Buy the Rumor, Sell the News: When Is News Priced In?

arXiv:2608.14014 · cs.AI, cs.LG, q-fin.ST · Submitted 2026-08-14 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Buy the Rumor, Sell the News: When Is News Priced In?".

Jane: The gist The news-aligned move sits before and at publication,

Tom: First, who's behind it and why it matters.

Title and authors: Jane: So, what do they actually find? The summary suggests that the news-aligned move—the actual price change related to a piece of news—sits before and right at the time of publication. They also look at rumor-flagged events and find that for those, the rumor day captures everything, while any confirmation after that doesn't add anything new.

Tom: That’s a big finding because it confirms one side of that old saying. The cumulative move in the news direction by the close of publication day is two point eight times its value twenty days later for all signed events, according to the paper <ref:2608.14014#pg2>.

Lu: And they introduce a way to measure this: they built this three-layer infrastructure involving tagged events, story structure, and a benchmark for publicity. The tagging process uses an Azure-hosted GPT teacher that labels a sample with seventeen event tags and five attributes, and then a small distilled classifier extends those labels to the full corpus at ninety-five percent fidelity <ref:2608.14014#pg1>.

Meng: Ninety-five percent fidelity is high, but I wonder how robust that distillation is when you’re dealing with such varied financial language. What's the cost of building that infrastructure?

Lalam: The paper says they achieved this pipeline with an estimated ninety-five percent corpus-wide teacher fidelity at a total cost of under twenty dollars for teacher labels, which is pretty low for this kind of work.

Jane: And that’s just the mechanics. But the core result is more interesting than just the setup. They also find that markets underreact to numbers but overreact to stories when you remove a placebo group of comparable stocks that drift with or without news, which is a key finding in Section two of their work <ref:2608.14014#pg2>.

Tom: That’s what I was hoping for. It suggests that we need to focus on the narrative structure rather than just the raw numbers when building our models.

The paper's summary: Jane: The paper doesn't just stop at those findings; they suggest a few ways to improve how we use this information for forecasting. They say that for building news-conditioned forecasting systems, durable information is the kind of event, the fact and intensity of coverage, and its width.

Lu: They propose ordering these event lines by measured per-tag, per-attribute, per-size drift priors rather than relying on source reputation heuristics. This means we should prioritize how much a specific tag or attribute actually influences drift over just trusting which news outlet is known for good stuff.

Meng: So instead of using some high-level reputation score, you use the actual measured drift data to decide what feature to focus on? That’s a practical shift for deployment.

Tom: Right. And they suggest encoding text as tagged, flagged event lines—tag, scheduled, rumor—instead of just sentiment scalars. They argue that the tag and flags carry the durable information better than just reading how positive or negative something sounds in a sentence.

Lalam: That makes sense from a modeling standpoint because it lets you know exactly which structural features matter and how fast that information decays over time. It shifts us away from relying on noisy sentiment signals.

Jane: And they also point out that the background-drift correction can be used to reinterpret market data by subtracting the placebo group, which helps remove apparent asymmetries between good and bad news and eliminates size gradients in post-event drift.

Tom: So, if you're building a system, you use this correction to filter out noise caused by general coverage patterns that look like reversals when they aren't real. That’s a useful tool for cleaning up the signals we feed into our models.

The paper's improvements: Jane: So, wrapping up the "Buy the Rumor, Sell the News: When Is News Priced In?" paper, we see that for news-conditioned forecasting, it’s not just about direction. It’s about durability and width—how much coverage there is and what kind of event it is.

Tom: The main idea is that direction gets priced in at publication, but the presence and width of the news flow persist over time. That's what we need to capture in our event lines instead of just sentiment scores.

Lu: The authors conclude that the exploitable object in news flow isn't necessarily direction, but presence and width. The focus shifts to how news is priced, and the direction is spent by the closing bell.

Meng: From an engineering standpoint, this means we should be structuring our input data around these tagged event lines rather than just relying on raw sentiment scalars for our models. That’s a concrete way to improve what we feed into production systems.

Lalam: I'm confident in the paper's answer because they provided a reproducible recipe for labeling those large corpora, and their final tag map shows that quantified fundamental tags continue to drift while soft news reverses, which is a really solid structural finding.

Jane: It’s clear this work gives us a better lens for understanding how we can build systems that actually learn from the structure of information flow rather than just the surface-level sentiment. That’s what we have today with this discussion on "Buy the Rumor, Sell the News: When Is News Priced In?".

Tom: Yeah. We’ll be looking out for how these structural insights translate into actual predictive models soon.

Conclusion: Jane: So we've looked at "Buy the Rumor, Sell the News: When Is News Priced In?" and it seems like for forecasting, we need to focus on how much coverage there is and what kind of event it is, not just direction.

Tom: That’s right. The paper shows that direction gets priced in at publication, but the presence and width of the news flow persist over time. We have to capture that structure instead of just sentiment scores for our models.

Lu: They conclude that the exploitable object in news flow isn't necessarily direction, but presence and width. The focus shifts to how news is priced, and the direction is spent by the closing bell.

Tom: Yeah. We’ll be looking out for how these structural insights translate into actual predictive models soon.

Lu: I'm curious to see if this event-time approach applies when we look at things that are inherently spatial or temporal, like in some of the work on three dee generation or action modeling.

Meng: Yeah, because when you’re shipping something, you have to worry about what breaks at scale. This paper gives us a better way to structure data before we even start building the model architecture.

Lalam: I'm ready for whatever comes next, but this research on event-time pricing seems like it sets a really strong foundation for how we think about information flow in these complex AI systems.

Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini, Sid Ghatak, Arman Khaledian

Tailstate Intelligence Ltd · Zanista AI Ltd · Increase Alpha, LLC

cs.AI, cs.LG, q-fin.ST

Submitted: 2026-08-14

Updated: 2026-10-18

Comments: Revised version accepted at the 3rd Workshop on LLMs and Generative AI for Finance (AI4F 2026), held in conjunction with ICAIF 2026, Milan, Italy. https://ai4f.org/

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

The gist: The gist The news-aligned move sits before and at publication, and for rumor-flagged events the rumor day captures everything How it works The study tests two old market sayings—that news is

Key concepts

News-aligned move: This refers to the price movement associated with news occurring before or at its publication time. The study found that the cumulative market move in the news direction by publication day is significantly larger than moves observed later, confirming that information is largely priced in when it breaks. This suggests anticipating prices around the moment of release is crucial.
Rumor-flagged events: These are specific types of news where a rumor is bought while the main news story is still being sold. The research demonstrated that for these events, the rumor day captures the entire price move, while subsequent confirmation adds little value. This highlights the immediate impact of speculative information versus confirmed facts.
Width and Direction: This concept explains that news impacts markets not just in a certain direction (up or down) but also in terms of how widely it spreads. Publicity increases volatility before publication, but this volatility subsides once the news is released because the uncertainty is resolved. The study suggests focusing on the breadth of coverage alongside its direction for better forecasting.

Summary

The gist The news-aligned move sits before and at publication, and for rumor-flagged events the rumor day captures everything

How it works

The study tests two old market sayings—that news is already priced in by the time it is published, and that the rumor is bought while the news is sold—on 4.57 million financial news articles covering roughly 3,000 US stocks (2023–2026) The researchers built a three-layer measurement infrastructure to test these claims This infrastructure included tagged events, story structure, and a benchmark for publicity

The tagging process involves two main steps First, an Azure-hosted GPT teacher labels a sample of articles using 17 event tags and five attributes Second, a small distilled classifier refined through active learning extends these labels to the full corpus at 95 percent fidelity This pipeline is designed for industrial scale, achieving an estimated 95% corpus-wide teacher fidelity at a total cost of under twenty dollars

Events are defined through embedding-based clustering Articles are clustered on the cosine similarity of their title-and-summary embeddings, restricted to the same primary tag, and matched against running stock stories An event is an aggregate of a (stock, trading day, tag) where the impact averages its articles

Key Findings

The first major finding confirms that the news-aligned move sits before and at publication Pooled across all signed events, the cumulative move in the news direction by the close of publication day is 2.8 times its value twenty days later For rumor-flagged events, the rumor day captures the entire move while subsequent confirmation contributes nothing

Second, markets underreact to numbers and overreact to stories Quantified fundamental news keeps drifting in the direction of the news for weeks, while soft story-driven news gives back its move This is revealed by subtracting a placebo of comparable stocks that drift with or without news, showing that markets underreact to numbers and overreact to stories

Third, news carries width as well as direction Publicity raises volatility before the publication day, and volatility declines once the news is out because publication resolves uncertainty

Measurement Details

Returns are measured as beta-adjusted abnormal returns around 1.68 million (stock, day, tag) events The abnormal return is calculated using a rolling 252-day OLS beta of the stock against the S&P 500 index The study uses a single-beta benchmark to measure residual drift

The analysis distinguishes between raw and adjusted means, where the adjusted mean subtracts the baseline drift of comparable stocks before applying the sign of the news direction This adjustment removes apparent asymmetries between good and bad news and eliminates size gradients in post-event drift

Implications for Forecasting

The study suggests that for building news-conditioned forecasting systems, durable information is the kind of event, the fact and intensity of coverage, and its width Event lines should be ordered by measured per-tag, per-attribute, per-size drift priors rather than by source reputation heuristics Since direction is mostly priced in at publication while width effects persist, text should be encoded as tagged, flagged event lines rather than sentiment scalars Any claimed news alpha should be benchmarked against a coverage-presence baseline measured here

The paper concludes that the exploitable object in news flow is presence and width, not direction The focus is on how news is priced, and the direction is spent by the closing bell

Limitations

Limitations include reliance on an LLM-generated corpus for sentiment and summaries The study does not explicitly model style effects such as size through a multi-factor treatment The corpus consists of public news, meaning some information is already priced by the time it appears The economic significance exercise is deliberately simple, focusing on how news is priced rather than designing a trading system

The study provides a reproducible recipe for labeling multimillion article corpora by distilling a large language model teacher into a compact classifier at 95 percent agreement with its Azure GPT teacher and negligible cost It also produces a per-tag, per-attribute, per-size drift table usable as a prior in news-conditioned forecasting models The study demonstrates that the background drift correction reinterprets post-news reversal and eliminates apparent asymmetries between good and bad news The final tag map shows that quantified fundamental tags continue to drift, while soft news reverses, indicating that hard, quantified disclosures keep drifting in their own direction for weeks after publication; soft narrative coverage gives back its move. The study is a descriptive account of where the news-aligned move sits in event time, not a causal claim about news moving prices The study is a descriptive account of where the news-aligned move sits in event time, not a causal claim about news moving prices The study is a descriptive account of where the news-aligned move sits in event time, not a causal claim about news moving prices

--- Page 1 ---

Buy the Rumor, Sell the News: When Is News Priced In? alireza kargarzadeh Tailstate intelligence ltd London United Kingdom alireza.kargarzadeh@tailstate.ai Nariman khaledian Independent Researcher Antibes France khaledian.nariman@gmail.com Navid parvini Zanista AI ltd London United Kingdom navid.parvini@zanista.ai Sid ghatak Increase Alpha Miami United States s.ghatak@increasealpha.com Arman khaledian Zanista AI ltd London United Kingdom arman.khaledian@zanista.ai Abstract Two old market sayings hold that news is already priced in by the time it is published, and that the rumor is bought while the news is sold Both are empirical claims about event time: they place the price move associated with a piece of news before and at publication rather than after it Whether the claims hold, for which kinds of news, and by how much are basic questions about how fast markets absorb public information We test them on 4.57 million financial news articles covering roughly 3,000 US stocks (2023–2026) A large language model teacher distilled into a compact classifier through active learning assigns each article one of 17 event tags and five attributes articles are clustered into stories to separate first reports from follow-up coverage and beta-adjusted abnormal returns are measured around the resulting 1.68 million stock-day events with 364,405 neutral-sentiment events as a placebo group Three results follow First, the price move associated with news concentrates before and at publication: pooled across all signed events, the cumulative move in the news direction by the close of publication day is 2.

Improvements for AI systems

What news is priced in at any given time can be determined by measuring when news is priced in using a two-step pipeline tags an Azure-hosted GPT teacher labels a sample, and a small distilled classifier, refined with active learning, extends the labels to all 4.57 million articles. This system allows AI models to distinguish between genuine information and market anticipation by classifying articles into 17 event tags and five attributes.
The system can identify which types of news drive price moves by measuring measured drift for each event tag, providing a per-tag, per-attribute, per-size drift table usable as a prior in news-conditioned forecasting models. This enables AI to rank the importance of different news categories for predictive tasks.
AI systems can be improved by using the measured information to structure text as tagged, flagged event lines (tag, scheduled, rumor, primary source, NEW or follow-up age) rather than as sentiment scalars, because the tag and flags carry the durable information. This allows models to prioritize structural features over noisy sentiment signals.
Forecasting systems can utilize the background drift correction to reinterpret market data by subtracting the placebo: Net of it, markets underreact to quantified fundamental news and overreact to soft stories, allowing AI to filter out spurious reversals caused by general coverage patterns.
AI can be trained using event-based representations instead of sentiment scalars, as this approach structures news into events rather than sentiment scalars and systems need to know which event tags matter and how fast their information decays. This leads to more accurate forecasting by focusing on the structure of the information flow.