AI Trading: What It Can and Can't Do
AI trading means using machine-learning models or AI assistants to research markets, write strategy code or place orders. The tool can speed up the work around a strategy, but it cannot see the future, and the CFTC warns that promised AI returns are a common fraud pitch.
This page is about the claims around AI in trading: which ones hold up, which ones regulators have flagged, and what the tools are good for.
Two other pages cover the neighboring ground, so this one does not repeat them: trading bots tests whether a rule-based system survives new data, and automated trading covers running one live.
How it works
“AI trading” covers three different things, and most confusion comes from mixing them.
The first is a statistical model trained on past prices to score what happens next. The second is a language model, the kind of assistant that writes text and code, used to research, summarize or build a strategy. The third is any of those connected to a broker account so that it places orders.
The first kind is pattern-fitting. It finds relationships in historical data. That can be useful, but every relationship it finds is a description of the past, and markets change what they reward. A model with many adjustable parts can fit past noise perfectly, which is the problem the overfitting page covers.
The second kind is a fast assistant. It can turn a written rule into Pine Script or Python, explain an error, or draft a checklist. It cannot tell whether the rule has an edge. Asked to “make a profitable strategy,” it will produce code that runs, which is not the same thing.
The third kind just removes the human from the order ticket. Whatever the signal is worth, automation executes it faster and without hesitation, including when it is wrong.
What regulators have warned about
The CFTC’s customer advisory is titled “AI Won’t Turn Trading Bots into Money Machines.” Its central line is that AI technology “can’t predict the future or sudden market changes.”
It warns that scammers claim AI algorithms “can generate huge returns—sometimes tens of thousands of percent—or yield 100 percent ‘win’ rates.” The CFTC announced it in press release 8854-24, dated 25 Jan 2024.
The same advisory describes one case in detail. It says Mirror Trading International promised “at least a 10 percent monthly return (or more than 200 percent per year)” and took nearly 30,000 bitcoins, which it puts at about $1.7 billion at the time.
FINRA, the SEC’s investor education office and NASAA issued a joint alert on 25 Jan 2024, titled “Artificial Intelligence (AI) and Investment Fraud.” Its examples include a pitch for a proprietary AI trading system said to be unable to lose, and it warns of deepfake audio and video used to impersonate executives or relatives.
Its first check is whether the person or firm is registered, on Investor.gov.
The SEC has also acted against firms that overstated their AI. On 18 Mar 2024 it charged two investment advisers, Delphia (USA) Inc. and Global Predictions Inc., with false and misleading statements about their use of artificial intelligence. The penalties were $225,000 and $175,000, $400,000 in all.
What AI can do in a trading workflow
Write and read code faster. Turning a rule into a script, then asking for a line-by-line explanation, is a genuine time saving, provided you read the explanation.
Check work you have already done. An assistant can look for look-ahead errors in a backtest, such as a signal that uses a bar’s close before the bar has closed.
Summarize documents. A filing, a fund prospectus or an exchange rulebook can be condensed quickly, as long as the numbers are checked against the original.
Keep a journal honest. Asking it to sort your own trade notes by the reason each trade was taken can surface a pattern you had not noticed. The judgment about what to change stays with you.
A worked example
Test the pitch in the CFTC’s case with arithmetic. A promise of 10% a month sounds modest until it compounds. $10,000 grown by 10% a month for 12 months is $10,000 times 1.1 to the 12th power, which is $31,384.28, a gain of 213.84% in a year. That matches the advisory’s “more than 200 percent per year.”
Carry it one more year. After 24 months the same $10,000 would be $98,497.33. A claim that doubles and redoubles money on schedule is a claim that the seller could fund itself without customers, which is the question to ask of any pitch like it.
Then a “100 percent win rate.” A strategy can show every trade as a winner by never closing losers, for example by averaging down or holding until price returns. The win rate is real; the open loss behind it is simply not counted. Ask for the largest open loss, not the share of closed winners.
The original data
In the 24,971-video corpus this site studies, 126 titles pair “AI” with a word beginning “trad”, from 95 channels, at a median of 13,652 views, and 21 of them passed 100,000. The counts are published as a CSV of AI-trading title counts, with the exact title rule for each row.
By tool name, 58 titles mention Claude at a median of 26,995 views, and 37 mention ChatGPT or GPT at a median of 4,246. Titles with “bot” or “bots” number 113, at a median of 9,476. The groups overlap, since a title can name a tool and a bot at once.
Claims in the titles
Of the 126 AI-trading titles, 18 carry a claim: a dollar figure, a percentage, “profitable,” “passive income,” “doubled,” “made more than” or “win rate.” Those 18 drew a median of 23,395 views. The other 108 drew a median of 11,257.5.
That gap is the incentive in one number. In this sample, titles that promise a result drew about twice the median views of those that do not. None of those 18 claims is checked here, and a view count says nothing about whether a claimed result happened.
On Michael’s own channel, the trading video with AI in its title in the 11 Aug 2026 lifetime export is “Best LuxAlgo AI Indicator Settings (2026 Update),” at 2,925 views. It is a settings walkthrough, and it carries no performance claim.
When it fails
The first failure is asking the tool to find the edge. A language model asked for a profitable strategy will return one that looks complete and has never been tested. The code runs; nothing about it shows an edge.
The second is a backtest with no out-of-sample test. A model tuned on five years of history will look excellent on those five years. Hold back a period it never saw, and judge it only there.
A third is trusting its numbers. Assistants can state a statistic, a rule or a date that is wrong with complete confidence. Every figure used in a decision has to be checked against the original source.
A fourth is the scam wearing AI’s clothes. The CFTC and FINRA alerts describe schemes where the AI is a label and the product is a deposit. Promised returns, fixed monthly percentages, pressure to recruit and difficulty withdrawing are the signs, whatever technology is named. The avoid a scam page walks through the checks.
And a fifth is automation without a stop. Connecting a model to a live account multiplies both its speed and its mistakes; the automated trading page covers the kill switch.
Related
Trading bots tests whether a rule found on one half of a history survives the other half, which is the test every AI-built strategy needs too. Automated trading covers what running a system live actually involves. And backtesting explains how to test an idea before any money is at stake.
Use an AI assistant for the work you can check: code you can read, a backtest you can rerun, a rule you can explain in a sentence. The moment it is being trusted for something you cannot check, such as where price goes next, it has stopped being a tool.
— Michael Whitman
This page is educational, not financial advice. Test every idea on your own charts before risking money.