Individual investors and independent researchers have spent two years wiring language models into stock and crypto trading. We traced 25 of their results to the source and scored each one on real money, real benchmarks, and whether anyone besides the author checked the work.
On October 18, 2025, a startup called Nof1 gave six leading AI models $10,000 each in real money and let them trade crypto futures. Each model got the same market data and the same prompt, and no human touched the trades. Sixteen days later, two were up. GPT-5, the model most people reach for when they ask a chatbot about stocks, had lost roughly 60% of its stake, and it finished last in every published version of the results.
That contest, Alpha Arena, is the cleanest real-money evidence anyone has published about language models making their own trading decisions. It also sits at the far end of the record from almost everything else: the academic papers, open-source frameworks and Substack experiments where the same kinds of models appear to beat the market, sometimes by a lot.
Two years in, the evidence splits along one line. Where a model reads and scores information and someone tests the result carefully, the signal is real and modest. Where a model is handed the trigger and a funded account, the record is short, noisy and mostly red. The retail pitch lives in the gap between those two findings.
We traced 25 results to their sources and scored each on three questions: was real money at risk, did the result beat its own benchmark, and did anyone other than the author check the work. Several widely repeated claims failed that tracing and are listed near the end, with what went wrong.
Season 1 of Alpha Arena ran from October 18 to November 3, 2025, on the Hyperliquid exchange. The six models traded perpetual futures on Bitcoin, Ether, Solana, BNB, Dogecoin and XRP, and every trade was public. Qwen3 Max, from Alibaba, and DeepSeek V3.1 finished in profit. Claude Sonnet 4.5, Grok 4, Gemini 2.5 Pro and GPT-5 all lost heavily.
Swipe the chart sideways to see the full range.
Even the scoreboard wasn't settled. Three outlets published final standings that differ by as much as 11 points for the same model, most likely because each captured the leaderboard at a different moment while positions were still open near the close. Nof1's own site refused every automated request we made for the primary figures, so the chart shows all three tallies instead of quietly picking one.
Qwen3 Max's gain of about 22% beat simply holding Bitcoin, which rose from roughly $106,500 to $110,500 over the window by CoinGecko's daily prices, under 4%, and gave most of that back the following day. DeepSeek cleared the same bar by about a point. It was also one 16-day window, in one of the most volatile markets there is, with six contestants, which is nowhere near enough to separate skill from luck.
Nof1 ran it again in late November and early December 2025, this time on US stocks, with more models and four different rule sets. Grok 4.20, a model xAI had not yet released, finished up roughly 10 to 12% and was reported as the only model to end in profit (ForkLog, Dec 8, 2025; Benzinga via Yahoo Finance, Jan 18, 2026). Across two seasons and two asset classes, the model that won each round was a different model.
None of this started with trading bots. It started with a finance professor asking GPT-4 whether a headline was good or bad news for a stock, and it took less than three years to get from there to agents that can place real orders in a brokerage account.
Red dots mark events where real money was on the line. GitHub star counts as of September 12, 2026.
Nearly every system in this record is wired one of five ways, running from a single prompt to an agent holding a funded brokerage account. Most of the GitHub popularity sits in frameworks whose own documentation tells you not to trade real money with them.
| Wiring | How it works | Example in this record | What it's asked for |
|---|---|---|---|
| Single prompt | Hand the model one headline or filing and ask for a score or rating | Lopez-Lira and Tang; GPT-4 stock ratings | Next-day direction, monthly ratings |
| Fine-tuned model | Train a model on news and the returns that followed | Guo and Hauptmann; FinGPT's news-spread model | Return ranking, price direction |
| Retrieval over filings | The system pulls relevant passages from 10-Ks, earnings calls and news before judging | MarketSenseAI 2.0 | Monthly stock picks |
| Multi-agent debate | Specialist agents argue the bull and bear case; a trader and a risk team decide | TradingAgents; ai-hedge-fund | Buy, sell or hold |
| Agent with brokerage access | The agent researches, then places orders through a broker's API or an MCP server | Alpha Arena; Robinhood Agentic Trading; Alpaca-based stacks | The trade itself |
MCP (Model Context Protocol) is the open standard that lets an AI assistant call outside tools, including a brokerage account.
TradingAgents, from Yijia Xiao, Edward Sun, Di Luo and Wei Wang, is the design most retail builders copy. It splits one decision across five stages of AI agents, each reading the previous stage's work.
The debate structure reads like prudence, and it may be. It also carries a risk that has nothing to do with markets. A 2026 theory paper by Maya Okawa, written about AI debates in general rather than trading, found that groups of AI agents can converge on a shared bias instead of the right answer, with the models' sampling randomness driving it. A bull and a bear built on the same model may agree with each other more than their names suggest.
That last gap is where AI Stock Market Impacts works, on purpose. Our engine doesn't read today's headlines to guess tomorrow's price. It scores how AI shifts relative value across 31 industries and 204 cross-industry effects over one to ten years, which is a slower question and a less crowded one. A signal that thousands of identical agents can find by reading the same news wire at the same moment is the kind of edge that disappears fastest; we wrote about that loop in The Ouroboros Trade.
The five-stage agent firm above. Runs on a simulated exchange.
A team of AI agents that produce trading decisions from market data pulled through the Financial Datasets API.
An agent platform for financial research, forecasting and report writing, from the team behind FinGPT.
Add-on skills that give Claude Code market analysis, screeners and chart reading.
Claude Code as an autonomous trader, able to place orders through Alpaca's brokerage API.
A maintained index of open-source trading agents, MCP data and execution servers, and agent skills.
Star counts from the GitHub API, September 12, 2026.
Every row below is a result someone published: a peer-reviewed study, a preprint, a benchmark, a real-money contest, or a personal experiment. Pick the factors you care about, hide the columns you don't, and click any column header to sort. The Venn diagram and the chart under the table follow the same filters.
Swipe the table sideways to see what happened, the verdict and the source.
"Beat benchmark" means the headline result beat the comparison its own authors chose, which ranges from buy-and-hold to a sentiment baseline. ± marks a split result. "Who checked it": peer review, or public trade logs anyone can audit, count as an outside check; preprints and personal posts do not.
Small numbers, read them as a pattern rather than a rate. Peer-reviewed studies mostly test whether a signal exists, not whether a real account made money, which is part of why they score higher than the public real-money contests.
Lopez-Lira and Tang had GPT-4 score news headlines for 4,123 US stocks between October 2021 and May 2024. The model called the market's initial reaction on about 90% of portfolio days. That first reaction happens too fast for an ordinary investor to trade, but the scores also predicted the drift that followed, and a simulated long-short strategy built on that drift made money. The effect was strongest in small stocks and bad news, and it grew with model size; older, smaller models showed none. The paper passed peer review at the Journal of Financial Economics, one of the two or three most demanding journals in the field.
MarketSenseAI, from George Fatouros, Kostas Metaxas and colleagues, runs separate agents for news, fundamentals, price action and macro, and reads SEC filings and earnings calls before picking stocks. Its 2023–24 backtest returned 125.9% against 73.5% for the S&P 100, after charging 10 basis points per trade. In April 2026 two of its authors published the system's out-of-sample record, built from signals generated live: an S&P 500 cohort beat the index by a compound 25.2% over 19 months, a statistically significant result (p = 0.003). An S&P 100 cohort beat its index by 30.5% over 35 months, but with only about ten picks a month that result did not reach statistical significance.
Sources: arXiv:2502.00415; arXiv:2604.17327. Both are preprints, and the out-of-sample evaluation was done by the system's own authors.
StockBench, which tested 14 models as trading agents on 20 Dow stocks from March 3 to June 30, 2025, found that most of them beat buy-and-hold in its main test, with Moonshot's Kimi-K2 ranked first. A 2025 preprint on the options "wheel," a strategy of repeatedly selling puts and covered calls, reported 15.3% a year with a Sharpe ratio (return per unit of risk) of 1.08 against 0.62 for its benchmark, across a backtest of nearly 19 years.
The personal experiments look good too. Robert Wade Thompson put $5,000 of real money into five stocks ChatGPT chose in January 2025 and matched it with $5,000 in an S&P 500 fund. A year later the ChatGPT basket was up 23.2% and the fund 20.5%. The five picks were Microsoft, Nvidia, Costco, Eli Lilly and JPMorgan Chase, a basket of household names that looks a lot like the index it beat, and it fell harder along the way: 14.1% at its worst point, against 9.3% for the fund.
"Our results on paper are much more optimistic than what the performance in reality would be with a reasonable investment size."Alejandro Lopez-Lira, co-author of the ChatGPT forecasting paper, to Morningstar, June 2025 (quoted by Futurism)
That sentence comes from the author of the most-cited positive result in the field, and it describes most of the scorecard.
Several claims circulate widely in this space, and some reached us from AI research tools while this page was being prepared. Each one below failed when we went to the source.
No US regulator licenses or bans the use of a consumer chatbot for trading decisions. The posture so far is investor education. FINRA's guidance on AI and investing, published January 25, 2024, warns that "AI-generated information might rely on data that is inaccurate, incomplete, or misleading", that it "can be faulty, or even completely made up," and that chatbot conversations "could lead you to make misinformed, emotionally-driven or impulsive investment decisions."
The brokers' own product design says something too. Robinhood's Agentic Trading confines the agent to a separate account the customer creates and funds, and the customer stays responsible for every trade it places.
The reading part is real. Language models extracting what news and filings mean, at a scale no analyst team could match, is a documented advance, and it has survived peer review. The deciding part has not earned trust with real money yet. The one fully public real-money test went badly for four of six models; the rematch had a single winner, and it was a different model. The best academic trading results are backtests, or evaluations run by the people who built the system.
The gap is widest exactly where the retail pitch aims: an agent with a funded account, running a strategy copied from a popular repository, reading the same news as every other agent built on the same handful of models. I'm confident about that mechanism, because Lopez-Lira and Tang measured the edge shrinking as it spread. I'm not confident about where the models go from here. MarketSenseAI's S&P 500 result is statistically significant and could keep holding, the models change every few months, and the whole record is barely two years old. Whether an autonomous agent can beat a plain index fund with real money over a full market cycle is still an open question, and the evidence on it is thin in both directions.
Our 31-industry AI-effects matrix measures how AI shifts relative value across industries and 204 cross-industry effects over one to ten years. It makes no price calls and no trades, and every parameter change since February 2026 is logged in public.
See the 31-industry matrixRead it alongside The Ouroboros Trade, on what happens when everyone's AI reads the same news, and Why Math-Only AI Stock Predictors Will Always Fail.