Whether artificial intelligence will make financial markets more efficient is no longer only an academic question. Asset managers are integrating large language models into research workflows, while securities regulators and central banks are asking how AI may change trading, supervision, and market stability. The stakes are real. If AI helps prices reflect public information faster, capital allocation improves. If it mainly lets the fastest and best-resourced institutions extract more value from information that everyone can technically see, the welfare case is less clear.

This debate often rests on a simple assumption: the main obstacle to market efficiency is whether investors can read and process information quickly enough. Our recent work suggests that this assumption misses an important part of the problem. The issue is not only whether investors have access to news. It is whether they can separate the genuinely surprising part of a news article from the part that was already predictable from what was known about the firm.

The starting point is well established. Tetlock (2007) showed that the tone of the Wall Street Journal‘s daily market column predicts next-day index returns. Loughran and McDonald (2011) gave researchers a financial-domain dictionary that became a workhorse for measuring sentiment in corporate disclosures. More recently, Chen et al. (2026) showed that representing each news article as a numerical embedding, a vector produced by a language model that summarises the meaning of the text, substantially improves cross-sectional return predictability. They concluded, however, that the resulting inefficiency is statistically real but economically modest.

In Didisheim et al. (2026), we argue that this conclusion understates the inefficiency by roughly an order of magnitude, because much of what reads like news in a news article is not truly new.

Consider a quarterly earnings article about a large software firm. A reader who knew only the firm’s identity, sector, size, valuation, profitability, and recent stock performance could already anticipate a large share of the language: cloud revenue, margins, AI investment, guidance, and the vocabulary of growth. The true news is the part that could not have been written before the article appeared: the unexpected guidance cut, the precise earnings surprise, the segment whose growth slowed. The same logic applies to Boeing aircraft-order stories, biotech clinical-trial reports, cybersecurity disclosures, and regulatory announcements. Each article has a predictable layer that follows from the firm’s profile, and a residual layer that contains the genuine surprise.

We make this distinction operational using around 6.7 million Reuters articles on US-listed firms between 1996 and 2022. We first convert each article into a language-model representation. We then ask how much of that representation could have been predicted from standard firm characteristics, such as size, valuation, profitability, past returns, and other variables in the Jensen et al. (2023) anomaly catalogue. The fitted component is what we call predictable news. What remains is pure news.

The split is economically revealing. Roughly 8-10% of the variation in news language is predictable from firm characteristics, and this predictable share rises over time as richer firm information becomes available. More importantly, the predictable component carries almost no return predictability. Investors appear to understand that some language in news articles is effectively boilerplate. They do not reward or punish firms merely because an article says things that could already have been inferred from the firm’s characteristics.

The surprising result is what happens to the residual. Pure news is an exceptionally strong predictor of future stock returns. A long-short portfolio formed on this residual delivers an out-of-sample Sharpe ratio of about 3.1 over our sample. To put that number in context, the S&P 500 has historically delivered a Sharpe ratio around 0.4, and the best individual anomaly among the 132 characteristics in the Jensen, Kelly and Pedersen catalogue delivers a Sharpe ratio of roughly 1.4. Even the ex-post optimal combination of all 132 anomalies falls short of the return predictability contained in the news residual.

This is not a small refinement to the anomaly literature. It suggests that public news contains a large amount of economically relevant information that prices incorporate only gradually. The effect also persists. The strategy retains statistically significant returns even when implementation is delayed by more than a year, well beyond the horizon at which classic post-earnings-announcement drift tends to fade.

What kinds of stories drive the effect? We use an AI-interpretability tool to break the high-dimensional news signal into roughly 5,000 narrow themes, such as cybersecurity disclosures, earnings guidance, Food and Drug Administration (FDA) approvals, meme-stock activity, cryptocurrency regulation, and bailout headlines. Two patterns emerge. Markets underreact to topics dominated by negative tone and dense numerical content, exactly the types of information that behavioural finance has long suggested investors may process slowly (Hong et al. 2000, Tetlock et al. 2008). Examples include negative earnings guidance, clinical-trial readouts, product recalls, and aircraft-order announcements. Markets overreact to topics characterised by ambiguity or very high media attention, consistent with research on noisy signals and limited attention (Fang and Peress 2009, Augenblick et al. 2025). Examples include bailout headlines, intraday price-swing coverage, and corporate COVID-19 response stories. In our decomposition, roughly 60% of the inefficiency comes from underreaction and the remainder from overreaction.

A natural concern is that a pre-trained language model may have indirectly seen information from the future, creating lookahead bias. Following He et al. (2025), we replicate the analysis using chronologically consistent language models trained only on information available as of each forecast date. The Sharpe ratio falls, because these point-in-time models are smaller and weaker, but the main conclusion is unchanged. A version of the same architecture with future information performs no better than the point-in-time version. The result is therefore not driven by model memorisation of the test period.

The broader message is that the stock market’s response to public news is much further from the textbook ideal than four decades of news-and-prices research had suggested. The technology that reveals this gap is itself part of the AI revolution. The same technology, if used widely enough, may eventually narrow the gap. But the route by which this happens matters for policy.

Financial regulation has traditionally focused on whether material information is publicly available. If two investors can see the same article at the same time, the playing field is often treated as broadly equal. AI changes that benchmark. Two investors may read the same Reuters story simultaneously, while only one has the model capacity, compute budget, and data infrastructure to extract the economically relevant signal. In that world, the relevant asymmetry is no longer simply between investors who see the information and investors who do not. It is between investors whose AI can process public information deeply and investors whose AI cannot.

Our results are strongly related to model capacity. Smaller Bidirectional Encoder Representations from Transformers (BERT)-style models capture only part of the news anomaly. Larger open-weight models capture much more of it. The largest frontier model we test extracts still more. Public information remains public, but the ability to exploit it scales with AI capability. As regulators in the US, EU, and UK consider AI use in asset management and financial markets (Korinek 2023, Foucault et al. 2025), one question should be added to the agenda: is equal access to information still the right benchmark for fair and efficient markets, or does equal access to processing capacity now matter too?

References

Augenblick, N, E Lazarus and M Thaler (2025), “Overinference from weak signals and underinference from strong signals”, Quarterly Journal of Economics 140(1): 335-401.

Bybee, L, B Kelly, A Manela and D Xiu (2024), “The structure of economic news”, Review of Economic Studies 91(2): 689-737.

Chen, A, B Kelly and D Xiu (2026), “Expected returns and large language models”, Review of Financial Studies 38: 3542-3579.

Didisheim, A, B Kelly, M Pourmohammadi and H Tian (2026), “The inefficient pricing of news”, NBER working paper.

Fang, L and J Peress (2009), “Media coverage and the cross-section of stock returns”, Journal of Finance 64(5): 2023-2052.

Foucault, T, L Gambacorta, W Jiang and X Vives (2025), “Artificial intelligence in finance”, VoxEU.org, 5 June.

He, Z, Y Lv, A Manela and J Wu (2025), “Chronologically consistent large language models”, working paper.

Hong, H, T Lim and J C Stein (2000), “Bad news travels slowly: Size, analyst coverage, and the profitability of momentum strategies”, Journal of Finance 55(1): 265-295.

Hong, H and J C Stein (1999), “A unified theory of underreaction, momentum trading, and overreaction in asset markets”, Journal of Finance 54(6): 2143-2184.

Jensen, T I, B Kelly and L H Pedersen (2023), “Is there a replication crisis in finance?”, Journal of Finance.

Kaniel, R and R Parham (2016), “Media attention and investment decisions”, VoxEU.org, 6 March.

Korinek, A (2023), “Large language models for economic research: Four key questions”, VoxEU.org, 27 July.

Loughran, T and B McDonald (2011), “When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks”, Journal of Finance 66(1): 35-65.

Meursault, V, P J Liang, B R Routledge and M M Scanlon (2023), “PEAD.txt: Post-earnings-announcement drift using text”, Journal of Financial and Quantitative Analysis 58(6): 2299-2331.

Tetlock, P C (2007), “Giving content to investor sentiment: The role of media in the stock market”, Journal of Finance 62(3): 1139-1168.

Tetlock, P C, M Saar-Tsechansky and S Macskassy (2008), “More than words: Quantifying language to measure firms’ fundamental information”, Journal of Finance 63(3): 1437-1467.