The 'biggest market anomaly ever found' — and why you still can't beat the market
- Robin Powell
- Jun 19
- 9 min read
Researchers claim to have identified the biggest market anomaly ever documented. It's more than double the size of anything previously known, and it goes on moving share prices for up to 18 months after the news breaks. Yet an ordinary investor can't profit from it, and that may be the most important finding of all.
A new study has identified what its authors describe as the largest price inefficiency ever documented in the stock market. Using a large language model to read 6.7 million Reuters articles, four researchers built a trading signal with a Sharpe ratio of 3.1 — more than double the strongest factor in the standard academic catalogue — that goes on predicting returns up to 18 months into the future. The obvious reading is that markets can be beaten after all, and that artificial intelligence is the tool that does it.
The reading is wrong, or at least badly incomplete. The people who found the anomaly did so with a seven-billion-parameter language model, millions of news articles and a statistical pipeline that strips each story down to the part that is genuinely surprising. None of that is available to an ordinary investor. And the most telling detail is that the researchers published the finding rather than quietly trading on it.
The paper, The inefficient pricing of news, is by Antoine Didisheim and Hanqing Tian of the University of Melbourne, Mohammad Pourmohammadi of Yale and the Swiss Finance Institute, and Bryan Kelly of Yale and AQR Capital Management. It was released in April 2026 as National Bureau of Economic Research working paper 35093, and has not yet been peer-reviewed — a point worth holding on to. But its central claim, that the biggest market anomaly ever found is also one an ordinary investor cannot use, rests on evidence the authors themselves set out in unusual detail.
Most financial news isn't news
Most of what looks like news in a financial story is not, in fact, news. A reader who already knows a company's sector, size, valuation and recent performance can predict a good deal of what any given article about it will say. The researchers put a number on this: around 10 per cent of the variation in article content can be explained by standard stock characteristics alone. That predictable layer, it turns out, moves prices barely at all.
To separate the predictable layer from the rest, the team used a large language model — an open-weight system called E5-Mistral-7B, with seven billion parameters — to turn each article into a string of 4,096 numbers that captures its meaning. They then stripped out the part of that mathematical fingerprint that could be predicted from 132 company characteristics drawn from a widely used catalogue of stock-market factors. What remained, the genuinely surprising residual, is what they call 'pure news'.
Pure news is the signal. The market reacts to it, but slowly and incompletely, absorbing it over months rather than days. The difference shows up directly in the numbers. A strategy built on the raw article fingerprints earns a Sharpe ratio of 1.1; one built on the residualised pure-news component earns 3.1. Strip out what was already knowable, and the signal nearly triples.
The biggest market anomaly in the catalogue
A Sharpe ratio measures return per unit of risk. A figure of 1.0 is considered very good, and most individual strategies never get close. The best single factor in the Jensen, Kelly and Pedersen catalogue — a database of 132 documented anomalies that serves as the field's benchmark — manages 1.4 over the period studied. The pure-news signal, across the full universe of stocks, scores 3.1: the biggest market anomaly ever measured, and nearly double the size of the next-largest in the catalogue.
That headline figure overstates what anyone could actually trade. Restrict the sample to companies large enough to deal in without moving the price — those above the median size on the New York Stock Exchange — and the pure-news Sharpe falls to 1.4, against 0.9 for the best conventional factor in the same large-stock universe. The edge is still there. It is simply much smaller once confined to stocks an investor could realistically buy and sell.
It shrinks again when you change the tool. The headline result relies on a frontier language model. When the researchers repeated the exercise with a smaller academic model — a version of GPT-2 trained only on text available at the time, ruling out any suspicion that it had seen the future — the Sharpe fell to 1.63, and to 1.61 even when the same model was allowed to peek at later data. The near-identical figures all but rule out hindsight as the cause. The fall was down to raw capacity. Mistral has seven billion parameters and was trained on seven trillion words; the academic model has 1.5 billion parameters and 71 billion words. The authors call this gap the 'compute divide'.
In its strongest form the strategy earns roughly 30 per cent a year at 10 per cent volatility, almost all of it unexplained by standard measures of risk. But the version available to someone without a frontier model and an institutional trading desk is a fraction of that.
An edge built for institutions
The obvious objection is that the authors have already dealt with all this. They have. The paper engages trading costs directly and concludes that the anomaly survives them. This is the strongest version of the case, and it deserves to be met head-on.
The pure-news strategy trades heavily. Its one-sided monthly turnover is 75 per cent — higher than any of the 132 factors in the catalogue, the next-highest being short-term reversal at 67 per cent. High turnover means high trading costs, and trading costs are where most paper anomalies die. So the authors calculate the strategy's returns net of costs, assuming a charge of ten basis points — a tenth of 1 per cent — for every dollar traded. Even then, its net performance beats every conventional anomaly. Averaging the signal over a six-month window cuts the turnover sharply while holding the Sharpe near 3.0; a 24-month window still delivers 2.4. Their conclusion is plain: the anomaly 'is not an artifact of unrealistic trading costs'.
The question is whose costs. The ten-basis-point assumption comes from a 2018 study by Andrea Frazzini, Ronen Israel and Tobias Moskowitz, who measured the real cost of trading using $1.7 trillion of live execution data from a large institutional money manager across 21 developed markets between 1998 and 2011. They found actual costs an order of magnitude smaller than earlier academic estimates — a median of around six basis points per rebalance on the New York Stock Exchange — and concluded that the main anomalies are 'robust, implementable, and sizeable'.
The institutional money manager in that study is AQR Capital Management. Bryan Kelly, a co-author of the news paper, is AQR's head of machine learning. There is no impropriety in this: the cost figure is real, and the firm that produced it is named in the open literature. But it sets the terms precisely. Ten basis points per dollar traded is a normal trading cost — for AQR. It is not a normal trading cost for an ordinary investor turning over three-quarters of a portfolio every month. Robert Novy-Marx and Mihail Velikov found in 2016 that strategies with one-sided monthly turnover above 50 per cent rarely survive real-world costs outside elite execution environments. The pure-news strategy sits at 75 per cent. The moat around it is not secrecy. It is infrastructure.
The decay has already begun
Even setting infrastructure aside, there is a clock running on the biggest market anomaly ever found. In 2016 R. David McLean and Jeffrey Pontiff examined 97 variables that academic studies had shown to predict stock returns, and tracked what happened to each once it was published. Returns were 26 per cent lower out of sample and 58 per cent lower after publication. The gap between those figures — 32 per cent — is the portion they attribute to investors reading the research and trading on it. Publication is itself an event that erodes the thing being published.
Two features of their finding bear directly on the news paper. First, the decay was greatest for the predictors with the highest in-sample returns — and the pure-news anomaly has the highest in-sample return ever documented. Second, formal publication understates how fast learning happens, because sophisticated firms monitor working-paper repositories as a matter of routine. The news paper was posted in April 2026. The window the authors describe is already narrowing.
Bad news travels slowly
Why should prices absorb genuine news so slowly? The paper's own decomposition supplies part of the answer: 62.1 per cent of the anomaly comes from underreaction, the rest from overreaction. Investors are slow to price in news that is negative in tone — cybersecurity breaches, criminal charges, product recalls — and news that is dense with numbers, such as earnings guidance and year-on-year comparisons. They overreact to news that is loud but ambiguous, such as bailout headlines and coverage of dramatic intraday price swings.
This matches what behavioural finance has been documenting for a quarter of a century. Harrison Hong and Jeremy Stein argued in 1999 that information diffuses gradually through the market: one group of investors is slow to react, causing underreaction, after which momentum traders pile in and push prices too far. The following year, Hong, Terence Lim and Stein found that momentum is strongest in stocks with little analyst coverage, and that such stocks 'react more sluggishly to bad news than to good news'. In 2007 Paul Tetlock showed that high media pessimism predicts downward pressure on prices the next day, followed by a reversion to fundamentals. The biases the news anomaly exploits are not new. They are deep, well-evidenced and human.
Equal access, unequal processing
Step back, and the anomaly looks less like a trading opportunity than like a map of a new kind of inequality. For most of its history, financial regulation has worked to guarantee equal access to information: everyone sees the same filings, the same announcements, the same Reuters story at the same moment. The news paper suggests equal access is no longer the binding constraint. Two investors can read the same article simultaneously; only one has the model, the compute budget and the data infrastructure to extract the signal buried inside it.
The authors put this directly. 'Public information remains public,' they write in an accompanying column, 'but the ability to exploit it scales with AI capability.' They go on to ask whether equal access to information is still the right benchmark for fair markets, 'or does equal access to processing capacity now matter too?' It is a question the paper poses rather than answers.
It is worth being plain about who is asking. Kelly is head of machine learning at AQR, and discloses consulting income from the firm exceeding $10,000 over the past three years. AQR supplied the trading-cost data that underpins the paper's claim that the anomaly is implementable. The researchers best equipped to harvest this signal are the ones publishing its existence, disclosing their interest and raising the regulatory question it poses. That is not a reason to discount the finding. If anything it is the opposite: this is positive economics, not a sales pitch.
Own the market, ignore the headlines
So the biggest market anomaly ever found turns out to be a poor prospect for the people most likely to hear about it. It is real, but only at full size in a universe of stocks too small to trade in quantity. It takes a frontier model and serious data infrastructure even to detect. Its turnover is high enough to surrender the edge to anyone without institutional execution. And it has been public since April, and the anomalies that decay fastest tend to be the ones, like this, that start out largest.
The sensible response is the one the evidence has pointed to for decades: hold a low-cost, globally diversified index fund and let the news look after itself. The people with the biggest edge are the ones telling you it exists — and telling you, in the same breath, how far out of reach it lies.
Resources
Didisheim, A., Kelly, B. T., Pourmohammadi, M., & Tian, H. (2026). The inefficient pricing of news. NBER Working Paper No. 35093.
Frazzini, A., Israel, R., & Moskowitz, T. J. (2018). Trading costs. Available at SSRN 3229719.
Hong, H., & Stein, J. C. (1999). A unified theory of underreaction, momentum trading, and overreaction in asset markets. Journal of Finance, 54(6), 2143–2184.
Hong, H., Lim, T., & Stein, J. C. (2000). Bad news travels slowly: Size, analyst coverage, and the profitability of momentum strategies. Journal of Finance, 55(1), 265–295.
Jensen, T. I., Kelly, B., & Pedersen, L. H. (2023). Is there a replication crisis in finance? Journal of Finance.
McLean, R. D., & Pontiff, J. (2016). Does academic research destroy stock return predictability? Journal of Finance, 71(1), 5–32.
Novy-Marx, R., & Velikov, M. (2016). A taxonomy of anomalies and their trading costs. Review of Financial Studies, 29(1), 104–147.
Tetlock, P. C. (2007). Giving content to investor sentiment: The role of media in the stock market. Journal of Finance, 62(3), 1139–1168.
Letting the news look after itself
If the lesson of this article — that the sensible response to a market this complex is to own it cheaply rather than try to outsmart it — is one you want to act on, How to Fund the Life You Want by TEBI editor Robin Powell and Jonathan Hollow makes the evidence-based case in full. Bloomsbury published the second edition, written for UK investors who want to build and spend their wealth without paying for an edge they cannot reach. Buy it on Amazon.
For readers who would rather have a professional alongside them, TEBI's Find an Adviser directory lists advisers who have publicly committed to the low-cost, evidence-based approach this article describes.