In May 2026, eight of the world's leading AI models got handed $10,000 each and set loose trading actual stocks against each other. Not a demo. Not a simulation with a marketing team behind it. It happened in public, documented by Bloomberg, with prices and outcomes anyone could check.
The portfolio lost about a third of its money. Across 32 total results, something finished in profit six times.
If the smartest models built by the biggest labs on earth can't reliably trade, what exactly is the $40 no-code bot you found on a forum supposed to be doing better?
Here's a harder number. A research collaboration between UC Berkeley and a blockchain analytics firm studied user data on a trading platform and found something specific. Bots lost 77 times more money per user than the humans on that same platform.
Not because the bots were unlucky. Some of them were never built to win at all.
This report builds the full mechanism from there. What a bot is once you strip the word "AI" off it. Why a backtest that looks great can be a curve fit to noise that already happened, nothing more. There's an original simulation showing exactly how a strategy tuned to the past falls apart once it meets the present. And the due-diligence framework for reading any bot's pitch the way its seller hopes you won't.
All of it below, with the exact page where each piece lives.
A Trading Habits Report
The AI Trading Bots Report
What a retail trading bot is mechanically, what happens when a rule tuned to the past meets a market that keeps moving, and what the research says about who gets paid.
- Length 20 pages, with 11 original charts and a 20,000-run simulation
- Author TradingHabits.com
- Format PDF, delivered as an instant download right after checkout
- Covers Bot mechanics, what "AI-powered" means once you open the hood, sourced research on bot losses, overfitting and cost drag, and an original backtest-into-live simulation
No bot vendor, signal service, or platform is named, ranked, compared, or recommended anywhere in this report, and no stock or ticker appears anywhere in it. Everything here is mechanics, sourced research, and the math.
What's Inside
20 Things This Report Says
- 01What a trading bot actually is mechanically, once you strip the word "AI" off the label.Page 4
- 02Why "AI-powered" and "rule-based script" are not the same claim, and how to tell the difference.Page 5
- 03How the technical skill required to build a bot collapsed to close to zero between 2016 and 2026.Page 6
- 04Why lowering the cost of trying an idea means more untested ideas get run with real money, not fewer.Page 6
- 05The search-interest and market-size numbers behind this year's bot boom, not the marketing version.Page 7
- 06What happened when eight frontier AI models actually traded real contests in May 2026, fully public and documented.Page 8
- 07The one detail buried in that contest's results that mattered more than any single trade.Page 8
- 08The research finding behind bots losing 77 times more money per user than humans, and precisely what platform it does and doesn't describe.Page 9
- 09The specific, non-obvious reason some of those losing bots were never built to win in the first place.Page 9
- 10The overfitting mechanism that turns a great-looking backtest into a losing live account, shown with an original diagram.Page 10
- 11Why a small, genuine edge can still turn negative once ordinary trading costs get applied honestly.Page 11
- 12How this report built its own simulation from scratch: 20,000 strategies, the same rule running in a backtest window and then live.Page 12
- 13The exact share of those simulated strategies that looked profitable in backtest against how many stayed profitable live.Page 13
- 14Why "removes the emotion" is a real advantage that still has nothing to do with whether the underlying idea has an edge.Page 14
- 15The 1990s research on retail trading performance that predates all of this technology, and the actual numbers behind it.Page 15
- 16What two separate day-trading studies, decades and continents apart, found about the odds of making money at all.Page 16
- 17Four specific claims worth treating as a stop sign before trusting any bot's marketing.Page 17
- 18A six-question checklist built to be asked before paying for any bot, signal service, or subscription.Page 18
- 19Seven terms defined once, so the rest of the report doesn't keep stopping to explain itself.Page 3
- 20Where every number in this report comes from, with the exact scope of each source stated plainly.Page 19
HABITS
★
Certificate of Guarantee
60-Day, No-Questions-Asked
If The AI Trading Bots Report doesn't earn its place on your desk, email us any time within 60 days of purchase for a full refund. No form to fill out. No reason required.
Instant Download
$9One-time payment. No subscription.
Order Now20-PAGE PDF · DELIVERED IMMEDIATELY AFTER CHECKOUT · 60-DAY GUARANTEE
Behind The Report
When Frontier AI Models Actually Traded
A Public Test, Not A Sales Page
In May 2026, the startup Nof1 ran a contest called Alpha Arena. Eight leading AI systems, including models from Anthropic, Google, OpenAI, and xAI, were each given $10,000 and set loose trading U.S. tech stocks over two-week windows, four separate contests in total. Bloomberg covered the results.
Across all 32 individual outcomes, a model finished in profit six times. The combined portfolio finished down roughly a third of its starting capital.
The Detail Past The Scoreboard
One model placed a small fraction of the trades a rival placed under the identical prompt, and still came out ahead of it. Trading more was not the same thing as trading better, even for systems built by the best-funded labs in the world.
No model, vendor, or bot product is named as a recommendation anywhere on this page or in the report. The full mechanism behind results like these, and the original simulation built to demonstrate it, is in the report itself.
The Alpha Arena Contest
A simplified rendering of the Alpha Arena result, matching the sourced chart on page 8 of the report. Background only. The report itself covers the mechanism, the research, and the simulation.
Try It: What Beating This Win Rate Takes
Default matches Alpha Arena's own result, a model finishing in profit 6 times out of 32. At that win rate, the average winning trade has to be more than four times the average losing trade just to break even, before counting any fees.
Background only. Nothing on this page or in this report is a recommendation to use, avoid, sign up for, or pay any specific trading bot, signal service, or platform.
Common Questions
What was Alpha Arena?
A May 2026 contest run by the startup Nof1. Eight leading AI systems, including models from Anthropic, Google, OpenAI, and xAI, each got $10,000 and traded U.S. tech stocks over four separate two-week windows. Bloomberg covered the results.
How many of the 32 runs actually made money?
Six out of 32. The combined portfolio across every model finished down roughly a third of its starting capital.
Did trading more often help?
No. One model placed a fraction of the trades a rival placed under an identical prompt and still came out ahead of it. Volume wasn't the edge.
Does this report recommend a specific bot or platform?
No. No model, vendor, or bot product is named as a recommendation anywhere on the page or in the report.