AI and automated trading bots

The research was done by AI agents that open web pages. Some sites refuse them; where that happened we say so. The small numbers point to the source list at the end. Where a figure is our own sum or guess, it is marked “our estimate” and the basis is given.

The claim

These are examples of the claim, copied from our collection of 47 tweets about this way of making money¹. (We call every method we examine a “scheme”, meaning a plan; the word is not a finding about anyone quoted.) View counts are as collected on 6 October 2026. Dates are worked out from each tweet’s ID number. A slash marks a line break in the tweet. The page judges the scheme, not the people who tweeted.

“I built a simple strategy using ChatGPT and made $71,500 trading $BTC / You can build the same trading bot in just 10 minutes”

@rektfencer, 24 April 2024, 1,769,357 views¹

“Paste this prompt into Claude Code. It will build you an Opus 5.5 + Jev trading bot. It will probably make you a lot of money…”

@milesdeutscher, 30 September 2026, 390,761 views¹

“A 19-year-old Japanese student built a trading bot with JEV & Claude in 2 days. Used his iPad as a second monitor. First night: $6,732 profit. Starting capital: $68. Total profit so far: $750,000.”

@codewithimanshu, 27 September 2026, 104,780 views¹

“I wrote a Bot, that’s made me $35K from $100 last night. This isn’t clickbait—just an AI Trading Bot powered by DeepSeek.”

@piyascode9, 25 September 2026, 13,555 views¹

“Jev” is an AI tool named in several of the tweets; we did not examine it. Claude, ChatGPT and DeepSeek are AI models.

The 47 tweets were viewed about 5.3 million times in total. Three accounts (rektfencer, RohOnChain and milesdeutscher) hold about 4.7 million of those views, which is about 89% (our estimate from the file)¹.

What the scheme is

A trading bot is a program that buys and sells for you. You give it access to an account at a broker or a crypto exchange, and it places orders by rule, day and night, without you clicking anything.

The tweets describe four versions.

• Build your own. You paste a ready-made prompt into an AI coding tool such as Claude Code or ChatGPT. The tool writes the program. Some versions also ask an AI model to decide each trade while the bot runs. This is the version with nearly all the views¹.
• Buy one. You pay for a finished bot. Two tweets offer an “EA” for $69.99¹. An EA, or “expert advisor”, is a bot that runs inside MetaTrader, a trading program used with foreign-exchange (forex) brokers. Other tweets link to a bot site, a Telegram bot, or a paid tool¹.
• Pay someone to run theirs on your account. One tweet promotes a monthly service whose seller runs its bots on the customer’s own account, which the tweet says stays under the customer’s control¹.
• Hand over the money. You buy into a pool or a token whose seller says a bot will trade for you. Two of the 47 tweets promote a token that comes with access to a bot¹.

Most of the crypto tutorials trade “perpetuals”. A perpetual is a bet on a coin’s price that never expires and is usually made with leverage. Leverage means trading with borrowed money: at 10 times leverage, a 1% price move changes your stake by 10%. If the price moves far enough against you, the exchange closes the position and keeps what is left of the stake. That is called liquidation.

The mechanism is real. Exchanges and some brokers publish tools for automated trading, and writing a bot with an AI assistant is now cheap. The question is whether the bot makes more than it costs, and for whom.

The arithmetic

Trading income is the money in the account, times the return per month, minus fees and running costs. A claim can be checked only if it gives all of these.

Step 1: what the tweets give you

We looked at the text of the 30 most-viewed tweets¹. All counts in this step are our estimate from reading the file. We did not see the images or videos attached to the tweets, which may show more.

• 17 of the 30 give no money figure at all, only words such as “PROFITABLE”, “prints” or “it will probably make you a lot of money”. Three more give only a price or a build cost. Ten state a profit figure.
• The text of the most-viewed tweet ($71,500 on bitcoin) gives no starting sum and no period; its attached image and thread were not viewed. The text of “Made $22,400 last month” gives a period but no starting sum.
• Five of the 30 give a usable starting sum and an end sum; one of them gives two end sums. A sixth says a bot started from nothing and reached $900K, which cannot be turned into a multiple.
• None of the 30 states in its text the fees paid, the worst losing stretch, or the leverage used.
• Several are stories about someone else, such as “a 19-year-old Japanese student” and “a Hong Kong marketer”.

Here is what those five tweets imply (our sums)¹.

The last row matters. That tweet describes its result as a simulated run on a DEX, meaning no real money was traded¹. (A DEX is an exchange that runs on a blockchain.)

For the top tweet, a crypto news site reported the day after that the poster “claims earning $71,500” from a strategy that ChatGPT wrote for the charting site TradingView⁷². The article gives no starting sum and no period, and does not say whether the figure was real profit or a test on past prices⁷². According to the same report, the thread also warned of the risks of trading and advised starting with small amounts⁷².

The US regulator for futures and derivatives, the CFTC, has published an advisory called “AI Won’t Turn Trading Bots into Money Machines”²³. It warns about offers of trading bots “that promise unreasonably high or guaranteed returns”, and the cases it describes involved money handed to someone else². The advisory does not name or concern anyone quoted on this page, and a large figure in a tweet is not evidence of fraud. We cite it only because it tells readers to treat such figures with caution.

Step 2: how much money the income needs

Turn the sum around. Capital needed = monthly income wanted, divided by the monthly return. The table is our estimate; the sources are for the return in each row.

The accounts in the tweets are $68 to $1,000¹. To earn $1,000 a month, an account that size has to at least double ($1,000 account), gain tenfold ($100) or about fifteenfold ($68) every month (our sums).

The fourth row is the point of comparison. About 8% a month is what the best 500 day traders in a whole country earned on the money they traded, before any living costs²⁰. The claims in Step 1 are tens to thousands of times above it. Side by side: a $1,000 account needs 100% a month to pay $1,000 a month, and Taiwan’s best 500 made about 8% a month (our sums)¹²⁰.

Step 3: what measured returns look like

Individuals who trade by algorithm. We found one study that measures them directly. It uses the records of every individual investor on India’s National Stock Exchange from 2012 to 2019, where orders placed by a program are tagged; the sample starts from 19 million investors⁷⁴. It cuts both ways.

• People’s algorithm trades did better than their own hand trades. Over ten days, hand trades trailed the market by 0.30% before these investors took up algorithms and 0.26% after; their algorithm trades trailed it by 0.15%⁷⁴. The authors say the gap “amounts to an annual return of 3.75% and 2.75%”⁷⁴.
• The algorithm trades were still below the market, on average⁷⁴.
• “unprofitable algorithmic traders are more likely to quit than profitable traders”⁷⁴.

The limits are large. It is a working paper, not yet published in a journal. It covers ordinary shares, not leveraged crypto. The algorithms were rule-based programs from before today’s AI models. We did not check how it treats trading costs. An earlier version of this page said that for AI-built bots or crypto perpetuals we found no such study. One has since been found, and it comes next.

AI agents run by ordinary users. A September 2026 research paper studied two groups (“fleets”) of AI trading agents that users funded themselves: 3,505 user-funded pools trading meme coins for 21 days, and 500 to 599 user-created agents trading perpetuals on Hyperliquid from June to August 2026⁹⁷. Its finding: “neither fleet shows a directional edge. The DXAP fleet is not profitable and trails a matched Hyperliquid retail benchmark (41% vs. 50% roundtrip win rate).”⁹⁷ In plain words, the agents on perpetuals were not profitable as a group, and won a smaller share of their trades than a matched group of ordinary Hyperliquid traders. The paper also reports that the risk setting each user chose shaped how the agents traded, meaning their leverage and liquidations, more than the strategy prompt did, and that in a replay of 416 recorded situations the decisions of leading AI models were “statistically indistinguishable” in quality⁹⁷. The limits: it has not been through journal review, we read only its abstract, and its authors appear, from the abstract, to be connected with the platforms studied⁹⁷.

For crypto perpetuals the nearest figure is weaker. A May 2026 analysis, which we read only as a repost on an exchange’s news page, says: “About 75% of addresses on Hyperliquid are in a loss-making position.”⁷⁵ It gives no sample size or method. It adds that one of the most profitable addresses made 261,000 trades in a month, which only a program can do⁷⁵. An address in profit on one day is not a person earning a living, and such addresses may belong to firms.

Hyperliquid itself publishes a leaderboard file, which listed 47,219 accounts on 7 October 2026⁸². Of the 18,987 that traded in the last month, 9,973, or 52.5%, show a profit for the month; of 44,095 accounts counted over their whole life, 48.0% show a profit (our counts)⁸². That looks better than the 75% above, but the two do not cover the same accounts. The file’s rule for inclusion is not documented, and it leans to large and very active accounts, including firms⁸². Neither figure separates bots from people.

Public trading pools on Hyperliquid. A “vault” is a pool that one operator trades, by hand or by program, and that outsiders can deposit into. Its results are public, so this is the nearest thing to a checkable track record for bot operators. We have two counts, both worked out on 7 October 2026.

• From the exchange’s own file: of 9,468 user-created pools, 6,385 (67%) are closed. 2,155 (22.8%) show a lifetime profit and 5,876 show a loss⁸³. Open pools hold about $98.6 million, 64% of it in the ten largest⁸³. Of 8,402 pools more than a year old, 275 are still open with at least $1,000 in them, and 117 of those are in profit⁸³. These are our counts, in dollars, and a pool may have closed for reasons other than losses.
• From an analytics site that adjusts for deposits and withdrawals: of 605 trader-run pools, 32.7% have a positive lifetime return and the median is a loss of 18.0%⁷⁹. Of the 218 that are at least a year old, 36.7%, about 80 pools, are in profit; the median is a loss of 30.7%; and 46 of them, 21%, lost 90% or more⁷⁹. The percentiles are our estimate from the site’s data file.

Three limits apply. A pool is not always a bot. The analytics site’s file appears to include only pools that at some point held about $5,000 or more, so the pools in the second count are not the $68 to $1,000 accounts in the tweets, and the share in profit among all pools is probably lower, as the first count suggests (our judgement)¹⁷⁹⁸³. And the returns are not compared with simply holding bitcoin.

An earlier draft of this page said every operator counted had paid a 10,000 USDC fee to open a pool. (USDC is a coin tied to the dollar.) That was wrong. Opening a pool costs 10,000 USDC today, but archived copies of the same Hyperliquid page from April 2025, June 2025 and February 2026 say 100 USDC, so pools at least a year old opened under the lower fee⁸¹. The winning side of these pools is described under “The upside”.

Bots on a prediction market. Some of the collected tweets and one freelance listing concern bots for Polymarket, a site where people bet on the outcome of events¹⁶⁹. A data firm counted all 2,347,464 wallets active there from September 2025 to mid-August 2026. About 1 wallet in 200 behaved like a machine, and those placed more than 6 in 10 trades⁹⁹. In a nine-day window, 11 to 19 August 2026, the 3,362 machine-run wallets as a group lost $2,822,122 after fees, and 34.4% of them finished ahead, against 37.8% of the 84,417 other wallets⁹⁹. The firm says machines are over-represented among both the big winners and the big losers⁹⁹. In a separate one-week count of Polymarket’s 15-minute crypto markets, 36.8% of 46,945 wallets finished the week in profit and the median wallet lost about $3¹⁰⁰. Both firms sell market data.

AI models trading real money. The best-known public test is Alpha Arena, run by a company called Nof1. In its first season six leading AI models each got $10,000 to trade crypto perpetuals on the exchange Hyperliquid³². It ran from 17 or 18 October (the reports differ) to 3 November 2025, about 17 days³⁰³². Four of the six lost money³⁰³¹. One news report gives the final balances from $12,231 for the best model, Qwen3 Max, down to $4,126 for GPT-5³⁰. A second report differs slightly³¹. Neither is the organiser’s own page, which did not show results to our agents. Both agree on the shape: the best model gained about 22%, and another with the same instructions lost more than half.

Three days into the contest the table looked different. Grok, DeepSeek and Claude Sonnet 4.5 were each up more than 25%³². Grok and Claude finished among the losers³⁰. A screenshot from a good day is not a result.

In a follow-up two-week season on US stocks, with four separate contests, the organiser’s site reports the winner at a 12.11% return³³. A news report quotes the organiser saying the winner “made money in all four competitions”, and adds that five days later it was the only model still in profit³³. It does not list the other models’ season results. The organiser’s site puts the winner’s gain at $4,844 in total across the four contests³³. A blog, which we did not check against the organiser, says that across the eight models and four contests of that season a model finished in profit in only 6 of 32 results¹⁰⁶. The same blog gives the first season’s result for GPT-5 as a loss of 62.66%, which does not match the $4,126 balance above, so the reports still disagree³⁰¹⁰⁶.

Two academic projects have also run AI models against live prices. One ran 21 models for 50 days and reports that “high LMArena scores do not imply superior trading outcomes”, meaning a model that ranks well in general did not trade better⁷⁶. The other reports that the design of the bot mattered more than which model was inside it⁷⁶. Neither abstract gives a profit figure or says real money was used⁷⁶. A third project, StockBench, ran AI agents over several months of real 2025 stock prices in a simulation. Its abstract says: “most models struggle to outperform the simple buy-and-hold baseline, while some models demonstrate the potential to achieve higher returns and stronger risk management”⁹⁶. Buy-and-hold means buying once and doing nothing.

These tests used 2025 models, not the newer ones named in the tweets. And a few weeks with one run per model cannot separate skill from luck.

Professional funds that trade by algorithm. One industry index of such funds, the Barclay Systematic Traders Index, returned -0.81% in 2023 and +1.66% in 2025, as relayed by a trade news site⁴⁴. We did not obtain the 2024 figure or a long-run average.

Individuals who trade short-term, mostly by hand.

• Brazil: among everyone who began day-trading futures in 2013 to 2015 and kept at it for at least 300 days, “97% of them lost money, only 0.4% earned more than a bank teller (US$54 per day), and the top individual earned only US$310 per day with great risk (a standard deviation of US$2,560)”¹⁹. That is roughly $6,500 a month for the single best person (our estimate: 310 x about 21 trading days), with daily swings eight times the average day’s gain.
• India: the securities regulator SEBI counted every individual trading share futures and options (“F&O”). Its headline: “93% of Individual Traders Incurred Losses in Equity F&O between FY22 and FY24”²¹. FY means financial year. Summaries of the study say there were over 10 million such traders over the three years, and that the losers’ average loss over the three years was about Rs 2 lakh²²²³. A lakh is 100,000, so that is roughly $2,400 at Rs 83 to the dollar (our estimate). “Only 1% of individual traders managed to earn profits exceeding Rs1 lakh, after adjusting for transaction costs”²². Later studies are reported to show about 91% losing in the year to March 2025²⁵ and about 88% the year after²⁶. A summary of that last study gives 87.7%, which leaves about 12.3% who did not lose (our sum)⁹⁵.
• Europe: the EU markets regulator found that on CFDs “74-89% of retail accounts typically lose money on their investments, with average losses per client ranging from €1,600 to €29,000”²⁸. A CFD (“contract for difference”) is a leveraged bet on a price, and it is the product most forex EAs trade. One large broker’s current disclosure is 70%²⁹.
• Meme coins: for the meme-coin version described in one tweet, a January 2025 report of blockchain data for the token site pump.fun found that 55,012 of 13.4 million wallets, about 0.4%, had realised $10,000 or more in profit⁴². The count leaves out coins bought after they leave the site’s launch stage. The site’s co-founder said the true number is “an order of magnitude LARGER”, which he did not document in the article⁴². About 30% of the wallets made a single sale and were probably not people⁴².

None of these four separates bot users from other traders.

Step 4: fees

Every trade pays a fee, win or lose. A bot can trade far more often than a person.

We use Hyperliquid’s fees as the example; we did not establish which exchange any tutorial uses. Its starting fee on perpetuals is 0.045% for an order filled at once (a “taker” order) and 0.015% for an order that waits in the queue (a “maker” order)⁸. Opening and closing a position, a “round trip”, with taker orders costs 0.09% of its size (our sum: 2 x 0.045%). On Binance an ordinary user pays 0.1% per side, 0.2% a round trip, on spot trades, meaning buying the coin itself without leverage¹⁴. On Binance’s futures an ordinary user pays 0.02% for a maker order and 0.05% for a taker order, so a taker round trip costs 0.10% (our sum), slightly above Hyperliquid’s 0.09%¹⁰². An earlier version of this page had not confirmed that rate.

What that does to an account depends on how often and how much the bot trades. This table is our estimate from Hyperliquid’s fees⁸ and a 30-day month. The trade counts are our assumption, not a measurement of any bot.

A slow bot is not sunk by fees. A “high-frequency” bot at 20 round trips a day has to earn about half the account every month before it has made anything. Leveraged positions also pay funding charges (regular payments between the two sides of a perpetual), so the table understates the cost.

An earlier version of this page said we had not sourced funding. On Hyperliquid it is paid every hour¹⁰¹. Its fixed part is set out in the exchange’s documents: “interest rate component is predetermined at 0.01% every 8 hours, which is 0.00125% every hour, or 11.6% APR paid to short.”¹⁰¹ APR means the rate per year. A “long” bets on the price rising and a “short” on it falling. So someone holding a long pays about 11.6% a year of the position’s size before any extra premium, and in extremes funding can reach the cap of 4% an hour¹⁰¹. At 10 times leverage the fixed part alone is about 116% of the stake a year, or about 9.7% a month (our estimate: 11.6% x 10 / 12)¹⁰¹.

Large traders pay less. Hyperliquid’s fees fall with the amount traded over the last 14 days. Above $5 million the taker and maker rates are 0.040% and 0.012%; above $500 million, 0.028% and nothing; above $7 billion, 0.024% and nothing⁸. The largest makers are paid a rebate of up to 0.003% (this row was read through a summarising tool)⁸. On Binance futures the lowest rates, nothing for maker orders and 0.017% for taker orders, need $25 billion of trading in 30 days¹⁰². So the largest firms on the other side of a newcomer’s trade pay about half the taker fee and nothing to leave an order waiting (our reading).

Two people who published real numbers found fees decisive. A developer’s open-source bitcoin bot predicted price direction better than chance and still turned 1 BTC into 0.955 BTC in two months, after 23,413 trades and 2.486 BTC paid in fees³⁸. His words: “I was never able to generate a robot that traded profitably after fees.”³⁸ A 2026 write-up of a bot run on about $40 put fees at an “estimated 20–30% of the account per month” at about 10 trades a day³⁹.

Step 5: running costs

A token is a piece of a word; it is the unit AI models charge by. The server price was read through a summarising tool¹⁰³.

Building the bot is cheap. One tweet says a bot was built for $6.80 using 1.45 million tokens¹. That is about $4.70 per million tokens (our sum: 6.80 / 1.45), which sits between the $4 and $20 prices above.

Running it costs anything from almost nothing to a lot. Suppose the bot asks a model for a decision, reading 5,000 tokens and writing 500 each time. The sizes and call rates are our assumption; the sums are our estimate from¹⁷.

Put Steps 4 and 5 together for a $1,000 account (our estimate). A slow bot on a free framework with the cheap model once an hour costs under 1% a month: about 0.5% for the model plus about 0.3% in fees. A busy bot on a $50 platform plan with five taker round trips a day needs about 18.5% a month just to stand still (5% plus 13.5%), and about 44.5% if it also calls Opus every five minutes. Cost does not stop the slow bot. Its problem is that nothing we found shows it beats simply holding the asset.

What people who tried it report

No verified record of a newcomer earning a living from a bot was found. We searched for a payment-verified, year-long income report from an individual running an AI trading bot. The search used only a few wordings, so we cannot say none exists. A second pass on 7 October 2026 searched in more wordings and found only vendors and books. It did find verified records of automated and bot-like accounts in profit. They belong to anonymous wallets, firms and one company-run pool, not to named newcomers, and they are set out under “The upside”.

In our own collection, no claim gives a starting sum, a period and a statement in its text. None of the 47 tweets links in its text to a broker statement or a public wallet, and one headline figure is a simulation¹. Most carry an image, video or link that we did not open, so proof could be there.

One tweet attributes a result to a professional trader. The trader is a well-known professional stock trader, whom we do not name because we could not check the claim. The tweet says a clip shows him explaining how his AI trading bot went from nothing to $900K¹. We did not open the clip and did not confirm the figure. If it holds, it is on the tweet’s own account a result for an experienced trader, not a newcomer.

First-person accounts with numbers are losses, and the cause is fees. These are the two write-ups in Step 4³⁸³⁹. A third builder, running six AI bots around the clock, wrote in April 2026: “Trading bot still isn’t profitable.” He gave no figures⁴⁰. None of the three pages offers a trading product for sale, though one carries affiliate links³⁸³⁹⁴⁰.

Tests on past prices do not predict live results. A backtest runs a strategy over old price data to see what it would have made. Quantopian, a platform where hobbyists wrote trading programs, compared backtests with later live results for 888 of them. It found that common backtest scores “offer little value in predicting out of sample performance (R² < 0.025)”³⁷. In plain words, a good backtest told you almost nothing. We read this through a summary site³⁷. A separate study re-tested published AI trading strategies over two decades and more than 100 stocks and found that “previously reported LLM advantages deteriorate significantly under broader cross-section and over a longer-term evaluation”³⁵. LLM means large language model, the kind of AI behind ChatGPT and Claude.

A good backtest is also easy to produce by accident. A 2014 article in a mathematics journal worked out how many versions of a strategy you need to try before one looks good by chance: “After trying only seven independent strategy configurations, the expected maximum SR IS is 1 for a two-year long backtest, while the expected SR OOS is 0.”¹⁰⁴ SR is the Sharpe ratio, a score of return against risk; IS means on the data used for testing and OOS means on new data. With five years of data the number is 45 versions¹⁰⁴. An AI coding tool can write and test dozens of versions in an afternoon (our observation), so keeping the best of many tells you little.

Quantopian’s own history points the same way. According to Wikipedia, the platform had over 210,000 members by August 2018 and in 2016 an investor committed up to $250 million to its members’ best algorithms; in February 2020 it “announced it would return investors’ money due to the underperformance of its investment strategies”, and its shutdown was announced in November 2020¹⁰⁵.

People who bought into an “AI agent” mostly lost. A May 2026 research paper measured 11 crypto projects that sell tokens tied to an AI “investment agent”, and 925,323 holders of those tokens. It found the projects’ own funds “retain over $30M in paper gains while their token holders collectively lost $191.7M, with the top 1% of wallets capturing 81.4% of all gains”³⁴. The paper has not been through journal review, and one author works at a crypto investment firm³⁴.

Copying someone else’s trades. A crypto news site reports a study finding that “Just 6.16% of the 292,000 Fomo wallets studied by DWF Ventures finished three months in profit” on one copy-trading app⁴³. We did not find the study itself.

Forex bots on “funded accounts”. Some EA sellers point buyers to prop firms. A prop firm sells a paid test, and those who pass trade the firm’s money for a share of profit. A company that supplies software to these firms says 7% of 300,000 accounts ever received a payout⁴¹. This is an industry statement that mixes bot users with people trading by hand, so it is only loosely related.

A seller’s own figure. AlgoTest, which sells automated-trading tools in India, says 45% of its customers were profitable in the 2023-24 financial year. It adds that this is “on a gross basis before brokerage, taxes, and charges. Also, it’s only for the positions booked through AlgoTest.”²⁷ No sample size is given. Bitsgap, which sells trading bots, reported 81,456 bots launched and 12.6 million trades in the first half of 2026; its report gives no share of bots that made money¹⁰⁷.

What this does and does not show. In one national stock market, automation appears to have reduced individuals’ shortfall compared with trading by hand, without lifting the average user above the market⁷⁴. Fast bots add fees³⁸³⁹. An earlier version of this page said nothing here shows that bots lose more than people do. Two newer sources point that way for particular groups: users’ AI agents on perpetuals won a smaller share of their trades than a matched group of ordinary Hyperliquid traders⁹⁷, and machine-run wallets on Polymarket were slightly less likely than other wallets to finish a nine-day window ahead⁹⁹. On the other side, a few hundred bot-like accounts are the documented big winners on that same market⁸⁴⁸⁵. Nothing shows a newcomer’s bot producing an income.

What the rules allow now

What is allowed

Running a bot on your own account is allowed at the venues we checked. A bot connects through an API key, a password that lets software act on your account.

• Robinhood (a US broker) now has a clause on AI agents in its customer agreement, revised 1 October 2026⁴⁸. One collected tweet describes running an AI agent on a Robinhood account¹.
• Binance.US documents API keys for third-party trading software, and Hyperliquid documents fees that bot builders can attach to orders⁵²⁵⁴.
• Small US stock accounts. The old rule that needed $25,000 to day-trade stocks was replaced in 2026. A compliance firm reports the change took effect on 4 June 2026, with brokers given until October 2027 to switch⁴⁹. Instead, “FINRA will require firms to monitor intraday margin exposure”⁴⁹. Margin is money borrowed from the broker. Under the proposal as published, a customer who “makes a practice” of not meeting shortfalls promptly, and leaves one unpaid after five business days, cannot borrow more for 90 days or until it is paid; shortfalls up to the lesser of 5% of the account’s equity or $1,000 do not count⁵⁰. We read the proposal, not the final rule text.

This makes it easier to start with little money. It does not change the odds.

Who carries the loss when the bot gets it wrong

You do. Robinhood’s agreement says:

“You acknowledge that AI agents may misinterpret your prompts, generate erroneous, fabricated, or unintended outputs … any instruction submitted to Robinhood from an authenticated session of your AI agent shall be final and binding upon you upon receipt by Robinhood, and Robinhood shall have no liability for any Losses arising out of or relating to any such error, malfunction, or unintended behavior of your AI agent.”⁴⁸

The same section makes the account holder responsible if the bot’s orders amount to market manipulation, “regardless of whether such conduct was intended by you”⁴⁸. Robinhood may also “terminate your account or disable your API/MCP connectivity for any reason” (MCP is one way of connecting an AI tool to an account), and says orders sent this way “may result in adverse execution quality and/or speed”⁴⁸. The access is “solely for your own personal use”. Using the programming interface directly, or developing a product on it, needs “Robinhood’s express written consent”⁴⁸.

What is restricted

• Hyperliquid from the United States or Ontario. Its terms bar “persons or entities who reside in, are located in, are incorporated in, or have a registered office in the United States of America or Ontario, Canada”⁵¹. They also prohibit “using technologies such as VPNs, proxies, or other methods to conceal your location”⁵¹. The terms say the operator may suspend or end a person’s use it finds “unauthorized, deceptive, fraudulent, improper or unlawful”⁵¹. So if a tutorial trades there, a US reader who follows it would be breaking the venue’s terms.
• Crypto derivatives in the UK. In 2020 the UK regulator, the FCA, published “final rules banning the sale of derivatives and exchange traded notes (ETNs) that reference certain types of cryptoassets to retail consumers”, in force from 6 January 2021⁵⁹. An ETN is a listed security that tracks a price. In 2025 the FCA reopened ETNs to ordinary customers from 8 October 2025 and said: “The FCA’s ban on retail access to cryptoasset derivatives will remain in place.”⁷⁸ Perpetuals are derivatives.
• Leverage in the EU. The EU’s 2018 measures capped leverage for ordinary customers on CFDs, from 30:1 on major currencies down to 2:1 on crypto²⁸. We did not re-check that the caps are unchanged in 2026.
• Using bots in exchange promotions. A news report quotes Binance: “over 600 accounts were banned last week for using unauthorized third-party tools” in October 2025⁵³. Binance said accounts breaking the rules may face “forfeiture of any profits earned in Alpha events”⁵³. This concerned Binance’s promotional events, not ordinary automated trading.
• Selling a bot service to others. In the US, trading other people’s money or selling futures or forex signals to the public generally needs registration with the CFTC⁶¹. We have this only from a law guide. A joint alert from three US regulators tells investors to check registration, and adds: “Even when based on accurate input, information resulting from AI can be faulty, or even completely made up.”⁶

We found no rule made specifically for AI trading agents used by ordinary customers. A news report, seen only as a search snippet, says US lawmakers wrote to the securities regulator about it in June 2026⁶².

Ways to lose money other than bad trades

1. Liquidation. On 10 October 2025, $19.3 billion of leveraged crypto positions were force-closed in one day, according to a research firm⁵⁸.
2. Handing over a key. Binance.US says: “API keys allow third-party platforms or entities to access your account and even perform actions - such as trades and withdrawals - on your behalf.”⁵⁴ In December 2022 the head of the bot platform 3Commas confirmed that its users’ keys had leaked: “We saw the hacker’s message and can confirm that the data in the files is true.”⁵⁵ The company asked exchanges to “revoke all the keys that were connected to 3Commas”⁵⁵. A security firm’s write-up puts the theft at “an estimated $20 million”, a figure it takes from a news report on an FBI investigation that we did not open⁵⁵.
3. Bots that act on your wallet. In September 2024, $3 million was taken from the wallets of 11 users of one Telegram trading bot, Banana Gun. The team said “we identified a potential vulnerability in the Telegram message oracle we use”, and refunded all 11 from its own treasury⁵⁶.
4. Bot code that steals. A security firm found a trading bot on the code site GitHub that took users’ wallet keys. It said the attacker used a batch of accounts “distributing malware while artificially inflating fork and star counts”⁵⁷. Stars are GitHub’s “like” button. One collected tweet cites a bot’s star count¹. We do not suggest that bot is unsafe; the point is that a star count is not evidence either way.
5. Pools that do not trade. See the regulator cases in the next section.
6. Tax paperwork. In the US, the tax office says: “If you exchange virtual currency held as a capital asset for other property, including for goods or for another virtual currency, you will recognize a capital gain or loss.”⁶⁰ A bot that makes hundreds of trades creates hundreds of entries to report.

Who makes money from it

The exchange, on every trade. Fees are charged whether the trade wins or loses⁸¹⁴.

Whoever referred you, on every trade. Hyperliquid’s documents say: “You will receive 10% of referred users’ fees, less any fee discount they receive.”⁵² Someone who distributes a bot can also add a fee of their own to each order it sends: “Builder fees charged can be at most 0.1% on perps and 1% on spot”⁵². “Perps” are perpetuals. Bybit, another exchange, says it “shares a portion of the trading fee profit generated from any eligible referral activity with its affiliates”, paid daily⁶⁶. An affiliate is someone paid for sending customers. In each case the referrer’s income depends on how much the newcomer trades, not on whether the newcomer wins.

Tool makers, every month. The platforms in Step 5 charge a subscription regardless of results¹⁵¹⁸. 3Commas says of itself: “3Commas provides software only.”¹⁵ Cryptohopper pays affiliates “up to 15% of the subscription costs as a commission”¹⁸. The MetaTrader Market, a shop for EAs, keeps a cut of each sale: “The Market service commission of 20% is charged on purchases.”⁶⁷

Operators of a pool or strategy that others pay into. Hyperliquid’s documents say: “Vault leaders receive a 10% profit share for managing the vault.”⁸¹ Darwinex pays a trader 15% of the profits of investors who follow the strategy⁸⁸. Managers on Collective2 charge subscribers $25 to $249 a month⁹⁰. The first two are paid only when followers profit; a subscription is paid either way.

Venues pay some bots directly. Polymarket’s documents describe daily rewards for accounts that keep orders waiting in its queue: “Rewards are distributed directly to maker addresses daily at midnight UTC.”⁹⁴ How the money is split is not published.

People with an audience. Here is what the 47 tweets steer readers to. The counts are our estimate from the collector’s labels and the tweet text; we did not open the posts on X¹.

The seven most-viewed tweets are all free how-to posts, from four accounts¹. For one of those accounts we followed the links. For the other three we found nothing for sale, which may be a gap in our search.

• A tweet by milesdeutscher links to his YouTube video, “Jev + Claude Opus 5.5 = The Most Powerful AI Trading Bot Ever”¹⁶³. The video’s description carries a link to his free community, a Bybit link advertising “$30K+ in welcome rewards”, and discount links for two data tools, a hosting company and a dictation app⁶³. We do not know whether each link pays him. The description also carries a risk warning, that trading can “involve significant risks, including the potential loss of your entire capital”, and a link to a disclosures page that we did not read⁶³.
• The community is free: about 2,300 members on the site Skool, with no paid tier on its public page⁶⁴. Skool charges the owner, not the members ($9 or $99 a month)⁶⁴.
• A separate paid crypto research community on the site Whop, “Miles High Club”, is listed at $199 a month (shown as reduced from $258.44) with 223 ratings⁶⁵. Its listing, re-read on 7 October 2026, introduces the seller with the words “My name is Miles Deutscher and I’m a crypto analyst and investor”⁶⁵. The listing describes trade ideas, research and mentoring, not a bot⁶⁵.

None of the pages we read for this account offers a bot for sale.

Tweets with near-identical wording. Eleven of the 47 tweets, from five small accounts, use near-identical wording, saying that a bot placed every trade without the poster doing anything by hand¹. Our agents followed their links and report that all lead to the same site⁷⁰. The site refused our agents by both routes, so we do not know what it sells or who runs it, and we draw no conclusion about it or about the accounts⁷⁰. Another tweet leads to a Telegram trading bot through a referral code¹. A tutorial for a bot of the same name says it “charges a standard transaction fee on every successful buy and sell order”; we did not confirm it is the same bot⁷¹.

Sellers of a service.

• The “HALO” service in one tweet runs the seller’s EAs on the customer’s own account. Its page says the company “does not accept or manage client funds”⁶⁸. It is sold at “Starting $800+ month (All EAs Included)” and advertises monthly averages of 6% to 15%⁶⁸. Those returns are the seller’s claim; we did not verify them. The page warns: “You may sustain losses equal to or greater than your initial investment.”⁶⁸ The page, re-read on 7 October 2026, gives two entry levels: “$5K minimum” under equity requirements, and “investment levels starting as low as $10,000”⁶⁸. At $800 a month, the fee is 8% of a $10,000 account each month and 16% of a $5,000 account (our sums), so on the smaller account the fee alone is above the top of the advertised 6% to 15% range. The page’s own small print says its performance examples “are hypothetical or historical in nature and do not represent actual trading results”⁶⁸.
• A custom-build offer in another tweet leads to a freelance listing priced at $750, with a 5.0 rating from 35 reviews⁶⁹. The builder is paid a fixed fee and the listing makes no profit promise⁶⁹.

What regulators have found when someone else’s bot held the money. These cases are about pooled money and paid programmes, where money left the customer’s own account. None is about a person building their own bot from a free prompt, and none is about a service run on the customer’s own account. These cases concern the companies named in them only. We found no regulator action against any account quoted or named on this page, and we do not suggest any connection.

• Mirror Trading International. According to the CFTC, for “as little as $100 in bitcoin” customers could buy into its operator’s pool, which “used a proprietary bot trading program that guaranteed at least a 10 percent monthly return”². According to the CFTC’s release, the court’s order finds that the operator ran “an international fraudulent multilevel marketing scheme”, which took bitcoin from at least 23,000 people in the US⁴. A US federal court entered a default judgment (the defendant did not contest) against the operator, Cornelius Steynberg, ordering $1,733,838,372 in restitution and a penalty of the same size⁴. The CFTC says the defendants “misappropriated all of the Bitcoin they accepted from pool participants”; its case against the company continued at that date⁴.
• In May 2026 the US securities regulator, the SEC, sued a Texas man. It alleges he raised about $12.3 million from about 150 investors for AI trading bots, and that his “bots did not function as represented”; its release describes about $5.5 million as “Ponzi-like payments”, meaning earlier investors paid with later investors’ money⁷. These are allegations as of the release of 29 May 2026, not a judgment; the case was pending at that date and we have not checked it since.
• The SEC charged a now-defunct company, 5 Fruits Enterprises, and its two operators with what it called “a scheme to defraud investors through false claims about 5 Fruits’ use of automated ‘bots’ in trading securities”, alleging they raised over $4.7 million from more than 140 investors⁹. The defendants agreed to final judgments “Without admitting or denying the allegations”, subject to court approval; we did not check whether the court approved them⁹.
• On 29 September 2026 the SEC filed two complaints over what it alleges were two investment frauds that took at least $15 million. Of the smaller one, TSAI, about $2.8 million, it alleges “there were no AI trading bots and deposited funds never were used to earn returns”¹⁰. These are allegations in complaints filed that day; the cases are pending.

Other cases follow the same lines: a forex pool whose operators, the CFTC alleged in a 2023 complaint, “falsely assured pool participants they could recoup the losses using artificial intelligence-based (AI) trading algorithms”⁵, and investment clubs that, the SEC alleged in December 2025, took at least $14 million while “no trading took place on the trading platforms, which were fake”¹¹. We did not check the outcome of either case. The US consumer regulator, the FTC, acted against a seller of Amazon and crypto money-making programmes, one of them pitched as a “#1 secret passive income crypto trading bot”; the proposed settlement of the whole case provided at least $2.6 million for refunds¹². On a paid day-trading course, the FTC alleged that “most customer accounts actually lost money”; the company settled the charges and agreed to pay $3 million for refunds¹³.

The upside

Automated trading is not a myth, and some people and many firms make money from it. This section sets out the best outcomes we could document. Each one says who checked it. “Chain data” means a third party worked the result out from public blockchain records. “Platform-recorded” means a broker or marketplace recorded the trades. “Own statement” means only the person or company says so.

Firms. Virtu Financial, a trading firm, told the SEC in its 2014 filing: “we have had only one losing trading day since January 1, 2008.”⁴⁵ It said “we lease co-location space at or near … the exchanges”, meaning its computers sit beside the exchange’s own⁴⁵. In India, summaries of the regulator’s study say “97% of FPI profits and 96% of proprietary trader profits” came from algorithmic trading in the year to March 2024, while individuals as a group lost²²²⁴. (FPI means foreign funds; proprietary traders are firms trading their own money.) One summary says 13% of the 9.57 million individual traders that year were counted as using an algorithm, a share it says is overstated because it includes positions that brokers closed by program²⁴. A newcomer’s bot trades in the same market as these firms.

Even among firms the gains sit with a few. A summary of the Indian regulator’s study for the year to March 2026 says proprietary traders and foreign funds made Rs 58,379 crore before costs, 99% of it by algorithm, and that the ten largest proprietary desks took 74.5% of proprietary profits⁹⁵. (A crore is 10 million.) On one Hyperliquid venue over 21 days in March 2026, a study by a market-making firm found that 363 wallets, 0.46% of active addresses, did 63% of all trading, and that 21 of them did 90% of that group’s volume⁹⁸.

Bot-like accounts on a prediction market (journalism and academic work, both from chain data). This is the strongest evidence that bots as a class win somewhere. Bloomberg News analysed every Polymarket wallet active since the start of 2025. The part we could read says more than 100,000 accounts lost at least $1,000, “almost twice the number that made at least that much”, and that “a majority of the profits were raked in by a tiny slice of what look to be automated bots”⁸⁴. A summary of the paywalled part says about 800 bot-like wallets cleared more than $100,000 each and took roughly $131 million together⁸⁵. An academic paper measured one bot method on the same market from April 2024 to April 2025: arbitrage, which means profiting from prices that do not add up. It found about $39.6 million taken in the year. The top account made $2,009,631.76 in 4,049 transactions and the tenth made $383,569.94⁸⁶. Nobody is identified, the owners may be firms, and we found nothing sold by them. Arbitrage of this kind rewards speed and capital, not price guessing (our reading)⁸⁶.

Public pools on Hyperliquid (chain data, our percentiles). Step 3 gives the base rate for the 218 pools at least a year old: 36.7% in profit, median down 30.7%⁷⁹. The winning side of the same group looks like this. The top quarter gained 20.7% or more over their life, which runs from one to 3.4 years, and the top tenth gained 77% or more⁷⁹. 18 of the 218, about 8%, at least doubled⁷⁹. Per year, 30% of them beat 10% a year and 8% beat 50% a year⁷⁹. Across all 605 pools of any age, the top tenth gained 45% or more and the top hundredth 265% or more⁷⁹. In the exchange’s own file of all 9,468 user-run pools, 276 show more than $10,000 of lifetime profit and 70 show more than $100,000 (our counts)⁸³. These gains are not compared with simply holding bitcoin over the same period, which included a strong market.

Named pools with long records, all from chain data:

• Growi HF, the largest user-run pool. Its description says: “Employs a quantitative, mean-reversion strategy, trades all available cryptos. Fully automated with advanced risk control”⁸⁰. It started on 10 July 2024. Over 2.2 years it has returned about 99% after fees, about 36% a year, and about 23.5% in the last year, with a worst fall of 32.4% on the way⁸⁰. About $13.4 million is deposited in it⁸⁰. These figures are as read on 7 October 2026 and move daily. The exchange’s file shows about $2.92 million of lifetime profit for the pool (our reading of the file)⁸³. Its operator takes a 10% share of depositors’ profits, so it earns from other people’s money as well as its own⁸⁰⁸¹. Who runs it was not checked at source. We name it as an example of a record, not as a recommendation.
• Three other pools older than a year and a half, given here under the names their operators chose: “Citadel” (2.9 years, +526%), “Delta_01” (2.1 years, +418%) and “Hyperrr” (1.8 years, +1,110%)⁷⁹. We know of no link between the pool called “Citadel” and the US firms of that name. We did not check whether these are automated or who runs them.

The pools at the top of the site’s list show the other side of a big number. One, BredoStrategy, showed about +3,900% a year after fees when its last three months are scaled up, but about +21% after fees over its eight-month life, with a worst fall of 67% (file read on 7 October 2026)⁷⁹. An earlier draft gave +5,129% and +45%; the pool’s figures move fast and the file no longer shows those. Its own description says: “last time I was lucky, this time I’ll lose everything”⁷⁹.

Large accounts on Hyperliquid (the exchange’s file, our counts). In the leaderboard file of 47,219 accounts, 11,484 show a lifetime profit above $100,000, 2,778 above $1 million and 360 above $10 million; 883 show a lifetime loss above $1 million⁸². In the last month, 834 accounts were up more than $100,000 and 677 were down more than $100,000, and the top 1% of gainers held 43% of all gains⁸². The file is a selected list that leans to large accounts and includes firms, market makers and the exchange’s own pool, and it does not mark bots⁸². Heavy activity alone is not the mark of a winner: among the 7,115 accounts with more than $100 million of lifetime trading, the most bot-like, 44% are in profit and the median is a loss of about $35,000⁸².

Long records in ordinary markets (platform-recorded).

• Darwinex is a UK-regulated broker that records every trade and pays a trader 15% of the profits made by investors who follow the strategy⁸⁸. Its home page shows four strategies. THA: +834% since January 2015, €11.1 million of investor money, €1.01 million earned in fees. SYO: +306% since December 2016, €959,000 in fees. ERQ: +328% since October 2014, €308,000 in fees. JTL: +157% since August 2019, €231,000 in fees⁸⁸. The returns are before the fees investors pay⁸⁸. For THA that is about 21% a year (our estimate). Averaged over 7 to 12 years of record, the fee income is roughly €30,000 to €100,000 a year (our estimate: fees divided by years)⁸⁸. These are the four the broker chose to show. We did not check whether they are automated.
• Collective2 is a US marketplace that tracks strategies and sells subscriptions to them. The top of its leaderboard shows ETF Timer at +16.1% a year since January 2008 with a worst fall of 43.9%; ares at +32.4% a year since 2020 with a worst fall of 14.1%; Long Horizon at +40.9% a year since December 2022 with a worst fall of 44.7%; and AELong at +26.2% a year since May 2023⁹⁰. Every one of these managers sells subscriptions to the strategy, at $25 to $249 a month⁹⁰. A leaderboard shows only those who lasted.
• The World Cup Trading Championships is a yearly contest in futures and forex whose standings are “subject to final audit”⁹¹. Its 2025 futures winner made +324.7% in the year, second place +206.4% and fifth place +69.3%⁹¹. In 2026, to 6 October, the futures leader stood at +214.4% and the forex leader at +302.6%⁹¹. Only the top five are shown and the number of entrants is not⁹¹. The site warns: “WCC entrants may trade more than one account in the competition.”⁹¹ It does not say which entrants use a program.

Taken together, a strong verified result is tens of per cent a year, held for years, with falls of a third or more along the way. It is not thousands of per cent in a night.

Some individuals, by a modest margin. The Indian stock study above is the direct evidence: individuals’ algorithm trades beat their own hand trades by about 3 to 4 points a year, and those who lost tended to stop⁷⁴. That is a smaller shortfall, not an income. The same paper says the gain from switching to algorithms was larger for less wealthy investors and for those with less varied holdings⁷⁴. Algorithm users were a distinct, far more active group: “the median number of trades for algo and all investors is 75 and 7, respectively” a year⁷⁴. It still gives no share of algorithm users who made money.

A small skilled minority. In complete records of Taiwan’s stock exchange, “the 500 top-ranked day traders go on to earn daily before-fee (after-fee) returns of 61.3 (37.9) bps per day” in the following year²⁰. A basis point (bp) is one hundredth of one per cent, so that is 0.379% a day after fees. The same abstract concludes: “Less than 1% of the day trader population is able to predictably and reliably earn positive abnormal returns net of fees.”²⁰ The working paper, read in full on this pass, adds the wider base rate: “While about 13% earn profits net of fees in the typical year, the results of our analysis suggest that less than 1% of day traders (1,000 out of 360,000) are able to outperform consistently.”²⁰ These were people trading by hand. An earlier version of this page said we found no verified record of an individual bot operator in this group. That needs correcting: the chain-data and platform records above show automated and bot-like accounts in profit over years. What we still have not found is a named, independently verified individual who earns a living from a bot they built with an AI tool.

There is also academic support for one AI method. Researchers found that scoring news headlines with ChatGPT predicted the next day’s stock moves. They also report: “Strategy returns decline as LLM adoption rises”³⁶. The edge shrinks as more people use it.

An automated pool is not the scheme. Hyperliquid runs its own pool, called HLP, which anyone can deposit into. It acts as the exchange’s market maker, a trader that always offers both to buy and to sell, a position no outside bot has. An analytics site that works from public blockchain data puts its return at 162.8% over 3.4 years, with a worst fall of 5.8%⁴⁶. The return has slowed to 14.7% over the last year and 0.9% over the last three months, and the money in it has fallen from a peak of about $604 million to about $180 million⁴⁶. A news report says the pool “showed a $4.9 million loss” after one trader’s moves in a single coin in November 2025⁷⁷. Depositing in it is not building a bot.

Being paid without trading your own money. Four programmes pay for models, research or a track record. None publishes what a typical participant earns.

• Numerai, a hedge fund, says on its home page that it has paid $35.7 million to data scientists and has 4,900 staked models (own statement)⁹². Participants must put up the fund’s own token as a stake, and “negative scores cause staked NMR to be burned”, meaning destroyed⁹². Payouts are in that token.
• WorldQuant BRAIN, a research platform, reports more than 550,000 users and more than 16,000 paid research consultants, about 3% of users (own statement; the share is our estimate)⁹³. It publishes no pay figures on that page⁹³.
• Darwinex Zero charges €45 a month to build a record with pretend money and pays 15% of profits on any money allotted to the strategy⁸⁸. A trade news site says that in December 2025 its entry tier had 9,418 participants and €60.59 million of seed money spread across 1,799 strategies, with 93 traders getting a first allocation⁸⁹. That is roughly one strategy with an allocation for every five participants (our estimate)⁸⁹. The amount paid out is not stated, and the site’s link to the programme is an affiliate link⁸⁹.
• Polymarket pays accounts that keep orders waiting in its queue. Its documents describe daily rewards, including $1 million for one month across short-dated crypto markets⁹⁴. How that is split, and what a small participant gets, is not published.

Bots where an AI model decides each trade. This is the version in the tweets, and here the evidence of upside is thin.

• The best result in the Alpha Arena stock season was the winning model’s 12.11% in two weeks, which the organiser puts at $4,844 in total across four contests³³.
• The co-founder of the firm behind Polystrat, an AI agent for Polymarket, told a news site that “over 37% of them” show a profit, “versus less than half that number for human participants” (own statement, from a company that distributes the agent)⁸⁷. The article puts the human share at 7% to 13% and attributes it to unnamed third-party data⁸⁷. The agents’ figure covers one month, and it means about 63% of the agents were not in profit (our sum)⁸⁷.
• In the one study of real users’ AI agents on crypto perpetuals, the group was not profitable⁹⁷.

A realistic good result (our estimate). For a competent newcomer who builds and runs their own slow bot, the likely first year is a result somewhat below simply holding the asset, with a small chance of a modest profit that lasts. For a fast or leveraged bot it is a loss. The estimate rests on a range: roughly 70% to 90% of leveraged retail accounts lose (the EU and broker figures, and the Hyperliquid analysis)²⁸²⁹⁷⁵; 93% to 97% lose in the Indian options and Brazilian day-trading studies¹⁹²¹; and in the one direct study of individuals’ algorithms, they trimmed the shortfall without reversing it⁷⁴. An earlier version of this page said no source measures this group directly. The study of users’ AI agents now comes close, and it found the group was not profitable⁹⁷.

A good result, as opposed to the likely one, is to be among the roughly one in three year-old public pools that are in profit over their life so far⁷⁹. Those pools are run by bot or by hand. Only the top quarter gained 20.7% or more, and the top tenth 77% or more, over one to 3.4 years, not in a first year⁷⁹. On $5,000 those gains would be about $1,000 to $3,900 (our sum). That is a useful sum, not a wage. The file appears to leave out pools that never held about $5,000, so the true share in profit is probably lower, and a newcomer’s odds are unlikely to be better than these operators’ (our judgement)⁷⁹.

The documented route to a wage is longer: a recorded track record of several years that draws in other people’s money. The Darwinex fee earners took 7 to 12 years⁸⁸. A Hyperliquid pool leader earns 10% of depositors’ profits⁸¹. The quick, large wins we could document, on Polymarket, went to a few hundred bot-like accounts competing on speed⁸⁵⁸⁶.

What it takes to compete

What those who win have. From the sources above: programming skill⁴⁷, capital in the tens or hundreds of thousands (Step 2 and the account-size counts below)⁸², low fees⁸, slow enough trading that fees do not dominate (Step 4), and months of testing⁴⁷. The main free bot framework tells its own users: “We strongly recommend you to have basic coding skills and Python knowledge.” It adds: “Always start by running a trading bot in Dry-run and do not engage money before you understand how it works and what profit/loss you should expect.”⁴⁷ Dry-run means the bot pretends to trade without using money.

The newer sources add detail on each.

• A role, not a guess. The documented winners mostly provide liquidity or do arbitrage. Providing liquidity means keeping buy and sell orders waiting for others to trade against. In one week of Polymarket’s 15-minute crypto markets, such waiting (maker) orders won only 47% of their individual trades and still made $728,501, while orders filled at once (taker orders) won 53% of theirs and lost money overall¹⁰⁰. A high share of winning trades is therefore not proof of profit.
• Speed, if you trade fast. On seven Hyperliquid markets, market makers held a position for a median of 2.4 minutes and placed about 19 orders for each one filled; ordinary traders held for about 83 minutes⁹⁸. On Polymarket, machines account for 55.3% of the money traded in markets under 15 minutes from settling, 26.1% in markets an hour to a day away, and 16.2% in markets more than a week away⁹⁹. A slow strategy meets far fewer bots than a fast one.
• Low fees. At very high volume the maker fee on Hyperliquid falls to nothing and the taker fee to about half a newcomer’s (Step 4)⁸.
• Capital. Opening a public pool on Hyperliquid now costs a fee of 10,000 USDC (it was 100 USDC until at least February 2026), and the leader must keep at least 5% of the pool⁸¹. In the exchange’s leaderboard file, among accounts active in the last month, 25.7% of those holding under $1,000 were in profit for the month, with a median loss of $148. The share was 36.0% for $1,000 to $10,000, 52.5% for $10,000 to $100,000, 70.2% for $100,000 to $1 million and 70.7% above $1 million (our counts)⁸². Read this with care: an account’s size is partly the result of winning or losing, one month is short, and the file is not all traders⁸².
• Risk rules written in code. In the study of users’ AI agents, the risk setting each user chose drove the agents’ leverage and liquidations; the strategy prompt mattered less, and in a replay of 416 situations leading models’ decisions were statistically indistinguishable⁹⁷. The abstract does not say that the setting decided who made money. The agents used a median of 5 times leverage however jumpy the market was, and one risk setting, 11% of the book, held 62% of the liquidations⁹⁷. They also let gains go: 43.2% of positions were at least 3% ahead at some point within 24 hours, and 49.3% of those closed at a loss⁹⁷. A fixed take-profit and stop rule recovered 0.39% per position⁹⁷. In plain words, the better operator adds hard exit and size rules instead of relying on the prompt or a newer model. Among ordinary traders on seven Hyperliquid markets, 21.4% of wallets were liquidated at least once⁹⁸.
• Time. The verified records above run for two years or more, and the fee earners’ for 7 to 12 years⁸⁰⁸⁸. A typical machine-run wallet on Polymarket ran for four months⁹⁹.

What it costs. Tools are cheap. A server of the size the free framework asks for is listed at $18 a month, the framework is free, and market data is free or $99 a month (Step 5)⁴⁷¹⁰³. The exchange connection itself costs nothing beyond trading fees. Tools are not the barrier. Trading costs are: 0.09% to 0.10% for each taker round trip, plus funding that starts at about 11.6% a year of a leveraged long position’s size^8101102. And capital: at measured rates of return, $1,000 a month needs tens of thousands of dollars or more at risk (Step 2).

The field is getting more crowded. On Polymarket, machines’ share of the money went from about 16% in September 2025 to about 29% in July 2026⁹⁹. In the nine-day window studied, machine-run wallets as a group lost money after fees⁹⁹. The study’s own summary: “Automation does not buy a better result here. It buys a bigger one.”⁹⁹

How to tell early which side you are on.

• Count costs first. Add a month of platform fees, AI charges and trading fees at the bot’s real trade count, and divide by your account size. If the answer is above a few per cent a month, the bot is working mostly for the exchange.
• Run it without money for months, then with a sum you can lose entirely, and judge it over months, not nights. A backtest or a simulated run is not income³⁷.
• Count how many versions you tried. If you tested many versions and kept the best, the good backtest may be chance. Seven tries on two years of data are enough to produce a good-looking score from a strategy with no skill¹⁰⁴.
• Compare it with doing nothing. If simply holding the coin or a share index did as well, the bot added work and fees, not profit.
• Look at which side of the trade you are on. If nearly all your orders are filled at once, you pay the higher fee on every trade and trade against the accounts that wait⁸¹⁰⁰.
• Treat a liquidation as a signal. One forced close in the first months means the leverage is too high for the strategy, whatever the prompt says (our judgement, from the pattern in⁹⁷).
• Check the claim, not the screenshot. A checkable claim gives the starting sum, the period, the fees and a statement or public wallet. None of the 30 most-viewed tweets gives all of these in its text¹.
• Never give a bought or downloaded bot a key that can withdraw⁵⁴.
• If you are asked to send money out of your own account to someone else’s bot or pool, you are no longer running a bot. That is the situation the regulator cases describe⁴⁷¹⁰.

It is a reasonable hobby for someone who already programs and treats the money as the cost of learning. It can become more than a hobby for someone who also has capital, a slow or liquidity-providing strategy, and the patience to build a record over years. It is not reasonable for anyone who needs the income soon, or whose plan is to paste a prompt or buy a $69.99 bot and collect.

What we did not verify

• Direct measurement. The one direct study of individuals’ algorithms is a working paper on Indian shares from 2012 to 2019⁷⁴. We did not read the whole paper or check its treatment of costs. The one study of users’ AI agents on crypto perpetuals is a preprint; we read only its abstract, and that its authors are connected with the platforms studied is our inference from the abstract⁹⁷. The abstract reports what drove the agents’ leverage and liquidations, not what drove profit, and its comparison group is described only as “a matched Hyperliquid retail benchmark”⁹⁷. The Hyperliquid loss share is from a repost with no method⁷⁵.
• Our own counts from public files. The pool percentiles and the leaderboard and pool counts are our own sums from raw files, not published figures⁷⁹⁸²⁸³. The rules for which pools and accounts each file includes were not found. For the exchange’s pool file we assumed the last value of each pool’s lifetime series is its total profit in dollars⁸³. None of these results was compared with simply holding bitcoin over the same period. The files do not say which pools or accounts are bots, who runs them, or whether their operators sell courses or tools.
• The pool fee. The date on which Hyperliquid raised the fee for opening a pool from 100 to 10,000 USDC was not found; it lies between 9 February and 7 October 2026⁸¹.
• Named pools and track records. Growi HF’s figures were read from the analytics site, not the exchange, and who runs it was seen only in a search snippet⁸⁰. Whether the Darwinex, Collective2 and World Cup examples are automated was not checked⁸⁸⁹⁰⁹¹. For Darwinex we read only the four strategies on its home page; the share of all its strategies that are profitable was not found⁸⁸. The Darwinex Zero figures are from a trade site with an affiliate link⁸⁹. For the World Cup contest, the number of entrants, the minimum account and whether accounts hold real money were not confirmed on the site⁹¹.
• Polymarket. Only the first paragraphs of Bloomberg’s analysis were readable; the 800 wallets and $131 million are from a summary⁸⁴⁸⁵. Claims seen on secondary pages that the time window for arbitrage has shrunk sharply, and that a small share of wallets took most profits, were not traced to a primary source and are not used. The Polystrat figures are the vendor’s own, and the human share it is compared with is attributed only to unnamed data⁸⁷. The dates of the two market-making-firm studies come from a search listing⁹⁸.
• A finding that cuts the other way. A reported university analysis saying bots lost far more per user than human traders on the platforms it studied was seen only in a secondary article and not opened. It is not used on this page.
• The tweets. We did not open the tweets on X or their images and videos. We followed the links in some tweets; those pages are in the source list. The images attached to the most-viewed tweets, which may show results, were not seen. The labels for what each tweet sells are the collector’s. Whether the $71,500 in the top tweet was a backtest or real profit is not established⁷². The $900K result that one tweet attributes to a professional trader was not checked¹, and we did not confirm the figure.
• The named tools. We did not look up the AI tools, bot products and token projects named in the tweets. We did not watch the YouTube video or establish which exchange any tutorial uses⁶³.
• Quotes. Most quotes reached us through a tool that reads the page and passes on the text, which can alter wording. Not every quote was matched a second time against its original page. The quotes used from the CFTC advisory, the FTC alert, the HALO page and the Whop listing were matched again on 7 October 2026^2136568.
• India. Only the headline of the regulator’s 2024 release loaded²¹. The average loss, the 1% figure, the algorithm shares and the 13% figure are from summaries²²²³²⁴. The 2025 figure is from a search summary of a page that refused us²⁵. The 2026 figures are from a news report and a blog summary, not the regulator’s report²⁶⁹⁵.
• The AI tests. Alpha Arena’s final figures come from news reports and a blog that disagree with each other^3031106. Per-model results for its second season were not found at the organiser; the “6 of 32” count is from a blog³³¹⁰⁶. No 2026 season was found. For the academic tests we read only the abstracts⁷⁶⁹⁶.
• Fund indexes. The Barclay figures are relayed by a trade news site, not read from the index publisher, and 2024 is missing⁴⁴.
• The Brazil study. We read the abstract only. Other versions of the Brazil and Taiwan papers give somewhat different figures¹⁹²⁰.
• Fees and costs. Fees at Coinbase, Robinhood crypto, Bybit and forex brokers were not obtained. Price slippage and liquidation costs were not sourced. The server price and Hyperliquid’s maker rebate were read through a summarising tool⁸¹⁰³. Prices at other hosting firms, other data sellers and Interactive Brokers were not obtained. Our fee and AI-cost tables rest on our own assumptions about how often a bot trades and calls a model.
• The HLP pool. Figures are from a third-party site, not checked against the exchange’s own page⁴⁶. Reports of an earlier loss to the pool in March 2025 were seen only in search results and are not used.
• Rules. We did not read the final US margin rule, the registration rule for sellers, Binance’s main terms, Coinbase’s or Kraken’s terms, or Robinhood’s separate agent agreements. Whether the EU leverage caps are unchanged rests on a 2018 page. Tax was checked for the US only, at one page.
• Sellers. The site that eleven tweets lead to blocked us⁷⁰. Bybit’s headline affiliate rate, Telegram-bot referral shares, the Whop community’s membership, the disclosures page linked from the video, and the EA seller’s sales page were not confirmed. We did not look into the HALO service’s regulatory status or track record; its page says the company “does not act as a broker, dealer, investment adviser, commodity trading advisor, or commodity pool operator”⁶⁸. Numerai’s and WorldQuant BRAIN’s figures are the companies’ own, and neither publishes how participants’ earnings are spread⁹²⁹³. Quantopian’s history was read on Wikipedia, not in the news reports behind it¹⁰⁵.
• Court outcomes. For the SEC and CFTC cases we read only the releases. They are allegations except where we say a court ruled or the defendants settled⁵⁷¹⁰¹¹. We did not check later court records for any of them, including whether the court approved the 5 Fruits judgments⁹. A 2024 UK case in which the FCA charged social-media promoters over an unauthorised trading scheme is known to us only from a search listing; the release itself did not load, we do not know the outcome, and it is not used on this page⁷³.
• Not researched at all: copy-trading marketplaces beyond the one leaderboard read⁹⁰, forex EA track records, profits of arbitrage bots on other blockchains, payout figures for automated accounts at prop firms, trading firms other than Virtu, X’s own rules on financial promotion, and rules outside the US, UK and EU.

Moved out of this list on 7 October 2026 (second pass): Binance’s futures fees, now read at source¹⁰²; funding charges on Hyperliquid, now read at source¹⁰¹; server costs, now priced¹⁰³; the full text of the Taiwan working paper, now read²⁰; and the absence of any study of AI-built bots on crypto perpetuals, now filled by one preprint⁹⁷.

Sources

“Read at source” means our agents opened the page itself on 7 October 2026, in most cases through a tool that extracts the text. “Secondary” means we read an article or summary about the source, not the source itself. “Estimate” means our own sum.

1. Does It Pay collection of 47 tweets on this scheme, collected 2026-10-06; individual tweets linked above. Read from our collected copy on 2026-10-07, not from the live tweets. Counts and all sums from tweet figures are estimates.
2. CFTC customer advisory, “AI Won’t Turn Trading Bots into Money Machines”. https://www.cftc.gov/LearnAndProtect/AdvisoriesAndArticles/AITradingBots.html . Page undated; issued 2024-01-25 per³. Read at source on 2026-10-07.
3. CFTC press release 8854-24. https://www.cftc.gov/PressRoom/PressReleases/8854-24 . 2024-01-25. Read at source on 2026-10-07.
4. CFTC press release 8696-23, on CFTC v. Steynberg (Mirror Trading International). https://www.cftc.gov/PressRoom/PressReleases/8696-23 . 2023-04-27. Read at source on 2026-10-07.
5. CFTC press release 8803-23, on Technical Trading Team LLC. https://www.cftc.gov/PressRoom/PressReleases/8803-23 . 2023-10-06. Read at source on 2026-10-07. Allegations; outcome not checked.
6. SEC, NASAA and FINRA investor alert on artificial intelligence and investment offers. https://www.investor.gov/introduction-investing/general-resources/news-alerts/alerts-bulletins/investor-alerts/artificial-intelligence-fraud . Page undated; 2024-01-25 per a search listing. Read at source on 2026-10-07 through the r.jina.ai reader; a direct fetch was refused.
7. SEC Litigation Release 26558, SEC v. Nathan Fuller. https://www.sec.gov/enforcement-litigation/litigation-releases/lr-26558 . 2026-05-29. Read at source on 2026-10-07. Allegations.
8. Hyperliquid docs, fees. https://hyperliquid.gitbook.io/hyperliquid-docs/trading/fees . Undated page. Read at source on 2026-10-07. Daily and monthly fee sums are our estimate.
9. SEC Litigation Release 26390, SEC v. Guess, Ligon and 5 Fruits Enterprises LLC. https://www.sec.gov/enforcement-litigation/litigation-releases/lr-26390 . 2025-09-05. Read at source on 2026-10-07. Settled: the defendants consented to final judgments without admitting or denying the allegations, subject to court approval.
10. SEC Litigation Release 26654, SEC v. Cryptoaiml and SEC v. TSAI Pro. https://www.sec.gov/enforcement-litigation/litigation-releases/lr-26654 . 2026-09-29. Read at source on 2026-10-07. Two separate complaints. Allegations.
11. SEC Litigation Release 26453, SEC v. Morocoin Tech Corp. and others. https://www.sec.gov/enforcement-litigation/litigation-releases/lr-26453 . 2025-12-22. Read at source on 2026-10-07. Allegations.
12. FTC press release on DK Automation. https://www.ftc.gov/news-events/news/press-releases/2022/11/ftc-takes-action-stop-dk-automation-kevin-david-hulse-pitching-phony-amazon-crypto-moneymaking . November 2022. Read at source on 2026-10-07.
13. FTC consumer alert on Warrior Trading. https://consumer.ftc.gov/consumer-alerts/2022/04/day-trading-earnings-werent-payday-warrior-trading-promised . 2022-04-25. Read at source on 2026-10-07.
14. Binance fee schedule. https://www.binance.com/en/fee/schedule . Undated page. Read at source on 2026-10-07 (spot rate only; the futures rate was not confirmed).
15. 3Commas pricing. https://3commas.io/pricing . Undated page. Read at source on 2026-10-07.
16. TradingView pricing. https://www.tradingview.com/pricing/ . Undated page, prices shown in euros. Read at source on 2026-10-07.
17. Claude pricing. https://claude.com/pricing . Undated page. Read at source on 2026-10-07. Running-cost sums are our estimate.
18. Cryptohopper pricing, https://www.cryptohopper.com/pricing (undated), and affiliate programme article, https://www.cryptohopper.com/blog/how-does-the-cryptohopper-multi-level-affiliate-program-work-7533 (2022-08-22, modified 2025-01-15). Read at source on 2026-10-07.
19. Chague, De-Losso and Giovannetti, “Day Trading for a Living?”, abstract on RePEc. https://ideas.repec.org/p/fgv/eesptd/525.html (FGV working paper 525, 2020); also https://ideas.repec.org/p/spa/wpaper/2019wpecon47.html (2019). Abstract read at source on 2026-10-07. The monthly sum is our estimate. The SSRN copy and the FGV repository refused us.
20. Barber, Lee, Liu and Odean, “The Cross-Section of Speculator Skill: Evidence from Day Trading”. Working paper, https://faculty.haas.berkeley.edu/odean/papers/day%20traders/Day%20Trading%20Skill%20110523.pdf , May 2011; published abstract (Journal of Financial Markets), https://ideas.repec.org/a/eee/finmar/v18y2014icp1-24.html , 2014. Abstracts read at source on 2026-10-07. The page quotes the published version; the working paper gives lower figures, 49.5 (28.1) basis points. On the second pass (2026-10-07) the working paper’s full text was read from the PDF, for the 13% and 1,000-of-360,000 figures.
21. SEBI press release PR No. 37/2024. https://www.sebi.gov.in/media-and-notifications/press-releases/sep-2024/updated-sebi-study-reveals-93-of-individual-traders-incurred-losses-in-equity-fando-between-fy22-and-fy24-aggregate-losses-exceed-1-8-lakh-crores-over-three-years_86906.html . 2024-09-23. Read at source on 2026-10-07: headline, date and number only; the body did not load, directly or through r.jina.ai.
22. Moneylife, on the SEBI study. https://www.moneylife.in/article/93-percentage-of-individual-traders-lost-rs18-lakh-crore-in-equity-fo-in-past-3-years-sebi/75210.html . 2024-09-23. Secondary. Dollar conversions are our estimate.
23. Taxmann, on the SEBI study. https://www.taxmann.com/post/blog/93-of-1-crore-fo-traders-saw-rs-2l-avg-loss-while-top-3-5-faced-rs-28l-avg-loss-incl-costs-sebi/ . 2024-09-23 (study date). Secondary.
24. TradingQnA (the broker Zerodha’s forum), on the SEBI study. https://tradingqna.com/t/sebis-latest-analysis-of-profits-losses-in-the-equity-derivatives-segment-fy22-fy24/173596 . 2024-09-23. Secondary.
25. Business Standard, on SEBI’s FY25 study. https://www.business-standard.com/markets/news/net-losses-of-traders-in-fo-widens-in-fy25-sebi-study-125070701221_1.html . 2025-07-07. Secondary: search summary only; the page refused us.
26. Business Today, on SEBI’s FY26 study. https://www.businesstoday.in/markets/story/rs91685-cr-lost-88-of-individual-traders-lost-money-in-fy26-options-drove-92-of-losses-550474-2026-08-21 . 2026-08-21. Secondary.
27. AlgoTest blog, “algo trading edge”. https://algotest.in/blog/algo-trading-edge/ . 2024-10-08, updated 2025-05-09. Read at source on 2026-10-07. Seller’s own figures.
28. ESMA, “ESMA agrees to prohibit binary options and restrict CFDs to protect retail investors”. https://www.esma.europa.eu/press-news/esma-news/esma-agrees-prohibit-binary-options-and-restrict-cfds-protect-retail-investors (also https://www.esma.europa.eu/node/84933 ). 2018-03-27. Read at source on 2026-10-07.
29. IG, CFD trading page. https://www.ig.com/en/cfd-trading . Undated page. Read at source on 2026-10-07. The percentage changes over time.
30. ForkLog, on Alpha Arena season 1. https://forklog.com/en/four-out-of-six-ai-models-suffer-losses-in-trading-tournament/amp . November 2025. Secondary.
31. GNcrypto, on Alpha Arena season 1 (in Russian). https://www.gncrypto.news/ru/news/qwen-wins-alpha-arena-season-1-with-22-percent-returns/ . 2025-11-04. Secondary.
32. Decrypt, on Alpha Arena three days in. https://decrypt.co/345006/ai-crypto-trading-showdown-deepseek-grok-winning-gemini-implodes . 2025-10-20. Secondary (news report), read on 2026-10-07.
33. Alpha Arena site, https://alphaarena.ai , season ended 2025-12-03, read at source on 2026-10-07 through the r.jina.ai reader (a direct fetch was refused); and ForkLog, https://forklog.com/en/ai-model-grok-4-2-triumphs-in-trading-tournament/ , 2025-12-08, secondary, read on 2026-10-07.
34. Paper on on-chain AI investment agents (authors at Pantera Capital and Stanford). https://arxiv.org/abs/2605.29174 . 2026-05-27. Read at source on 2026-10-07. Preprint.
35. “Can LLM-based Financial Investing Strategies Outperform the Market in Long Run?”. https://arxiv.org/abs/2505.07078 . 2025-05-11, revised 2026-06-26. Abstract read at source on 2026-10-07.
36. Lopez-Lira and Tang, “Can ChatGPT Forecast Stock Price Movements?”. https://arxiv.org/abs/2304.07619 . First posted April 2023, latest version 2025-10-28. Abstract read at source on 2026-10-07.
37. Quantpedia, reproducing the abstract of Quantopian’s study of 888 algorithms. https://quantpedia.com/quantopians-academic-paper-about-in-vs-out-of-sample-performance-of-trading-alg/ . 2016-05-04. Secondary; the paper’s host refused us.
38. pytrader, open-source bitcoin bot by “owocki”. https://github.com/owocki/pytrader . 2016-03-26; shelved 2017-03-20. Read at source on 2026-10-07. Self-reported.
39. tildalice.io, lessons from a machine-learning crypto bot. https://tildalice.io/ml-crypto-trading-bot-lessons/ . 2026-06-04. Read at source on 2026-10-07 through the r.jina.ai reader; a direct fetch was refused. Self-reported; the page carries affiliate links.
40. Indie Hackers post on running six bots. https://www.indiehackers.com/post/started-building-bots-because-i-was-scared-of-falling-behind-on-ai-ended-up-with-six-of-them-running-24-7-on-a-mac-mini-1c26e805cf . 2026-04-05. Read at source on 2026-10-07. Self-reported.
41. Finance Magnates, https://www.financemagnates.com/forex/only-1-in-20-traders-pass-prop-firm-challenges-reports-the-funded-trader/ , 2025-03-18; and TradeInformer, https://www.tradeinformer.com/prop-weekly/what-ftmo-tells-us-about-prop-firm-finances , 2025-03-07. Secondary (industry statements relayed by trade press).
42. Decrypt, on pump.fun wallets. https://decrypt.co/300403/pump-fun-traders-millionaires . 2025-01-10. Secondary.
43. Yellow.com, on a DWF Ventures study of copy-trading wallets. https://yellow.com/news/fomo-copy-trading-94-percent-wallet-losses . 2026-08-28. Secondary; the study itself was not found.
44. The Full FX, on Barclay CTA indexes: https://thefullfx.com/ctas-end-2025-on-a-positive-note/ , 2026-01-27 (2025 figure), and https://thefullfx.com/currency-managers-highlight-cta-performance-in-2023/ , 2024-01-24 (2023 figure). Secondary (trade press relaying the index publisher), both read on 2026-10-07.
45. Virtu Financial, Form S-1 filed with the SEC. https://www.sec.gov/Archives/edgar/data/0001592386/000104746914002070/a2218589zs-1.htm . 2014-03-10. Read at source on 2026-10-07 (first part of the filing).
46. Trading Strategy, Hyperliquid HLP vault page, https://tradingstrategy.ai/trading-view/vaults/hyperliquidity-provider-hlp , and vault listing, https://tradingstrategy.ai/trading-view/vaults . Live pages, 2026-10-07. Read at source on 2026-10-07. Third-party analytics; a CoinGecko article seen only in search results gives different lifetime figures. Annualised and capital sums are our estimate.
47. Freqtrade documentation. https://www.freqtrade.io/en/stable/ . Undated page. Read at source on 2026-10-07 through the r.jina.ai reader; a direct fetch was refused.
48. Robinhood Customer Agreement, section 29. https://cdn.robinhood.com/assets/robinhood/legal/Robinhood-Customer-Agreement.pdf . Revised 2026-10-01. Read at source on 2026-10-07 (text extracted from the PDF).
49. ACA Global, “FINRA ends the pattern day trader rule”. https://www.acaglobal.com/industry-insights/finra-ends-the-pattern-day-trader-rule/ . 2026-05-06. Secondary.
50. Federal Register, SEC Release No. 34-104572, notice of FINRA’s proposed change to Rule 4210. https://www.govinfo.gov/content/pkg/FR-2026-01-14/html/2026-00519.htm . Release dated 2026-01-09, published 2026-01-14. Read at source on 2026-10-07. Proposal, not the final rule.
51. Hyperliquid terms. https://app.hyperliquid.xyz/terms . Date not shown. Read at source on 2026-10-07 through the r.jina.ai reader; a direct fetch returned an empty page. Quotes are fragments of clauses 1.6, 3.1.5 and 6.2.
52. Hyperliquid docs, referrals and builder codes. https://hyperliquid.gitbook.io/hyperliquid-docs/referrals and https://hyperliquid.gitbook.io/hyperliquid-docs/trading/builder-codes . Undated pages. Read at source on 2026-10-07.
53. The Block, on Binance account bans. https://www.theblock.co/post/375278/binance-bans-more-than-600-accounts-over-unauthorized-third-party-tools . 2025-10-20. Secondary.
54. Binance.US help, API key safety. https://support.binance.us/en/articles/9842812-binance-us-api-keys-best-practices-safety-tips . 2026-06-03. Read at source on 2026-10-07.
55. Cointelegraph, on the 3Commas key leak, https://cointelegraph.com/news/3commas-ceo-confirms-api-key-leak-following-warning-from-cz , 2022-12-28, news report read on 2026-10-07; and Halborn, https://halborn.com/explained-the-3commas-breach-december-2022/ , secondary, read on 2026-10-07, for the loss figure, which it takes from a CoinDesk report on an FBI investigation that we did not open.
56. The Block, on the Banana Gun exploit and refund. https://www.theblock.co/post/318074/banana-gun-exploit-refund . 2024-09-25. Secondary.
57. Cointelegraph, on a GitHub trading bot that stole wallet keys. https://cointelegraph.com/news/solana-trading-bot-github-malware-scam . 2025-07-04. Secondary (reports a SlowMist analysis we did not open).
58. Sentora research article on the October 2025 liquidations. https://sentora.com/research/articles/structural-shifts-in-crypto-the-black-friday-drawdown-and-hyperliquid-s-bid-to-challenge-cexs . October 2025. Secondary.
59. FCA, “FCA bans the sale of crypto-derivatives to retail consumers”. https://www.fca.org.uk/news/press-releases/fca-bans-sale-crypto-derivatives-retail-consumers . 2020-10-06. Read at source on 2026-10-07. For the 2025 change see⁷⁸.
60. IRS, frequently asked questions on virtual currency transactions. https://www.irs.gov/individuals/international-taxpayers/frequently-asked-questions-on-virtual-currency-transactions . Updated 2026-06-30. Read at source on 2026-10-07.
61. Terms.law, guide to registration as a commodity trading advisor. https://terms.law/Trading-Legal/guides/cftc-cta-startup-kit.html . Date not checked. Secondary.
62. WFTV, “What is agentic trading?”. https://www.wftv.com/news/is-agentic-trading/NV64RNJKZQ2MXN4LWQNOY6DS7I/ . 2026. Secondary: search snippet only; the page was blocked to us.
63. YouTube, Miles Deutscher Finance, “Jev + Claude Opus 5.5 = The Most Powerful AI Trading Bot Ever”. https://www.youtube.com/watch?v=Pk7W7BKMwqo . 2026-09-24. Description read at source on 2026-10-07 through the r.jina.ai reader. The video itself was not watched.
64. Skool, “Miles Deutscher Finance” community page, https://www.skool.com/milesdeutscherfinance , and Skool pricing, https://www.skool.com/pricing . Undated pages. Read at source on 2026-10-07.
65. Whop, Miles High Club. https://whop.com/miles-high-club/miles-high-club/ . Undated page. Read at source on 2026-10-07. No member count is shown.
66. Bybit affiliate programme FAQ. https://www.bybit.com/en/help-center/article/Bybit-Affiliate-Program-General-FAQ . Undated page. Read at source on 2026-10-07 through the r.jina.ai reader; a direct fetch timed out.
67. MetaTrader 5 help, selling in the Market. https://www.metatrader5.com/en/terminal/help/market/market_sell . Undated page. Read at source on 2026-10-07.
68. The Trading Hub, HALO sign-up page. https://discover.thetradinghub.com/opt-in-page . Undated page. Read at source on 2026-10-07. Returns shown are the seller’s claim; the fee shares are our estimate.
69. LaborX freelance listing for custom bots. https://laborx.com/gigs/i-will-create-ai-openclaw-polymarket-bot-robinhood-bot-crypto-trading-bot-volume-bot-pumpfun-bot-122338 . Undated page. Read at source on 2026-10-07. Price shown as 750, payable in several crypto coins.
70. blockedge.live. https://blockedge.live/ . Tried on 2026-10-07; refused directly and through r.jina.ai. The link destinations were followed by our agents; the tweet count is our estimate from¹.
71. Tutorial for the Based Telegram bot. https://intercom.help/actech/en/articles/16826218-based-bot-tutorial-your-first-trade-from-scratch . October 2026. Secondary (third-party tutorial for a bot of the same name; the match to the linked bot was not confirmed).
72. BeInCrypto, on the ChatGPT trading-bot thread. https://beincrypto.com/crypto-trader-chatgpt-trading-bot/ . 2024-04-25. Secondary (news report on the thread), read on 2026-10-07 through the r.jina.ai reader; a direct fetch was refused.
73. FCA press release on charges against social-media promoters of a trading scheme. https://fca.org.uk/news/press-releases/finfluencers-charged-promoting-unauthorised-trading-scheme . 2024-05-16. Secondary: search listing only. On 2026-10-07 the address returned the FCA news index, not the release.
74. Ghosh, Li and Zhang, “When Human Meet Algorithm: the Adoption and Impact of Retail Algorithmic Trading”. https://afajof.org/management/viewp.php?n=152460 . Working paper, 2025-03-15. Read at source on 2026-10-07 (text extracted from the PDF; abstract and pages around Table 2).
75. HTX news, repost of an analysis by Stacy Muur of Hyperliquid addresses. https://www.htx.com/news/data-75-of-traders-on-hyperliquid-are-losing-money-what-are-R2h1lIXH/ . 2026-05-20. Secondary, read on 2026-10-07 through the r.jina.ai reader. No sample size or method given.
76. Two academic live trading tests of AI models: “When Agents Trade: Live Multi-Market Trading Benchmark for LLM Agents”, https://arxiv.org/abs/2510.11695 , 2025-10-13; and “LiveTradeBench: Seeking Real-World Alpha with Large Language Models”, https://arxiv.org/abs/2511.03628 , 2025-11-05. Abstracts read at source on 2026-10-07; the full papers were not read.
77. Cointelegraph, on the POPCAT trades and the loss to Hyperliquid’s HLP pool. https://cointelegraph.com/news/hyperliquid-hlp-popcat-attack-3m-wipeout . 2025-11-13. Secondary (news report), read on 2026-10-07.
78. FCA, “FCA opens retail access to crypto ETNs”. https://www.fca.org.uk/news/press-releases/fca-opens-retail-access-crypto-etns . 2025-08-01. Read at source on 2026-10-07.
79. Trading Strategy, data file of vaults by chain, https://top-defi-vaults.tradingstrategy.ai/top_vaults_by_chain.json (file generated 2026-10-07), and Hyperliquid vault listing, https://tradingstrategy.ai/trading-view/vaults/chains/hyperliquid (data updated 2026-10-07). Read at source on 2026-10-07. Third-party analytics from public chain data. Shares, medians and percentiles are our estimate computed from the file; the file appears to include only vaults that at some point held about $5,000 or more (the smallest peak value among its Hyperliquid vaults is exactly $5,000).
80. Trading Strategy, Growi HF vault page. https://tradingstrategy.ai/trading-view/vaults/growi-hf . Live page, 2026-10-07. Read at source on 2026-10-07 (page through a summarising tool; figures also read from the data file at⁷⁹).
81. Hyperliquid docs, for vault leaders. https://hyperliquid.gitbook.io/hyperliquid-docs/hypercore/vaults/for-vault-leaders . Undated page. Read at source on 2026-10-07 through the r.jina.ai reader (fee shown: 10k USDC). Archived copy of the same page, https://web.archive.org/web/20260209055421/https://hyperliquid.gitbook.io/hyperliquid-docs/hypercore/vaults/for-vault-leaders , captured 2026-02-09, read at source on 2026-10-07 (fee shown: 100 USDC); the April and June 2025 captures were reported by a reviewer and not re-opened.
82. Hyperliquid, public leaderboard data file. https://stats-data.hyperliquid.xyz/Mainnet/leaderboard . Live file, 2026-10-07. Read at source on 2026-10-07. All counts and shares are our estimate from the file; its inclusion rule is not documented.
83. Hyperliquid, public vault data file. https://stats-data.hyperliquid.xyz/Mainnet/vaults . Live file, 2026-10-07. Read at source on 2026-10-07. All counts are our estimate; we read the last value of each vault’s all-time series as its total profit in dollars, which is not documented.
84. Bloomberg News, analysis of profits and losses of Polymarket wallets. https://news.bgov.com/crypto/prediction-market-users-suffer-broad-losses-as-bots-reap-gains . 2026-04-28. Read at source on 2026-10-07 through the r.jina.ai reader: first paragraphs only; the rest is paywalled.
85. Newser, summarising the Bloomberg analysis at⁸⁴. https://www.newser.com/story/388532/polymarkets-biggest-winners-arent-human.html . 2026-05-10. Secondary.
86. Academic paper on arbitrage carried out on Polymarket (authors at IMDEA Networks and Oxford). https://arxiv.org/abs/2508.03474 (full text at https://arxiv.org/html/2508.03474v1 ). 2025-08-05. Read at source on 2026-10-07.
87. CoinDesk, on AI agents trading on prediction markets. https://www.coindesk.com/tech/2026/03/15/ai-agents-are-quietly-rewriting-prediction-market-trading . 2026-03-15. Read at source on 2026-10-07 through the r.jina.ai reader. The figures are a vendor’s own statement.
88. Darwinex home page. https://www.darwinex.com/ . Undated live page. Read at source on 2026-10-07 through the r.jina.ai reader. Platform-recorded returns, shown before investors’ fees; yearly rates and yearly fee income are our estimate.
89. Traders Union, on Darwinex’s December 2025 allocation results. https://tradersunion.com/news/brokers-news/show/1250805-darwinia-results-highlight-global/ . 2026-01-11. Secondary (trade news site; its link to the programme is an affiliate link).
90. Collective2 leaderboard. https://collective2.com/leader-board . Live page, 2026-10-07. Read at source on 2026-10-07 through the r.jina.ai reader. Platform-tracked; every listed manager sells subscriptions.
91. World Cup Trading Championships standings. https://www.worldcupchampionships.com/world-cup-trading-championship-standings . Live page, standings to 2026-10-06. Read at source on 2026-10-07 directly and through the r.jina.ai reader. Standings are subject to the organiser’s final audit.
92. Numerai home page, https://numer.ai/ , and staking rules, https://docs.numer.ai/numerai-tournament/staking . Undated pages. Read at source on 2026-10-07 through the r.jina.ai reader. Company’s own figures.
93. WorldQuant BRAIN. https://www.worldquant.com/brain/ . Undated live page. Read at source on 2026-10-07 through the r.jina.ai reader. Company’s own figures; the share is our estimate.
94. Polymarket docs, liquidity rewards. https://docs.polymarket.com/developers/market-makers/liquidity-rewards . Undated page. Read at source on 2026-10-07 through the r.jina.ai reader.
95. Sahi blog, on SEBI’s FY26 study. https://www.sahi.com/blogs/sebi-fy26-fo-report-retail-traders . 2026-08-20. Secondary (summary of the regulator’s report), read on 2026-10-07 through the r.jina.ai reader.
96. “StockBench”, a test of AI agents on real 2025 stock prices. https://arxiv.org/abs/2510.02209 . 2025-10-02, revised 2026-03-02. Abstract read at source on 2026-10-07 through the r.jina.ai reader; the full paper was not read.
97. Paper on two fleets of user-funded AI trading agents (DX Terminal and DXAP). https://arxiv.org/abs/2609.05663 . 2026-09-04. Abstract read at source on 2026-10-07. Preprint; the authors appear to operate the platforms studied.
98. Arrakis (a market-making firm), study of who trades on seven Hyperliquid markets, https://arrakis.finance/blog/who-is-trading-on-hip3 , 2026-04-08, and its study of Trade.xyz, https://arrakis.finance/blog/trade-xyz , 2026-04-29. Read at source on 2026-10-07 through the r.jina.ai reader. Dates are from a search listing; the pages show none. The firm sells market-making services.
99. Bitquery, study of how much of Polymarket is machine-run. https://bitquery.io/investigations/how-much-of-polymarket-is-bots . 2026-08-20. Read at source on 2026-10-07 through the r.jina.ai reader. The firm sells blockchain data.
100. Telonex, on top traders in Polymarket’s 15-minute crypto markets. https://telonex.io/research/top-crypto-traders-polymarket-15m . 2026-02-13. Read at source on 2026-10-07. The firm sells prediction-market data.
101. Hyperliquid docs, funding. https://hyperliquid.gitbook.io/hyperliquid-docs/trading/funding . Undated page. Read at source on 2026-10-07. The 10-times-leverage sums are our estimate.
102. Binance futures fee schedule. https://www.binance.com/en/fee/futureFee . Undated page. Read at source on 2026-10-07 through the r.jina.ai reader; a direct fetch showed no records. The round-trip sum is our estimate.
103. DigitalOcean server pricing, https://www.digitalocean.com/pricing/droplets (read through a summarising tool), and Alpaca market-data plans, https://alpaca.markets/data (read through the r.jina.ai reader). Undated pages. Read at source on 2026-10-07.
104. Notices of the American Mathematical Society, article on how trying many strategy versions produces good backtests by chance. https://www.ams.org/notices/201405/rnoti-p458.pdf . May 2014. Read at source on 2026-10-07.
105. Wikipedia, “Quantopian”. https://en.wikipedia.org/wiki/Quantopian . Read on 2026-10-07 through a summarising tool (events 2016 to 2020). Secondary; the news reports behind it were not opened.
106. ai2.work blog, on Alpha Arena’s seasons. https://ai2.work/blog/ai-trading-bots-hand-retail-investors-real-hedge-fund-tooling . 2026-08-05. Secondary, read on 2026-10-07 through the r.jina.ai reader; not checked against the organiser.
107. Bitsgap blog, first-half 2026 results. https://bitsgap.com/blog/bots-that-delivered-bitsgaps-h1-2026-results-revealed . 2026-07-24. Read at source on 2026-10-07 through a summarising tool. Seller’s own figures.

Corrections

If you are quoted, named or described on this page and think something is wrong, want your reply shown beside it, or want to be removed, write to [email protected]. We aim to reply within 14 days, and always within one month. Factual errors are corrected with a dated note here. A sentence that is seriously disputed comes down while we check it.

Change log:

• 2026-10-07: Second research pass. Replaced “When it can work” with “The upside” and “What it takes to compete”. Added sources 79 to 107. Corrected four statements, each marked in the text with “An earlier version of this page said”: that no study of AI-built bots on crypto perpetuals existed, that funding charges and Binance futures fees were unsourced, that nothing showed bots doing worse than people, and that no verified record of a bot operator in profit had been found.
• 2026-10-07: Review of the second pass. Corrected the statement that every pool operator counted had paid a 10,000 USDC fee: archived copies show the fee was 100 USDC until at least February 2026⁸¹. Restated the AI-agent study to match its abstract⁹⁷. Updated the BredoStrategy and Growi HF figures to the file as read that day⁷⁹⁸⁰. Made clear that the one-in-three figure is pools in profit over a life of one to 3.4 years, by bot or by hand, not bots in profit after one year. Source 81 gained an archive link; no source was added or renumbered.
• 2026-10-07: wording about named accounts and companies reviewed; the regulator advisory is no longer quoted beside the tweets’ figures, a trader is no longer named and a site is no longer named in the text, undecided cases are marked as alleged with outcomes where known, and the HALO and Whop passages were re-read at source.