Bloomberg published a report on how programmers — and even enthusiasts with no technical background — are training AI agents to trade stocks, crypto, and prediction market contracts. Journalists concluded that, so far, these experiments are running into the fundamental limitations of language models themselves.
The $100,000 Experiment
A 29-year-old U.S. developer spent two and a half weeks teaching Claude his own trading logic, including risk levels, entry points, and position sizing. He then deployed the agent on a simulator tied to the Alpaca brokerage platform with a virtual $100,000 portfolio.
Over 30 days, the bot returned about 7%, compared with 4.5% for the S&P 500 over the same period. But the ride was far from smooth: drawdowns hit 22%, and the agent repeatedly gravitated toward blue-chip stocks instead of riskier trades, forcing its owner to intervene manually.
One story stood out in particular: when Nvidia shares suddenly surged, the agent debated itself and decided not to chase the rally. The call proved correct — buying at the top could have cost the portfolio roughly $10,000.
Caution Is a Built-In Limitation
Language models are trained on enormous volumes of financial text and tend to reproduce a mainstream view of investing. Traders looking for something beyond the usual buy an index ETF approach have to deliberately push these systems beyond their conservative bias — and even then, results are far from guaranteed.
Meanwhile, exchanges are eagerly embracing the trend. Polymarket, OKX, Bybit, and Kraken have already launched interfaces optimized for bots, while X and Telegram have spawned an entire genre of posts claiming 1,000% returns in just days. One such post racked up 4.7 M views before another AI-run account debunked it. Some of these viral posts lead directly to malware.
Why AI Performs Poorly Where Edge Matters Most
An agent that scrapes Google and bets on the most likely outcome does not create new knowledge — it simply recycles information already available online. If too many bots operate this way, a prediction market stops functioning as a forecasting tool and turns into an echo chamber of consensus opinion.
The article’s main subject also tested his agent on Kalshi, giving it $30 to bet on sports events — with no success. Bitcoin price-range predictions worked slightly better, with about 60% accurate trades, but in the end, the bot lost that money too.