Sometimes, in narrow and well-defined tasks, but the public evidence says general-purpose AI trading agents mostly do not beat simple strategies. In a live contest in late 2025, six leading AI models each traded $10,000 of crypto with no human help for 17 days; four of them lost more than 40%. Academic studies from 2025 and 2026 point the same way: small edges at best, and often worse once trading costs are counted.
What does the public evidence actually show?
The best public evidence shows mixed results, with a few winners, many losers and small margins either way. Here are five of the most useful data points, from live contests to academic benchmarks.
| Evidence | Set-up | What it found |
|---|---|---|
| Public live contest, late 2025 | Six leading AI models, $10,000 each, crypto futures, 17 days, no human intervention. | Two finished up, by about 22% and about 5%. Four lost more than 40% each; the worst lost more than half. |
| An October 2025 academic stock-trading benchmark | 13 models trading 20 large US stocks daily for 82 trading days, March to June 2025. | Eight of the 13 edged past buy-and-hold’s 0.4% return; the best made 2.5%. Even so, the authors’ summary says most agents struggle to beat buy-and-hold. |
| A November 2025 live-trading benchmark | 21 AI models over 50 days of live US stock and prediction-market data. | High scores on general AI leaderboards did not predict better trading. |
| A May 2026 reproduction study | Two published research trading agents re-run over one year on five stocks, with real costs. | Both ended below buy-and-hold after costs, which beat them on four of the five stocks. |
| A crypto VC report, April 2026 | AI agents across decentralised finance. | A yield agent earned 9.75% a year; in open-ended trading, agents lagged humans by up to 5 to 1. |
Buy-and-hold means simply buying the assets at the start and doing nothing. It is the bar any active strategy has to clear, and it costs nothing to run.
The direction is consistent, even where margins differ. Agents can occasionally post a good run. What they have not yet shown, in public, is a reliable edge over doing nothing, after costs, over a long period.
Why do general-purpose AI models struggle at trading?
General-purpose AI models are built to reason in language, and markets are mostly numbers, noise and other traders adapting to you. Several specific problems show up in the evidence.
They are weak with price data. The organiser of the late-2025 contest said the models do not handle numerical time-series data very well, which was the main input they were given. They also had a narrow set of assets and a limited set of possible actions.
They have habits. The same organiser reported that each model showed consistent biases across many prompt variations, much like an investing personality. A model that leans aggressive stays aggressive, whatever the market does.
Knowledge is not skill. The October 2025 benchmark found that strong results on financial knowledge tests did not translate into successful trading. Knowing what a moving average is and trading well are different abilities.
Costs eat thin edges. In the May 2026 reproduction study, one agent’s Sharpe ratio, a measure of return per unit of risk, roughly halved from 0.43 to 0.22 once fees, spreads and market impact were included. On one stock, its return went from a 2% loss before costs to a 10% loss after them.
Backtests can flatter. The same study warned that current public evidence cannot yet separate real predictive ability from contamination, where a model has already seen the test period in its training data. It also found that across five published frameworks, 35 of 40 combinations of system and real-world cost were left unmodelled.
Where do agents actually do better?
Agents do better where the task is narrow, the rules are clear and the goal can be measured. Yield routing is the clearest example in the data.
The April 2026 crypto VC report found that a stablecoin yield agent, which moves funds between lending platforms to chase the best rate, earned users 9.75% a year. That beat the rates on the platforms it compared against. The same report estimated that autonomous agents drive more than 19% of on-chain activity, most of it narrow bot work such as routing stablecoins.
The contrast is sharp. In open-ended trading, the report found agents lose to humans by up to a 5-to-1 margin, and in one stock-trading contest the top human beat the top agent by more than five times. An investment associate at the firm behind it summed up the pattern: agents struggle when the situation is not clearly defined.
How should I judge an AI trading agent’s performance claim?
Judge it the way a careful fund analyst would: ask for the evidence that would be hard to fake. The May 2026 reproduction study proposed minimum reporting standards for exactly this reason. These questions follow the same logic.
- Was it live or a backtest? Live results with real money are far stronger evidence than a simulation of the past.
- Is it net of all costs? Fees, spreads, funding and slippage, the gap between the expected and actual price, can turn a gain into a loss.
- How long, and how many trades? A few weeks of good returns is mostly noise. The same study notes that over a year of daily data, the statistical uncertainty can be wide enough to cover the whole gap between an agent and buy-and-hold.
- Compared to what? Ask how it did against simply holding the same assets.
- What was the worst drawdown? A strategy that rose 20% after first falling 50% is not one most people could live with.
- Did the test start after the model’s training data ended? If not, the model may have seen the answers.
If a seller cannot answer these, treat the performance claim as marketing.
What could go wrong?
The biggest danger is not a real agent that underperforms. It is a fake one that promises not to. US regulators have warned about this directly.
In January 2024, the CFTC’s advisory on AI trading scams warned that fraudsters use public interest in AI to sell trading bots and signal services with unrealistic returns. It stated plainly that AI cannot predict the future or sudden market changes. It described one alleged scheme that took more than $1.7 billion in bitcoin from at least 23,000 people, promising at least 10% a month from a trading bot, while very little was actually traded.
The advisory’s red flags are worth memorising: guaranteed returns, claims of huge returns or 100% win rates, referral bonuses, and hype from influencers or strangers online. A real agent cannot honestly promise any of those.
Real agents carry their own risks too. The late-2025 contest traded perpetual futures, a product that lets traders use leverage, or borrowed money, and leverage turns a model’s bad habits into large losses quickly. Our inference is also that one good season proves little: with enough agents competing, some will post strong results by chance.
The trust takeaway
AI trading agents can make money in narrow, rule-based jobs, and occasionally in short live runs. The broad claim, that an AI can out-trade the market for you, is not supported by the public evidence as of November 2026. Anyone selling certainty is selling something else.
Before you trust any AI trading agent with money, check:
- Proof: Is there a live track record, net of costs, over months rather than weeks?
- Baseline: Did it beat simply holding the same assets?
- Downside: What was its worst drawdown, and could I live with it?
- Red flags: Does anyone promise guaranteed returns, high win rates or referral rewards?
- Permissions: If it underperforms, is the damage limited to a small, trade-only account?