Call · 30 min
Behind the curtainMattia Esposito12 November 202612 min read

Do trading bots work? 521 hours in the markets, nine AI agents, 18 bots and zero euros: what I learnt, and what the regulators say.

Anyone selling a trading bot shows you a chart going up, and almost never tells you how many attempts it took to draw it. Here is what the regulators write, what happened to me over 521 hours and the checklist to fill in first, in Word and PDF.

In brief

None of the regulators' warnings read for this page says a bot makes money on its own, and the American ones say the opposite of what the ads promise. The CFTC, January 2024: “AI technology can’t predict the future or sudden market changes”.

The weak spot is the backtest. A study published in the Notices of the American Mathematical Society in 2014 shows that a handful of configurations is enough to produce a brilliant result on past data, and the people selling it rarely say how many they tried.

I spent 521 hours on it, first with nine AI agents and then with 18 bots, and by 22 September 2026 I had made zero euros. The story is below, with the mistakes and what I changed.

The ten-field checklist is free to download below, in Word and PDF, with no email required. This is not financial advice.

Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.

The checklist before you put money into a bot, ready to download

A sheet with ten questions to ask anyone selling a bot, or yourself before you switch one on, in Word to fill in and PDF to print. Each field comes from a regulator's public warning, a study on backtesting or a mistake I made myself. It deliberately contains no return figures.

The ten-field checklist to fill in before you put money into a trading bot, with the source of each field
FileContentsLink
Wordto fill in, in Italian

The ten fields with room for your answers and the source of each.

i-bot-di-trading-funzionano-checklist.docx

PDFto print, in Italian

The same checklist, to fill in by hand.

i-bot-di-trading-funzionano-checklist.pdf

N.What to writeWhere it comes from
1

Who sells it and how they earn: commission, subscription, broker affiliation.

CFTC, SEC and FINRA

2

Who authorised it, with the number, checked against the public register.

Consob, FCA, Investor.gov

3

What return it promises. High or guaranteed is the first sign of fraud.

CFTC, SEC and FINRA

4

How many configurations were tried before the backtest you are shown.

Bailey and others, 2014

5

Which period was kept out of the test, never seen.

Bailey and others, 2014

6

Commissions, spread, slippage; on CFDs, the provider's share of accounts losing money.

ESMA, 2018

7

The biggest loss it has had, and what the bot does when it gets there.

My mistake no. 1

8

What it does after a loss: doubles up, raises the leverage, stops.

My mistake no. 1

9

How many days it has been running on prices that did not exist when the rule was written.

My post of 9 October

10

How much you can lose without changing your life, written down before you start.

Common sense

In March 2026 I decided to build my own understanding of finance. From then on I spent 521 hours in the markets: first deciding by hand, then with algorithms, then by hand again with the algorithms alongside. On 22 September I did the sums and wrote them up in public: “Zero euros. 521 hours. A lot of lessons.”

The first lesson came from some salespeople. They were pitching an algorithm, with a commission for themselves, that doubled the stake after every loss, like at the roulette table. It wins almost every time, then loses once and takes everything with it. Field 8 of the checklist comes from there. The second I learnt on my own, at night, with the computer left on testing combinations: the full post lists all ten.

Do trading bots work? What the regulators write

The market regulators do not write that bots don’t work: they write that nobody can promise they will, and that anyone promising it is usually selling something. The clearest warnings are American, from January 2024, and the European ones focus mainly on firms using AI with their clients.

RegulatorWhat they writeWhen
CFTC (United States)

AI cannot predict the future or sudden market changes; high or guaranteed returns are a sign of fraud. It cites Mirror Trading International: more than $1.7 billion in bitcoin taken from at least 23,000 people.

25 January 2024

SEC, FINRA and NASAA (United States)

Be wary of anyone promising returns through AI, even if registered; check the registration on the public registers.

25 January 2024

ESMA (European Union)

For investment firms, responsibility for decisions stays with management, whether a person or an AI tool makes them.

30 May 2024

ESMA (European Union)

On CFDs, in the national regulators’ analyses, between 74% and 89% of retail accounts lose money; each provider must state its own percentage.

March 2018

Consob (Italy)

Blocks unauthorised websites, including one that presented itself as an AI-based investment; flags AI-generated content used for fraud.

19 March and 10 September 2026

FCA (United Kingdom)

Warns that a firm called “Ai Trader Bot” is not authorised and advises people to avoid it.

6 June 2025

The SEC and FINRA alert asks people to be wary of anyone promising that AI can guarantee amazing returns, “even from registered firms and professionals”: being properly registered does not change it. And one figure about people shows how narrow the door is: in a study of Taiwanese day traders from 1992 to 2006, around 360,000 people, fewer than 1% beat the market consistently. They are people, not bots, but a bot trades in the same market.

How a trading bot works

A trading bot is a program that buys and sells on its own by following a rule: when the price does a certain thing, it opens or closes a position. The European MiFID II directive calls it algorithmic trading when an algorithm sets the order parameters with little or no human involvement.

The bot knows nothing the rule doesn’t contain. If the rule has a real edge, it applies it tirelessly and without emotion; if it doesn’t, it loses with the same discipline. So the work that matters all happens beforehand: working out whether the rule has an edge, and how to avoid fooling yourself into thinking you have found one.

AI agents add another layer. An agent is a language model that can read, search and use tools, and someone puts it in charge of placing trades. The model is still an LLM: it produces plausible text, it can make things up, and it knows nothing about tomorrow’s prices. The difference between an agent and a fixed automation is explained on the page about AI agents and chatbots.

Trading with artificial intelligence: the nine agents and the 28 wasted days

On 9 July a friend from university passed me his bot, and we put nine AI agents with different roles inside it: one read the news, one was risk-averse, one the opposite. Before every trade they voted. It looked like a miniature investment committee.

On 6 August we tested it on 1,212 trades from 2010 onwards. On average, each trade made zero. Nine agents, plenty of debate between them, no edge. In the title of the post I called it what it was: 28 wasted days.

Looking back, the reason is simple. A group of models debating over a chart produces opinions, and an opinion on tomorrow’s price is worth as much as anyone else’s. Adding more agents to the vote adds no information: it adds noise that looks like reasoning. Using AI to make the decision is exactly the case the regulators warn about.

Overfitting: the backtest that always tells you you’re right

A backtest runs a rule on past prices, and the more rules you try, the easier it is to find one that worked on the past by chance. Bailey, Borwein, López de Prado and Zhu put it in a 2014 paper: “high performance is easily achievable after backtesting a relatively small number of alternative strategy configurations”.

The problem, they write, is that a backtest is rarely shown with the number of configurations tried before the good one, and without that number nobody looking at it can judge. This is field 4 of the checklist. I fell for it for months. I put it like this: “Try a million combinations and the best one wins by definition. Not because it works: because it was the best of a million.”

There is a newer version of the same mistake, and it concerns LLMs. A model trained on text up to a certain date has already seen, in some form, the very years you might use to test it. A January 2026 study on arXiv, written by a vendor that sells models designed to avoid this, measures exactly that effect. The practical fix is an old one: test on a period the model cannot know.

The method I changed halfway: the rule written first, four years of study, three years kept hidden, attempts counted, thirty days on new prices The rulewritten beforethe results The studyfour yearsof data The taskthe next three years,never seen The counthow many attemptsin total The future30 days onnew prices against falling in lovewhere to lookagainst overfitting the tossed cointhe real exam
How I changed the tests after the first 28 wasted days. Each step guards against a different way of fooling yourself; none of them guarantees a profit.

What I changed: agents as researchers

After the 28 days I stopped letting the agents decide and put them to work on research, with written rules to keep me from fooling myself. They read academic papers. They look for situations where someone is forced to buy or sell, such as a fund that has to rebalance or a company that has to change currency. And a log keeps every idea, including the discarded ones, with the reason.

There are four rules against self-deception, and they are in the diagram above. The rule is written before seeing the results. The study uses four years of data, and the three years after that stay blindfolded, used only as the final test. Attempts are counted, because a coin tossed a hundred times will always throw up a run of heads somewhere. And at the end you look at the future, the one exam that cannot be rigged.

The numbers, as of 9 October 2026

59 days of testing without a single rule passing the final test. Then 32 rules that make money on paper, and from those, 18 bots.

Two of them passed the final test. For the first, the probability of a result like that by pure chance was about 1 in 900. It measures the odds of luck, not the odds of making money.

Then 30 days of simply watching them. In the post I admitted it: “Not even the earlier test really counts. The real exam is the future.” I have published no further numbers since 9 October, and this page adds none.

What holds true regardless: “AI can’t decide on its own. It researches, checks, remembers.” The post of 9 October.

Are trading bots legal? Italy, the European Union, the UK and the US

The rules read for this page concern anyone offering a bot or an investment service to the public, not someone running a program on their own account. In Italy, providing investment services to the public on a professional basis is reserved to authorised firms, and doing it without authorisation is unlawful. For personal use I found no primary source: what follows is not legal advice.

WhereWho may offer the service, and what to checkSource
Italy

Investment services to the public reserved to investment firms (SIMs), banks and authorised firms; Consob blocks unauthorised websites and publishes the list.

Consolidated Law on Finance, art. 18; Consob

European Union

Firms engaged in algorithmic trading must have effective systems and risk controls.

MiFID II, art. 17

United Kingdom

FCA authorisation is required; unauthorised firms end up on its warning list.

FCA

United States

Registration with the SEC and FINRA, which you can check on Investor.gov; the CFTC oversees derivatives and currencies and takes complaints.

SEC, FINRA, CFTC

The wording of the European directive is clear about who it addresses: “An investment firm that engages in algorithmic trading shall have in place effective systems and risk controls”. Selling a bot or signals to third parties, on the other hand, can become an investment service, and then the restrictions above apply.

How to spot an AI trading scam

Three answers settle it before any other check: a high or guaranteed return, a seller who isn’t on the public register, a bot that doubles the stake after every loss. The first two come from the regulators’ warnings, the third from my own first mistake.

On doubling up, arithmetic is enough. Start with 1 euro and double after every loss: after ten losses in a row the next stake is 1,024 euros, and when it finally wins, the profit is still 1 euro. It only works with infinite money and infinite time. Consob, meanwhile, flags a newer risk: AI-generated videos, voices and images used to make the seller look credible.

The practical check is field 2 of the checklist: company name, regulator, number, and a search on the public register. If the seller can’t be found, the conversation is over. If they can, the other eight fields remain.

What I took from it, beyond trading

For me, trading was a training ground for one thing: testing an idea against data without falling in love with it. I wrote it in the post: “You can’t lie to the numbers. You can only choose which ones to look at.” The same rules, the rule written first and the attempts counted, apply to a campaign, a website or a product idea.

The other thing that stays is where to put people. The agents read, search and remember every idea; the decision stays with the person accountable for it, as ESMA writes for firms. In systems it is called human in the loop, and it matters even more when money is involved. The other tests told with numbers are in the Behind the curtain section: for example the theme page with 632,288 likes and what the law says about collecting contacts from social media.

Questions and answers

Do trading bots work?

None of the regulators’ warnings read for this page says a bot makes money on its own, and the US CFTC writes that AI cannot predict the future or sudden market changes. A bot follows a rule: if the rule has no real edge, it just loses money faster.

In my experience, 521 hours up to 22 September 2026, the result was zero euros. This is not financial advice: you can lose all of your capital.

Can you make money with trading bots?

No official source rules it out and none promises it. On CFDs, European national regulators counted between 74% and 89% of retail accounts losing money in 2018. Among the Taiwanese day traders studied from 1992 to 2006, fewer than 1% beat the market consistently. Those figures are about people, not bots, but they show how narrow the door is.

Are trading bots legal?

It depends on who offers the bot, to whom and with what authorisation. In Italy, providing investment services to the public is reserved to authorised firms (Consolidated Law on Finance, art. 18), and Consob blocks unauthorised websites, including ones that present themselves as AI investments. The rules read here concern the provider of the service; for use on your own account, this is not legal advice.

Which AI should you use for trading?

The right question comes before the name. A language model can read news and papers, write code and keep a log of ideas; it does not know what prices will do. And if it was trained on the years you then use to test it, the test is skewed, because it has already seen those prices.

How can I tell if a trading bot is a scam?

Three signs settle it: a high or guaranteed return, a seller missing from the public registers of Consob or of their own country’s regulator, a bot that doubles the stake after every loss. The ten-field checklist on this page covers the rest.

Notes on sources

  1. CFTC, release 8854-24, 25 January 2024, with the customer advisory on AI trading bots and the Mirror Trading International case; the quotation comes from here. Read on 12 November 2026.
  2. FINRA, SEC and NASAA, Artificial Intelligence (AI) and Investment Fraud, 25 January 2024; the same alert on Investor.gov.
  3. ESMA, Public Statement on AI and investment services, 30 May 2024; ESMA, measures on CFDs and binary options, March 2018: the 74-89% figure comes from the national analyses of the time, and covers CFDs only.
  4. Bailey, Borwein, López de Prado, Zhu, Pseudo-Mathematics and Financial Charlatanism, Notices of the AMS, vol. 61, no. 5, May 2014.
  5. Barber, Lee, Liu, Odean, The Cross-Section of Speculator Skill, 2011 working paper, published in the Journal of Financial Markets in 2014: Taiwanese day traders, 1992-2006.
  6. Consob, press release of 19 March 2026 and of 10 September 2026 (in Italian); the up-to-date list of blocked websites is on the blocking page.
  7. FCA, warning on Ai Trader Bot, 6 June 2025. Directive 2014/65/EU (MiFID II), original text, Articles 4 and 17; the directive has been amended since 2014. Consolidated Law on Finance (Legislative Decree 58/1998), art. 18, on Normattiva.
  8. Benhenda, Look-Ahead-Bench, arXiv, January 2026: a study by an author linked to a model vendor.
  9. My numbers come from the two LinkedIn posts of 22 September and 9 October 2026, and stop there.
·Behind the curtain

Trading is just the frame here. The method is the same one I use for everything else.

This page is part of Behind the curtain: the things I have actually tried, with the numbers, even when the numbers say zero. Mattia Esposito, founder of Itria AI.