What it takes for an AI to trade well

A person researching a stock might spend days on filings, news, charts and opinion, then decide with judgment built over years. An AI model brings a different kind of power to the same decision, and it is easy to describe that power loosely. So we measured one decision.

At 9:45 a.m. Eastern on October 13, 2026, Mellifera (Melli), the decision agent behind our flagship swing swarm, made her scheduled morning decision. Her decision log records exactly what she was given: 158,797 characters in the first pass, and another 162,452 when she asked for a second look at 13 more symbols. That is about 80,000 tokens, or roughly 55,000 words. A person reading at 250 words a minute would need about 3.7 hours. She took 3 minutes 37 seconds, then placed three orders: a stop order on her Apple position, and new positions in UnitedHealth and Goldman Sachs.

Two things stand out. Most of what she reads is structured data, not prose: option chains alone are 48% of the first pass. And none of it is raw web text. Each scout finding arrives dated, cited and scored for how well it is sourced, and prices come only from the market feed. She receives the filtered result of other agents' searching, not the firehose.

A person and a model, on one scale

Per decision, she reads about what a diligent analyst reads in an afternoon. The difference is everything around that afternoon. She reads it in minutes, on a schedule, without fatigue, and she starts from training text roughly 15,000 times larger than everything a person reads in a lifetime.

Raw processing power is a less useful comparison than it sounds. Estimates of the brain's compute span four orders of magnitude, and a brain runs on about 20 watts, roughly 0.8 kWh over a 40-hour research week, far more energy than one model decision uses. What the model adds is breadth, speed and consistency. What it lacks is judgment about what matters. The next section shows what happens without it.

Alpha Arena, run by the research lab Nof1, gives frontier models $10,000 each and lets them trade. In the round we studied, eight models each traded under four sets of instructions: a baseline, a cautious “monk mode”, one that gave them more context about the competition, and one that pushed maximum leverage. Typical positions were held for minutes to a few hours at about 10x leverage.

The 32 accounts started with $320,000 and finished with $207,420, a loss of 35%. 26 of 32 lost money, 13 lost more than half, and the worst lost 96%.

Three findings matter for anyone building an AI trader.

They traded far too much

The 25 accounts with full trade data averaged 529 trades each, and up to 976. More trading went with worse results (a correlation of −0.40), and fees of $33,299 more than doubled the losses: before fees those accounts lost $28,842, and after fees $62,141.

Win rate didn't separate winners from losers

Every account won between 27% and 41% of its trades, profitable or not. What separated them was the size of their wins, and the next test shows how little of that was skill.

The winners were carried by one trade

Take away each profitable account's single best trade, and 5 of the 6 finish in the red. We ran the same test on our flagship swarm's last 30 days. It stays positive.

The fastest way to lose an investor's trust is a blown-up account. In six months on broker paper accounts, none of our three swarms has come close. The worst is down 2.7%. The flagship's deepest drawdown was 7.7%, smaller than QQQ's 11.3% over the same dates, and its worst single trade cost 1.6% of the account.

This is not luck or an unusually timid model. The limits below are checked in code when an order is placed, outside the model. She can argue for a trade. She cannot argue past these.

She is also never asked to grade herself. Our scouts used to rate their own findings, and gave 92 to 99% of them a confidence of 0.8 or higher, which told us nothing. Confidence is now computed from the evidence behind a finding, so a newswire item outranks an undated web page whatever the model thinks of it.

SwarmTrade has run autonomous swarms since March. Most of what we know came from finding out, often the hard way, where a capable model still goes wrong. Six lessons shaped the platform.

Every release is listed in the changelog.

Three swarms trade on Alpaca broker paper accounts with public records. The flagship is a swing swarm that holds stocks and options for days to weeks. The other two trade options intraday.

The flagship against the market

Since April 7, 2026 the flagship has returned +26.9%, against +18.1% for SPY and +26.2% for QQQ. It matched the Nasdaq-100 with a shallower drawdown (−7.7% against −11.3%) and beat the S&P 500 by about 9 points. Its edge is recent: over the past 90 days it returned +20.9% against +3.1% for SPY and +3.5% for QQQ. For most of the months before that, it trailed both.

We hold ourselves to the test we applied to the Arena. In the last 30 days the flagship closed 9 trades, 7 of them winners, for a realized +$7,367. Its best trade, an AMD call, made $4,775 of that. Without it, the month is still +$2,592.

Over its whole life the flagship has closed 584 trades. It wins 43% of them, and its average win ($178) is nearly twice its average loss ($93), a profit factor of 1.45. That is the shape of a disciplined swing strategy: many small losses cut early, and fewer, larger gains.

A research page that only shows the good numbers isn't research. These are the limits of the record today.

Software is no longer a moat: a capable team with AI tools can rebuild features quickly. A long, verifiable track record can't be rebuilt quickly, so we are building ours in public.

• The strategy farm. 58 swarms have public records today: 3 on broker paper accounts and 55 simulated against real quotes. Losers stay on the leaderboard.
• Decisions that can't be quietly edited. Farm decisions are published to a hash-chained log the moment they are made. Each entry's hash covers the one before it, so any edit, deletion or reordering breaks the chain, and anyone can check it.
• Graduation. Strategies move from simulation to broker paper accounts, and then to real money, starting with our own capital, only after they earn it.
• Better data for published records. Next is full options-tape quotes for the accounts whose records we publish.

The goal is not an AI that replaces the investor. It is an investor who sets the mandate and gets a research desk and an execution desk that follow it every day, with rules they can see and a record they can check.