Source-linked AI summary
Buy the Rumor, Sell the News: When Is News Priced In?
Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini, Sid Ghatak, Arman Khaledian
TL;DR
The paper asks when different kinds of financial news are priced in and measures event-time returns across a large, placebo-controlled news corpus. It finds that price moves concentrate before and at publication, while quantified news drifts and story-driven news reverses afterward.
Problem
The study asks when, for which kinds of news, and by how much markets absorb public information, amid limited tagged-event, story-structure, and publicity benchmarks.
Method
The paper classifies 4.57 million articles, clusters repeated coverage into events, and measures beta-adjusted abnormal returns against neutral-sentiment placebo events.
Results
2.8 times: pooled news-aligned cumulative returns by publication-day close exceed their value 20 days later; quantified news drifts, while story-driven news gives back its move.
Takeaways & Limitations
For news-based forecasting or trading, event type, coverage intensity, and volatility remain informative, while directional information is largely spent by the closing bell.
Takeaways & Limitations
The study is descriptive and does not establish that news causes price movements.
Abstract
from arXiv · showhide
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 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 2.8 times its value 20 days later, and for rumor-flagged events the rumor day captures the entire move while the subsequent confirmation contributes nothing. Second, measured against the placebo of comparable stocks, markets underreact to numbers and overreact to stories: quantified fundamental news (earnings, dividends, guidance, analyst actions) keeps drifting in the direction of the news for weeks, while soft story-driven news (launches, macro commentary, leadership) gives back its move. 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. The study also produces a table of measured drift for each event tag, usable as a prior in news-conditioned forecasting models.
1 Introduction
The paper develops infrastructure to classify and de-duplicate multimillion-article news flow, then measures price moves relative to publication using a placebo design. It finds that news-aligned moves occur mainly before and at publication, while background drift explains apparent post-news reversal and quantified fundamentals continue to be underreacted to.
- Main findings: News-aligned price moves sit overwhelmingly before and at publication, and rumor-flagged events have their entire move on the rumor day.These results support both sayings that news is priced in by publication and that the rumor is bought while the news is sold.
- Measurement infrastructure: 17 news tags and five attributes form a taxonomy distinguishing economic objects such as earnings, lawsuits, and promotional coverage.The taxonomy was designed iteratively to identify distinctions that matter for price behavior.
- Measurement infrastructure: 1.68 million (stock, day, tag) events receive beta-adjusted abnormal-return measurement around publication.The analysis uses a full-corpus tagging pipeline and measures returns around resulting events.
- Main findings: 364,405 neutral-sentiment events provide a built-in placebo for publicity and reveal background drift unrelated to news.The placebo removes drift that otherwise appears as widespread post-news reversal.
- Main findings: Markets underreact to quantified fundamental news after background-drift correction, while the correction reinterprets apparent post-news reversal and good-bad-news asymmetry.The introduction presents these as consequences of separating news effects from return drift.
2 Related work
Related work spans LLM sentiment signals, structured event representations, and methodological approaches that distill LLM labels into compact classifiers. This study contributes an empirically measured drift prior from 1.68 million events and an industrial-scale tagging pipeline.
- LLM sentiment prediction: LLM equity-news research uses extracted sentiment to predict returns, market and macro outcomes, trading signals, fundamental analysis, and stock ratings.This strand includes headline scoring, open-model fine-tuning, and LLM applications to fundamental analysis and ratings.
- Event-based forecasting: Structured-news research represents articles as events, using event representations, event memory, and end-to-end event-driven trading policies.These systems require knowledge of which event tags matter and how quickly their information decays.
- Event-based forecasting: The study’s measured drift table supplies an event-information prior estimated on 1.68 million events rather than learned end to end.The prior addresses event-tag relevance and information decay for structured-news forecasting systems.
- Methodology: Methodological work finds that distilling LLM-generated labels into small supervised classifiers can match human annotation quality at a fraction of the cost.Related work also reduces retrieval costs over financial text and documents look-ahead risks in LLM text-based forecasting.
- Methodology: The tagging pipeline is an industrial-scale distillation approach with per-tag validation, extending methodological work on compact supervised classifiers.The passage explicitly characterizes this pipeline as an industrial-scale instance.
3 Data
The study uses NewsWitch, a commercial financial-news product processing extensive coverage of roughly 3,000 US-listed stocks. Its sample retains 4.57 million deduplicated, financial articles directly related to covered stocks from a much larger raw corpus.
- Data: 4.57 million retained articles form the sample after deduplication and financial-content filtering.The corpus contains more than 70 million raw articles overall.
- Data: Roughly 3,000 US-listed stocks are covered by the NewsWitch news corpus.Daily processing draws from more than half a million sources and handles over a quarter million news URLs.
- Data: NewsWitch has operated in daily production since the beginning of 2024 as the commercial source of the corpus.The product processes financial-news URLs concerning the covered stocks.
4 Method
The method tags financial articles, groups repeated coverage into stock-level stories and events, and measures beta-adjusted abnormal returns against neutral-event placebos. It also tests robustness to follow-up coverage while treating the results as descriptive rather than causal.
- Method overview: The pipeline has three stages: distilled-LLM article tagging, story and event grouping, and placebo-based abnormal-return measurement.This design separates event identification from return measurement around those events.
- Article classification: 17 event tags and five binary attributes are assigned by a compact classifier trained through teacher labeling and active learning.A GPT teacher labeled 29,472 random articles; the distilroberta-base student has 82M parameters and was trained on 600,000 fresh articles.
- Story and event construction: Embedding-based clustering links same-tag articles within stock-days and continues events across days when later coverage remains close to a story’s frozen centroid.Within stock-days, clustering uses title-and-summary cosine similarity with a 0.80 threshold.
- Return measurement: Abnormal returns use a rolling 252-day OLS beta against the S&P 500, with cumulative windows spanning days −5..0, day 0, days +1..+5, and days +6..+20.Signed events count returns in the news direction as positive, and inference uses 5,000-date cluster-bootstrap draws.
- Identification and limitations: +1.41% to +1.02%: restricting to first reports reduces the earnings pre-window, showing follow-up coverage inflates measured anticipation by roughly a third.All post-publication conclusions are unchanged; the study describes event-time placement rather than claiming that news causally moves prices.
5 Results
News-aligned price moves concentrate before and at publication, with rumor confirmations adding little. After removing background drift, quantified fundamental news continues while soft, attention-driven news reverses, and publication also resolves uncertainty by reducing volatility.
- 5.1 Timing and rumors: +1.41% pre-publication and +0.42% on publication day for earnings contrast with −0.05% afterward, illustrating the priced-in signature.Analyst actions show the same shape: +1.12% pre and +0.22% on the day against −0.13% afterwards.
- 5.1 Timing and rumors: 94% of 18,618 rumor-flagged events receive same-tag confirmation within 60 trading days, but the rumor day delivers +0.36% while confirmation adds +0.01%.The interval between rumor and confirmation contributes −0.09%, and days +6 to +20 after confirmation give back −0.06%.
- 5.2 Background drift: −0.92% over the following month among small caps receiving neutral coverage reveals a background drift that also appears on quiet stock-days.The corresponding neutral-news drifts are −0.58% for mid caps and −0.34% for large caps.
- 5.2 Background drift: 0.00% adjusted drift follows both positive and negative news over days +6 to +20, eliminating apparent good-news overreaction and bad-news underreaction.Raw drift was −0.62% after positive news and +0.63% after negative news; both become 0.00% after subtracting the baseline.
- 5.3 Adjusted drift: +0.35% adjusted drift for capital returns and +0.22% for earnings contrast with −0.34% for macro read-throughs and −0.18% for product launches over days +6 to +20.Quantified fundamental tags continue, whereas soft attention-driven tags reverse; launch and partnership coverage is 96–97% positive.
- 5.4 Width and volatility: 1.36 times normal absolute movement occurs on days with ten or more articles versus 1.05 for single-article days, while scheduled disclosure events widen and then narrow volatility.Neutral guidance moves are 1.29 times normal on publication day and 1.23 times over the following week; neutral earnings move 1.11 times.
6 Implications for news-conditioned forecasting
The paper’s measurements imply that news-conditioned forecasting systems should rank news by measured drift priors and represent news as tagged event lines. This design reflects direction being mostly priced in at publication while width effects persist.
- Ranking: Forecasting systems should rank event lines by measured per-tag, per-attribute, per-size drift priors rather than source-reputation heuristics.This applies when compressing a day’s news into a short context.
- Representation: News should be encoded as tagged, flagged event lines to preserve event-level information for forecasting.The passage specifies event lines with tag and flags such as scheduled and rumor.
- Representation: Direction is mostly priced in at publication, whereas width effects persist.This distinction motivates representing news for both directional and width-related forecasting effects.
7 Limitations
The study’s first limitation is that its sentiment, summaries, and event tags depend on vendor-generated and distilled-LLM labels, which introduce classification disagreement. These disagreements are concentrated in low-consequence junk categories, but the passage does not state the completed teacher-only robustness result.
- Labels: About one article in eight receives a label from the distilled classifier that disagrees with its LLM teacher.The event tags are layered on top of the corpus vendor’s LLM-generated sentiment and summaries.
- Labels: Label disagreements concentrate in price commentary and promotional content, where misclassification matters least.These are identified as the study’s two junk categories.
- Labels: The passage notes a teacher-only check in Table 4 but does not provide its completed conclusion.Accordingly, no specific robustness result can be reported from the supplied text.
8 Conclusion
Across 1.68 million tagged events, news-price moves occur before and at publication, with confirmation adding no further movement after a rumor. The conclusion emphasizes placebo-adjusted measurement and event type, coverage intensity, and volatility as durable inputs for news-based forecasting or trading systems.
- Conclusion: 1.68 million tagged events show that anticipation builds before publication, publication absorbs what remains, and subsequent weeks add nothing new.The study concludes that both market sayings survive its full-news-flow test.
- Conclusion: Rumor confirmation adds no further price movement: by confirmation, the market has finished trading the rumor.Waiting for certainty therefore means buying after the market has already incorporated the information.
- Conclusion: A placebo of comparable stocks is necessary because beta-benchmark drift occurs in stocks with and without news, creating misleading reversal, continuation, and sentiment-fading results.The supplied conclusion says subtracting this placebo makes those apparent effects disappear together.
- Conclusion: For news forecasting or trading, the durable information is event type, coverage fact and intensity, and volatility, while directional information is spent by the closing bell.The conclusion characterizes volatility as news “width” alongside direction.