABER tested LLM-based trading strategies across more than one hundred financial instruments and approximately twenty years of market data.
Previously reported advantages deteriorated significantly when the tests were expanded across longer periods and more diverse markets.
The models were too cautious during bull markets while remaining too aggressive during bear markets. They suffered substantial losses precisely when caution would have been most valuable.
There is something almost human about it.
Being afraid while everything is rising and recovering your confidence shortly before disaster.
What About Research Showing That LLMs Can Predict Markets?
Such research does exist.
A study by Alejandro Lopez-Lira and Yuehua Tang found that advanced models could extract short-term signals from financial news headlines, particularly for smaller companies and negative news.
That does not mean we can ask ChatGPT, “Should I buy Tesla tomorrow?” and begin selecting the color of our yacht.
In the study, the model was part of a controlled process. It received a specific type of data, performed a narrowly defined task, and its outputs were aggregated and evaluated statistically.
That is an application of LLMs for text processing.
It is not a digital Warren Buffett sitting at a Bloomberg terminal.
An LLM can be valuable as a supporting tool. It can:
- classify financial news,
- summarize reports and earnings calls,
- identify risks in corporate documents,
- help write and test code,
- compare scenarios,
- monitor logs and anomalies.
AI is already used in finance for data analysis, risk management, surveillance, fraud detection, and supporting the development and testing of strategies.
At the same time, the CFTC emphasizes the need for oversight, auditing, data quality controls, and human involvement in autonomous financial systems.
That is a reasonable role.
An LLM can act as a fast analyst whose work still needs to be verified.
It should not act as a fund manager who has just been given the brokerage password.
What Is Missing from the Miraculous Bot Videos
A video demonstrating a real strategy should present considerably more than a green chart.
It should include long-term testing across different markets, data genuinely unknown to the model, out-of-sample performance, walk-forward testing, commissions, spreads, slippage, maximum drawdown, parameter stability, and several months of live trading.
Preferably with results the creator cannot quietly delete before recording another video.
Without that, we are watching a software demonstration, not evidence of a market advantage.
The bot may work.
It may connect to an exchange.
It may buy and sell assets.
It may even make money for a few weeks.
A coin can also land on heads five times in a row. That does not mean we have created artificial intelligence capable of predicting coins.
An LLM is an excellent tool for building trading systems.
That is one of the worst possible reasons to allow it to trade independently.
Give it documents, data, code, and logs.
Keep the brokerage password to yourself.