Do AI Trading Bots Actually Work? What the Evidence Says

AI trading bots work as execution tools: they carry out a strategy's rules faster and more consistently than a person can, but they cannot create an edge the strategy lacks, so a bot makes money only when the strategy inside it already does. An AI trading bot is software that places orders automatically from fixed rules or a statistical model, marketed with artificial intelligence as the headline feature; an edge is a repeatable advantage that stays profitable after costs.

Key takeaway

An AI trading bot automates how a strategy is executed and cannot supply an edge the strategy lacks. Because the best of many zero-skill backtests looks skilled by luck alone, trust a bot only on an audited live record that discloses how many variants were tested, never on its backtest.

What does an AI trading bot actually do?

An AI trading bot automates the execution layer, which turns decisions into orders, and leaves the strategy layer, which makes the decisions, unchanged. The strategy layer reads market data and outputs entry, exit and position size; the execution layer sends them to the broker as orders, places and moves stops, and runs while the trader sleeps. Bot marketing usually attaches the AI label to the strategy layer, the one part a buyer cannot inspect.

A bot is excellent at the right-hand box: discipline, speed and consistency. Whether it makes money is decided in the left-hand box, which the "AI" label describes and the buyer usually cannot see.

Automation does improve execution: a bot trades without fear or revenge, reacts in milliseconds and sizes the hundredth trade exactly like the first. No bot can turn a strategy with negative expectancy (an average trade that loses money after costs) into a winner; automation only delivers the losses more consistently. As the CFTC's customer advisory on AI trading bots puts it: "AI technology can't predict the future or sudden market changes."

Why do trading bots fail live after a great backtest?

Trading bots fail live after a great backtest mainly because the backtest was the best of many attempts, and the best of many attempts is mostly the luckiest. A backtest simulates a strategy on historical prices; backtest overfitting is tuning it until it fits past noise, and multiple testing (running many variants and keeping the winner) makes overfitting almost automatic, because every extra variant is another lottery ticket.

Bailey, Borwein, López de Prado and Zhu formalised the problem in "Pseudo-Mathematics and Financial Charlatanism" (Notices of the American Mathematical Society, 2014). For a zero-skill strategy, the paper's Equation 2.3 implies that the annualised Sharpe ratio (average excess return divided by its volatility) estimated from y years of data is approximately normal with mean 0 and standard deviation 1/√y, so one year of data gives a standard error of about 1. A one-year backtest Sharpe of 1.5 is therefore only 1.5 standard errors from zero, short of the 1.96 needed for 95% confidence, and a zero-skill strategy reaches 1.5 or more 6.7% of the time. The authors name the missing fact: "A researcher that does not report the number of trials N used to identify the selected backtest configuration makes it impossible to assess the risk of overfitting."

One year of daily data measures a Sharpe ratio with a standard error of about one. A backtest at 1.5 sits comfortably inside what pure chance produces, before anyone has tried a second variant.

Across 100 zero-skill strategies tested on the same year, an average of 2.3 show a Sharpe above 2.0, and a vendor who publishes only the top one is selling noise that looks like skill. Simulated fills widen the gap further (the choice between tick data and bar data can decide whether a scalping backtest means anything); for your own ideas, see how to backtest a futures strategy without overfitting.

How many random strategies does it take to fake a Sharpe of 2.5?

About 100: among 100 zero-skill strategies tested on one year of data, the best has an expected backtest Sharpe ratio of 2.51 and a 90.0% chance of showing 2.0 or more, while its expected live Sharpe is 0.

The table applies the paper's distribution: the expected best of N zero-skill strategies is E[max of N standard normal draws] ÷ √y, computed here by exact numerical integration (the paper's closed-form approximation gives a slightly higher 1.57 for N = 10), and on a one-year backtest the chance that the best clears a Sharpe of s is 1 - Φ(s)^N, where Φ is the standard normal cumulative distribution function. Inputs are illustrative: independent strategies, normal returns and a true Sharpe of 0 for every strategy.

Strategies tried (N)	Expected best Sharpe, 1 year	Chance best is 2.0 or more, 1 year	Chance best is 2.5 or more, 1 year	Expected best, 3 years	Expected best, 5 years
1	0.00	2.3%	0.6%	0.00	0.00
10	1.54	20.6%	6.0%	0.89	0.69
100	2.51	90.0%	46.4%	1.45	1.12
1,000	3.24	>99.99%	99.8%	1.87	1.45
10,000	3.85	>99.99%	>99.99%	2.22	1.72

Seven on/off settings create 128 variants, enough to manufacture a 2.59 Sharpe from nothing in a one-year test. Longer backtests shrink the illusion but never remove it; only a count of variants tried lets you discount it.

Seven on/off parameters already produce 2^7 = 128 variants (expected best 2.59), and an optimiser sweeping three parameters across 10 values each runs 1,000 backtests. Longer histories help slowly: the expected best zero-skill Sharpe falls to 1 only after 2.37 years of data for 10 trials, 6.29 years for 100 and 10.51 years for 1,000. The paper sets the same budget for holding the expected best zero-skill Sharpe at 1: with 5 years of data, "no more than 45 independent model configurations should be tried".

Real strategy variants are correlated, so the effective number of independent trials is below the raw count, and fat tails, autocorrelation and costs shift every value; the table shows how luck scales, not a forecast for any bot.

Do trading bots make money for the people who buy them?

Buyers have little evidence that they do: no reliable public statistic shows what share of retail trading bots make money, and the often-repeated claim that 90% of bots lose has no primary source we could trace. Wiecki, Campbell, Lent and Stauth (2016, The Journal of Investing) studied 888 algorithms built on the Quantopian platform, each with at least six months of out-of-sample results, and found that backtest Sharpe ratios explained less than 2.5% of the variation in out-of-sample performance (R² below 0.025), while backtest volatility and maximum drawdown carried more predictive value; the more backtesting a quant had done on a strategy, the larger the gap.

The launch date is chosen after the backtest looks best, which is exactly when selection bias is largest. On 888 real algorithms, the backtest explained almost none of what happened next.

Vendor selection puts the buyer on the wrong side of the backtest-to-live gap. The vendor shows the in-sample winner, 2.51 in the 100-trial case; the buyer receives the live expectation of 0, minus commissions, slippage and the subscription fee, so a slow loss is the expected result of a zero-skill bot. For the slower route of building your own edge, see how long it takes to become a profitable trader.

Are AI trading bots a scam? What regulators say

Not every trading bot is a scam, but the CFTC, the SEC, FINRA and NASAA have all issued warnings or brought cases about AI claims in trading and investing since January 2024. The CFTC advisory "AI Won't Turn Trading Bots into Money Machines" (announced January 25, 2024) opens: "Fraudsters are exploiting public interest in artificial intelligence (AI) to tout automated trading algorithms, trade signal strategies, and crypto-asset trading schemes that promise unreasonably high or guaranteed returns." The case it cites is Mirror Trading International, whose bot trading program "guaranteed at least a 10 percent monthly return (or more than 200 percent per year)"; according to the advisory, the scheme operated as a Ponzi scheme and took more than $1.7 billion in bitcoin from at least 23,000 people.

On March 18, 2024 the SEC charged two investment advisers, Delphia (USA) Inc. and Global Predictions Inc., with making false and misleading statements about their use of AI; both settled, without admitting or denying the findings, for civil penalties of $225,000 and $175,000 respectively. Gary Gensler, SEC Chair at the time, named the practice: "Investment advisers should not mislead the public by saying they are using an AI model when they are not. Such AI washing hurts investors."

A joint SEC, NASAA and FINRA alert, "Artificial Intelligence (AI) and Investment Fraud" (January 25, 2024), gives the test for any bot sales page: "Investment claims that sound too good to be true usually are."

How do you check a trading bot before paying for it?

Check a trading bot by asking five questions in order and stopping at the first answer the seller cannot document.

Each question removes one way a backtest can mislead: luck, hidden variants, missing costs, hidden drawdowns, and custody. The last one is the one that protects the deposit when everything else is fiction.

Is there a live, independently verified record? Broker statements or third-party verification of at least a year of real-money trades. A backtest, demo account or screenshot does not count.
How many variants were tested, and what was the out-of-sample period? Out-of-sample data is data never used for tuning. The deflated Sharpe ratio of Bailey and López de Prado (The Journal of Portfolio Management, 2014) discounts a Sharpe ratio for the number of trials and for non-normal returns; in the authors' words, "DSR helps separate legitimate empirical findings from statistical flukes."
Are all costs included? Commissions, exchange fees, slippage and the subscription. A futures scalper averaging a tick or two per trade can lose its whole edge to costs.
What was the worst live drawdown? Maximum drawdown, the largest peak-to-trough loss, shows whether your account or a prop firm's loss limit would have survived.
Who holds the money? The bot should trade an account in your name at a regulated broker; Mirror Trading International customers paid bitcoin into the pool itself. For futures, check whether the seller is registered in the National Futures Association's free BASIC database.

Can you use a trading bot on a prop firm account?

Often yes, but automation rules vary by firm, so the same bot can be allowed at one firm and grounds for removal at another. As of September 2026, Topstep's TopstepX API Access page says "Custom automated strategies and bots are allowed via the TopstepX / ProjectX API", subject to its ban on high-frequency trading, but also that "All trading activity must originate from your personal device"; it prohibits VPS, VPN and remote-server use and warns that automation on a VPS can lead to suspension or removal.

Topstep's Prohibited Trading Strategies page (updated June 10, 2026) also bans software or AI that manipulates, abuses or gains an unfair advantage, and scalping algorithms built to exploit unrealistic simulated fills; those fills are one reason an evaluation pass says little about funded results (see why sim results don't match live trading). Rules change often, so check the firm-by-firm rules on automated and algo trading, and read why prop firms ban copying signal providers before buying anyone's signals.

Is a trade copier the same as a trading bot?

No: a trade copier copies decisions rather than making them, sending one master account's trades to other accounts, so it inherits whatever edge, or lack of edge, the master has. Copying a losing strategy to five accounts loses five times over. Thor, this blog's own product, is a copier, and whether copy trading works comes down to the same question as whether a bot works.

A copier earns its place when a trader with a tested approach runs several funded accounts and wants the same trade in each, sized per account. Thor runs server-side, so copying continues when the trader's computer is off, which makes it the wrong tool at any firm that requires trading to originate from the trader's own device, as Topstep's API page does with no stated exception for copiers. Confirm the firm's rules first, then test the copier on sim accounts before going live.

Go deeper

Frequently asked questions

Is there any AI trading bot that actually makes money?

Profitable automated strategies exist, mostly inside professional trading firms that do not sell them. A strategy with a durable edge is usually worth more to its owner traded with capital than rented out for a monthly fee, so a retail listing is weak evidence of one.

Can ChatGPT or another AI chatbot predict the market?

No. A general chatbot has no knowledge of future prices; it can write strategy code or explain indicators, but any strategy it suggests needs the same out-of-sample testing as a human idea. Asking a chatbot for 50 variants and keeping the best backtest is multiple testing at high speed.

Are AI trading bots different from ordinary algorithmic trading?

Mostly in the marketing. A rule-based algorithm and a machine-learning model face the same overfitting arithmetic, and a flexible model with many parameters has more ways to fit noise, so it needs more out-of-sample data, not less.

How long should a bot's live track record be before I trust it?

Roughly 4 divided by the square of the claimed Sharpe ratio, in years, before the result sits two standard errors above zero: 16 years for a Sharpe of 0.5, 4 years for 1 and 1 year for 2. The rule uses a standard error of about 1 divided by the square root of the years traded, and choosing among many vendors raises the bar further.

Does a trading bot keep working when the market changes?

Not reliably: a model fitted to past data has no way to know when the patterns it learned stop holding. Strategies tuned in one volatility regime can break when volatility or trend behaviour shifts, so a live record should span different market conditions, not one calm year.

Is a 100% win rate trading bot possible?

Only by hiding losses, usually by never closing losing trades or by adding to them (martingale averaging), which turns many small wins into one large loss later. A win rate says nothing about profit without the size of the average loss.

Are trading bots worth it for beginners?

Rarely, because a bot automates a strategy and a beginner usually has no tested strategy to automate. Buying one also means judging a backtest, the skill beginners have least of.

Is it legal to use a trading bot?

Yes, running your own automated strategy in your own account is legal in US markets as long as it does not manipulate prices, for example by spoofing, and it follows your broker's and any prop firm's rules. Selling futures trading advice to the public is different: the seller may need CFTC registration as a commodity trading advisor and NFA membership unless an exemption applies.

Can I compute a deflated Sharpe ratio myself?

Yes, with a spreadsheet or a few lines of code. It needs the observed Sharpe ratio, the number of independent trials, the variance of Sharpe ratios across those trials, the number of return observations, and the skewness and kurtosis of returns; without an honest trial count the result is meaningless.