Machine learning finds real patterns in stock returns and language models read the news fast. Here is what the peer-reviewed research adds up to, where AI forecasts break, how a fund picked by AI has done since 2017, and a test on real market data you can repeat.
November 16, 2026
19
min to read
The short answer
Not reliably. Machine-learning models find real but small patterns, and language models read news fast. But the predictable part is tiny, fades as more investors use it and can vanish once trading costs are counted, and shocks like the COVID-19 crash had no precedent to learn from. Test any AI pick on data it never saw.
It depends on what you want it to predict. “The stock market” can mean tomorrow’s direction, next year’s return, which stocks will beat the others, or the next crash, and the research gives a different answer for each. Predicting tomorrow’s direction is also the day trader’s basic problem; our guide to AI day trading covers what the research on human and AI day traders found.
| What you want predicted | What the evidence says | Where it comes from |
|---|---|---|
| Up or down tomorrow | Close to a coin flip. On the S&P 500 perp, a rule that always said “up” was right 53.1% of the time from March to November 2026 | Our test on Hyperliquid data, below |
| The whole market’s return over the coming months | Most published predictors fail on new data; a few still work | Goyal, Welch and Zafirov, 2024 |
| Which stocks will beat other stocks | Machine learning finds real but small patterns, with much of the profit in hard-to-trade stocks | Gu, Kelly and Xiu, 2020; Avramov, Cheng and Metzker, 2023 |
| How the market reacts to news | Language models read headlines well, but the first reaction comes too fast to trade, and the edge in the slower drift shrank as more investors used them | Lopez-Lira and Tang, 2026 |
| A crash or a shock | No. COVID-19 hit the stock market harder than any earlier pandemic, so the data held nothing like it | Baker and coauthors, 2020 |
Behind every row is the same idea, the efficient market hypothesis: prices already reflect what investors know, so the next move depends mostly on news nobody has yet. Markets aren’t perfectly efficient, and AI is good at finding the cracks. But a crack that many people trade tends to close, and the biggest moves come from surprises. The US derivatives regulator, the CFTC, put the limit in one line in a January 2024 advisory: “AI technology can’t predict the future or sudden market changes.” What AI trading can and can’t do covers this limit across every kind, not just prediction.
Finance researchers have tested machine learning on stock returns for years, in peer-reviewed journals, with decades of data. The results split into two camps, and both are right.
The two camps meet in the middle. Theis Ingerslev Jensen, Kelly and Lasse Heje Pedersen found that most published factors, broad drivers of returns such as value or momentum, can be replicated and work out of sample in data from 93 countries (Journal of Finance, 2023). So patterns exist. What the evidence doesn’t support is the step from “there is a pattern” to “this model knows where a stock is going.”
Language models such as ChatGPT, Claude and Gemini were trained to read and write, not to forecast prices, and the tests reflect that. The public trading experiments with OpenAI’s models, from benchmarks to a $100 portfolio, are covered in can ChatGPT trade for you?. How Grok did in nof1’s Alpha Arena contests is in our guide to whether Grok can trade for you. Three findings matter most for anyone weighing an AI prediction.
Alejandro Lopez-Lira and Yuehua Tang had GPT-4 score news headlines dated after its knowledge cutoff (Journal of Financial Economics, 2026). GPT-4’s scores agreed with the direction of the market’s immediate reaction on about 90% of portfolio-days, but that reaction happens before anyone can trade on it. The scores also forecast some of the slower drift afterwards, especially in small stocks and after negative news, and the returns of a strategy built on them declined as more investors adopted language models. The authors read that as the market becoming more efficient in real time.
A model trained on years of text may already know what happened after a headline. Paul Glasserman and Caden Lin tested this with GPT sentiment on financial news (2023). Inside the training window, headlines with the company’s name removed did better than the originals, a sign that what the model knew about famous companies got in the way of reading the news itself. Outside the window, the look-ahead problem goes away, though the distraction can remain. Lopez-Lira, Tang and Mingyin Zhu went further and showed that models can recall exact market figures from before their cutoff; our stock trading guide covers that study. The rule that follows: only results on dates after the model’s training cutoff count.
In 2024, a working paper by Alex Kim, Maximilian Muhn and Valeri Nikolaev reported that GPT-4, given anonymous financial statements, beat financial analysts at predicting the direction of firms’ future earnings. In February 2025 the authors temporarily withdrew it from circulation for review after a co-author “identified inconsistencies in the data and analyses while attempting to replicate past analyses.” A finding is not settled until someone else can reproduce it.
On questions about future world events, language models have closed in on the best human forecasters. A system built by Danny Halawi and coauthors, tested only on questions published after its models’ cutoffs, “nears the crowd aggregate of competitive forecasters” (2024). An ensemble of 12 models was statistically indistinguishable from a crowd of 925 human forecasters on 31 questions (Science Advances, 2024). And on ForecastBench, a benchmark run by the Forecasting Research Institute, “several models are already statistically indistinguishable from superforecaster-level accuracy,” the institute wrote on July 16, 2026, adding that the results are “more consistent with superforecaster parity than with outperformance” and that its superforecaster predictions date from 2024.
A stock price is a harder target than a question about a future event, because the price is itself a forecast: it already combines the views of everyone trading, with money behind them, and it moves the moment news arrives. To profit, a model has to be right where that price is wrong, again and again, and the evidence above says that is where most edges fade.
Strip away the forecasts and AI is very good at the work around them.
The tools that do these jobs for stock traders, from screeners to chart readers, are compared in our guide to AI stock trading.
AI predictions fail in four recurring ways, each with a well-documented case.
A model learns from the past, and the market’s worst days often have no past. Scott Baker, Nicholas Bloom, Steven Davis and their coauthors found that “no previous infectious disease outbreak, including the Spanish Flu, has affected the stock market as forcefully as the COVID-19 pandemic,” looking at large daily market moves back to 1900 (Review of Asset Pricing Studies, 2020). A model trained on that history had never seen anything like early 2020.
During the week of August 6, 2007, “a number of quantitative long/short equity hedge funds experienced unprecedented losses,” Amir Khandani and Andrew Lo wrote (NBER, 2008). Their explanation: similar funds unwound similar positions at the same time, and market makers pulled back, which made the price moves worse. When many models learn from the same data, they crowd into the same trades and run for the same exit.
In 2011, a study reported that the mood of Twitter posts predicted the daily up and down moves of the Dow Jones Industrial Average with 87.6% accuracy. When Michael Lachanski and Steven Pav re-examined it, they found the effect in the original sample but not in a longer one, “a result consistent with data snooping,” and “no evidence that our measures of Twitter mood aid in predicting the stock market out of sample” (Econ Journal Watch, 2017). Data snooping means trying so many variations on the same data that one looks brilliant by chance. Campbell Harvey, Yan Liu and Heqing Zhu argue that because researchers have tried hundreds of factors, a new one should clear “a much higher hurdle, with a t-ratio greater than 3.0,” instead of the usual 2 (NBER, 2014; Review of Financial Studies, 2016).
A signal that changes its mind every day needs many trades, and each one pays a spread. Avramov and coauthors traced part of the problem to “high turnover and extreme positions” in machine-learning portfolios. A prediction can be right more often than not and still lose money once it is traded.
Backtests are easy to publish. Live records, made before anyone knew the results, are rare. The longest public one we found is a fund: the Amplify AI Powered Equity ETF (AIEQ), which has traded since October 17, 2017. The stocks in its index are picked by a model from EquBot that uses IBM Watson to “analyze data points across news, social media, industry and analyst reports, and financial statements on thousands of U.S. companies.” Our guide to AI investing sets funds like this next to robo-advisors and AI stock scores.
| Return a year, at net asset value | AIEQ | SPY (S&P 500 fund) |
|---|---|---|
| 1 year | 6.58% | 11.98% |
| 3 years | 19.68% | 23.24% |
| 5 years | 3.13% | 11.93% |
| Since AIEQ began in October 2017 | 9.37% | Not published for these dates |
Sources: Amplify’s AIEQ page and State Street’s SPY page, month-end figures as of October 31, 2026. Over three years AIEQ returned almost 20% a year, a strong number on its own, and still trailed a plain S&P 500 fund over every period both funds report. One fund can’t settle the question; a different model might do better. But it is a real-money answer to a question most AI stock pickers only answer with backtests. For the public tests of chatbots picking stocks, see what the tests of ChatGPT found.
Keep survivorship in mind when you look for more examples. Funds, newsletters and apps whose AI did badly tend to close or go quiet, so the ones still advertising look better than the average attempt.
“Right 60% of the time” sounds impressive until you compare it with rules that know nothing. We took the daily closes of four perpetual futures, contracts that track a price, on trade.xyz, an exchange built on Hyperliquid: xyz:SP500 (the S&P 500), xyz:XYZ100 (trade.xyz’s own XYZ U.S. 100 Index), xyz:NVDA (Nvidia), xyz:TSLA (Tesla). XYZ100 is not the Nasdaq-100; our QQQ article explains the difference. Each candle runs midnight to midnight UTC, weekends included, because these markets trade every day. Then we scored three rules that need no AI at all: always say “up”, say whatever happened yesterday, and say the opposite of yesterday.
| Rule | xyz:SP500 | xyz:XYZ100 | xyz:NVDA | xyz:TSLA |
|---|---|---|---|---|
| Always say “up” | 53.1% | 53.1% | 50.0% | 50.0% |
| Say what happened yesterday | 42.1% | 50.0% | 49.0% | 49.9% |
| Say the opposite of yesterday | 57.9% | 50.0% | 51.0% | 50.1% |
| Days in the test | 241, from March 19, 2026 | 397, from October 14, 2025 | 366, from November 13, 2025 | 366, from November 14, 2025 |
Daily candles from Hyperliquid’s public API, taken November 15, 2026; flat days left out. Two lessons follow.
On both index perps, prices rose on more days than they fell, so on xyz:SP500 the rule “always up” was right 53.1% of the time over 241 days, with no model at all, and 53.1% on xyz:XYZ100. An AI that is “right 55% of the time” on the same days has beaten a rule anyone could follow by about two points, which a lucky streak could explain.
The best of the 12 results was 57.9%: on xyz:SP500, betting against yesterday’s direction was right 139 times out of 240. If the rule had no edge at all, a result at least that far from 50% would turn up about once in 60 tries, and it held in both halves of the period, at 56.3% and 60.0%. That is how a pattern earns a second look. It is not how a pattern earns trust:
The only honest next step is the one this article keeps coming back to: months the rule has never seen. Our guide to quantitative trading puts a published signal through that test on Hyperliquid data. Hit rate is not profit either. A rule that is right on many small days and wrong on a few big ones can lose money, so judge it on its returns, its worst drawdown and what it costs to trade, the checks in our guide to backtesting a trading strategy.
Whether the claim comes from an app, a newsletter, a fund or a chatbot, put it through the same eight checks.
Up or down, a return, or a ranking? Over a day, a month or a year? “Accurate” means nothing until you know what was predicted, and over what horizon.
A backtest is hypothetical performance. The SEC’s Marketing Rule lets registered investment advisers show it only with policies that keep it relevant to the people who see it, and with the information behind it. A live record, made in public before the results were known, is worth far more.
A language model has read about everything before its training cutoff, so a test inside that window can reward memory instead of foresight. Only dates after the cutoff count.
Always “up”, and simply holding an index fund, on the same days. If the AI doesn’t beat both, it hasn’t shown an edge.
How many models, markets and settings were tested before this one was shown to you? The more tries, the higher the bar should be.
A signal that trades often can lose its edge to trading costs. Ask for results with realistic trading costs included.
The worst drawdown, the longest losing streak and what happened in a shock such as the COVID-19 crash. A record without a bad month is a short record.
A registered adviser has to stand behind its claims. An anonymous channel selling “AI picks” doesn’t, and promised returns are the classic sign of a scam.
Here is a claim about an AI stock prediction tool: [paste the claim, the results and any fine print]. Grade it against these questions and quote the part of the text that answers each one: What exactly is predicted, and over what horizon? Are the results live or a backtest? Do the test dates fall after the model’s training cutoff? What baseline is it compared with: always “up”, or an index fund over the same dates? How many versions were tried before this one? Are trading costs included? What was the worst drawdown? Who is accountable for the claim? Say plainly which questions the claim leaves unanswered, and don’t tell me whether to buy anything.
When “AI picks” come with promised returns, a deposit on someone’s platform or pressure to act now, the question is no longer accuracy. Our guide to whether AI trading is legit covers the scams that use AI as bait.
We build Ethy, so weigh this part accordingly. Ethy is an AI trading agent you ask before you trade, and it is built to check a prediction rather than sell you one.
Ethy trades Hyperliquid perps, crypto as well as the stock, index, commodity and FX markets listed on trade.xyz, such as xyz:SP500 above. Leverage on perps magnifies losses as much as gains. Ethy’s backtests are estimates on past data, and Ethy is software for research and analysis that doesn’t give investment advice.
Not reliably. In research, GPT models scored the tone of news headlines in line with the market’s first reaction, but that reaction is too fast to trade, and the slower edge shrank as more investors used such models. Tests inside a model’s training period can reward memory instead of foresight. Treat a ChatGPT forecast as an idea to test, not a prediction to act on.
We know of no independent ranking of them, and accuracy alone is a weak test. On the S&P 500 perp from March to November 2026, a rule that always said “up” was right 53.1% of the time, so an AI that claims 55% has shown very little. Ask for a live record on dates after the model’s training cutoff, compared with an index fund, with every pick included.
Not in a way you can rely on. The sharpest shocks tend to have no precedent: researchers found that no earlier infectious disease outbreak, the Spanish Flu included, hit the US stock market as hard as COVID-19. A model trained on the past had nothing like it to learn from. Prepare for crashes with diversification, position size and stops rather than forecasts.
Only after you test them. The longest public record we found, the AIEQ fund, picked by AI since 2017, trailed an S&P 500 fund over one, three and five years to October 31, 2026. If you use an AI score or pick, treat it as one input, check it against an index fund over the same dates, and size positions as if it could be wrong.
Because a stock price is already a forecast. On ForecastBench, AI models have likely matched expert superforecasters on questions about future events, according to its July 2026 results. But a market price already combines the views of everyone trading, with money behind them, and moves as soon as news arrives. To profit, a model must be right where that price is wrong.
Ethy is built to test predictions, not to sell them. It reads charts on four timeframes, finds out why a market moved, backtests your rule on up to 365 days of history against buy and hold, and runs it in paper mode on live prices before any money moves. Its backtests are estimates on past data, and it is software for research and analysis that doesn’t give investment advice.
Trading involves risk, including the loss of your money. Ethy AI is a research and trading tool and does not give investment advice or recommendations. Backtests are simulations on historical data and do not predict future results.
Deploy your AI trading agent in one click.
