AI-run funds have existed since 2017, and researchers have tested language models on stock news. Here is what the track records show, where AI genuinely helps investors, and how to judge an AI stock-picking claim.
The public evidence so far says no, not consistently. AI-driven funds have had good stretches and bad ones, and over their full histories the best-known one trailed the S&P 500 index. That is the same result most human stock pickers get: according to S&P Dow Jones Indices’ SPIVA scorecard, most active large-cap funds trail the index over long periods. A new tool does not change the underlying problem. Prices already reflect what most participants know, and when everyone uses similar tools, their edge spreads thin.
That does not make AI useless for investors. It means the useful question is not “can AI predict prices?” but “which parts of investing does AI do better than a person with a spreadsheet?”
The longest public record belongs to the AI Powered Equity ETF (AIEQ), launched in October 2017 to pick US stocks with an AI model; today it seeks to track the AI Powered Equity Index. Its results are published, which makes it a fair test.
| Checkpoint | AIEQ | S&P 500 | Source |
|---|---|---|---|
| Launch (Oct 2017) to May 2021 | About +64% | About +75% (total return index) | Bloomberg, via WealthManagement.com |
| Launch to Sep 30, 2026 | 9.61% a year (NAV) | About 14.4% a year (SPY, end-2017 to Sep 2026) | Amplify ETFs; Exent |
The second comparison uses a near, not identical, window, but the gap is wide either way. Sources: Amplify’s AIEQ fund page and Bloomberg’s 2021 report on the fund. The fund also charges a 0.75% expense ratio, against an asset-weighted average of 0.14% for index equity ETFs in 2025 according to the Investment Company Institute, which explains part of the gap.
No. Buying companies that build AI is a sector bet; letting AI choose stocks is a method. The two get mixed up in headlines. In Exent’s data to Sep 30, 2026, a robotics and AI theme fund (BOTZ) returned 9.1% a year over 10 years against 15.2% for SPY, while a generative AI fund (CHAT) returned 51.9% a year over 3 years. Same theme, very different outcomes, both driven by which stocks the fund held and when. Neither result says anything about AI’s ability to pick stocks. More on theme funds in thematic ETFs.
The most cited study is Lopez-Lira and Tang, “Can ChatGPT Forecast Stock Price Movements?”. The authors asked language models to score news headlines as good or bad for a company. In the latest version, GPT-4’s scores captured the initial market reaction to headlines published after its training cutoff and also predicted some of the drift that followed, especially for small stocks and bad news. Simpler, older models could not do this.
Two findings matter as much as the headline. First, much of the reaction happened too fast to trade on. Second, the returns of the strategy declined as more investors adopted language models, which the authors read as markets becoming more efficient. That is the usual life cycle of an edge: once a method is public and cheap, its returns shrink. See alpha decay.
You can ask it for ideas, but a general chatbot is a weak stock picker on its own. It writes fluent answers whether or not the facts behind them are current, it can state numbers it has not calculated, and it does not see your portfolio, your risk or what you already own. A stock idea is only good or bad in the context of the rest of the portfolio. The trade-offs are covered in this comparison of ChatGPT with a portfolio tool, and how AI ratings compare with human analysts in analyst vs AI stock ratings.
A practical rule is “quant for math, AI for words”. Calculations about risk and return should come from quantitative models run on market data, where every number can be reproduced. Language models are best used for the work that is mostly reading and writing:
| Task | Better done by | Why |
|---|---|---|
| Volatility, Sharpe ratio, drawdown, correlation | Quantitative models | Exact arithmetic on price data; reproducible |
| Crash replays and optimization | Quantitative models | Needs full return histories, not text |
| Reading a screenshot of holdings | AI | Turning an image into a clean list is a language and vision task |
| Summarizing news and filings | AI, with sources | Fast reading; the sources let you check it |
| Explaining what a result means | AI | Plain-language explanation of numbers the models produced |
| Predicting next month’s price | Neither, reliably | The evidence above |
A sensible split: let quantitative models do the calculations, and use AI to read, summarize with sources and explain. Be wary of any tool that advertises market-beating returns. A more realistic aim is to measure your portfolio correctly and improve the return you get for the risk you take; see risk-adjusted outperformance.
Some have over short stretches. The longest-running public AI stock-picking ETF, AIEQ, returned 9.61% a year from its October 2017 launch to Sep 30, 2026, which trailed the S&P 500 over a similar window.
Research by Lopez-Lira and Tang found that GPT-4 can read news headlines and anticipate short-term price reactions, but much of that reaction was too fast to trade and the strategy’s returns declined as more investors used similar models.
There is no public evidence that AI picks consistently beat the market or top analysts after costs. Judge any source, human or AI, by its live, risk-adjusted track record against a fair benchmark.
Reading and summarizing large amounts of text, such as news, filings and earnings calls, turning screenshots into data and explaining results in plain language. Risk and return calculations are better done by quantitative models.
Nobody can say. Historically, when a method becomes widely available its excess returns shrink, and research on language models and news already shows that pattern.
Market Intelligence reads the market regime and tells you whether the backdrop favors taking risk or protecting capital.
Educational content, not investment advice. Examples are illustrative; past performance does not guarantee future results.