PaperlayerView exact revision on arXiv

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

arXiv:2608.14014v1cs.AIcs.LGq-fin.ST

TL;DR

The paper asks when stock prices incorporate news and how timing varies by event type. It labels and clusters 4.57 million financial news articles, then measures beta-adjusted event returns against a comparable-stock placebo. Price moves concentrate before and at publication, but quantified disclosures subsequently drift while narrative coverage reverses, and volatility declines after publication.

  • Problem

    Evidence remains limited on when news is priced into stocks and how timing varies by event type, questions relevant to forecasting and trading systems.

    Source: §1 · p. 1

  • Method

    Using 4.57 million articles, the study labels news, clusters stories, and measures event-time beta-adjusted abnormal returns against a comparable-stock placebo.

    Source: §1 · p. 1, §4.2 · p. 3, §8 · p. 8

  • Results

    News-aligned moves concentrate before and at publication; placebo-adjusted results show subsequent drift after quantified disclosures, reversal after narrative coverage, and lower volatility after publication.

    Source: §8 · p. 8

  • Takeaways & Limitations

    Event-specific drift estimates can serve as ranking and importance-sampling priors for news-conditioned forecasting models.

    Source: §1 · p. 2

    The single-market-beta return benchmark does not explicitly model style effects such as stock size.

    Source: §7 · p. 8

Abstract

from arXiv · show

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

News-aligned price moves concentrate before and at publication, while placebo-adjusted results distinguish underreaction to quantified fundamentals from overreaction to soft stories. The study enables this measurement through large-scale event labeling, story clustering, and a neutral-news placebo.

1 Introduction

  • Full-news-flow measurement requires three previously missing layers: accurate event labels, consolidated stories, and a benchmark separating publicity from news direction.Existing LLM financial-news research documents predictability but does not build these layers. The study asks why favorable coverage can accompany falling prices and what news timing implies for forecasting and trading.Source: §1 · p. 1
  • 4.57 million articles covering roughly 3,000 US stocks (2023–2026) receive 17 news tags and five attributes through teacher–student labeling.An Azure-hosted GPT teacher supplies sample labels; an active-learning-refined compact classifier extends them at 95 percent fidelity to the teacher and negligible cost. Story clustering separates first reports from follow-up coverage.Source: §1 · p. 1
  • 1.68 million (stock, day, tag) events support beta-adjusted abnormal-return measurement, with 364,405 neutral-sentiment events providing a publicity placebo.A bootstrap resamples trading dates to assess significance. The measurement also examines pure publicity effects and how news moves volatility.Source: §1 · pp. 1–2
  • News-aligned price moves concentrate before and at publication, with rumor-flagged events capturing their entire move on the rumor day.These findings support both market sayings about news being priced in and buying rumors while selling news.Source: §1 · p. 2
  • Removing background benchmark drift reveals underreaction to quantified fundamental news and overreaction to soft stories, rather than widespread post-news reversal.The placebo removes residual benchmark drift present with or without news. This correction also reinterprets apparent good–bad news asymmetry and a class of news-fading strategies.Source: §1 · p. 2
  • A per-tag, per-attribute, per-size drift table provides a ranking and importance-sampling prior for news-conditioned forecasting models.The paper also contributes a reproducible recipe for distilling a large language model teacher into a compact corpus-labeling classifier.Source: §1 · p. 2

2 Related work

The study measures when news-aligned price moves occur rather than building a return predictor. It contributes event-level drift estimates and uses LLMs for classification rather than forecasting.

Sentiment and return prediction

  • Unlike sentiment-based return prediction, this study measures news-aligned moves in event time and uses placebo controls to avoid misleading interpretations of post-news drift.Reviews by Jadhav and Mirza and Cao et al. [1] describe the literature’s predictive focus. Chen et al. find short-horizon news momentum across 16 markets, persisting for days in small stocks, echoed by this study’s adjusted earnings and analyst drifts.Source: §2 · p. 2

Structured news events

  • 1.68 million events underpin a measured drift table that supplies event-driven forecasting systems with priors on which event tags matter and how quickly information decays.Li et al. [15] learn structured event representations, Wang et al. [20] maintain event memory, and Li et al. [16] train event-driven trading policies; this study estimates its prior rather than learning it end to end.Source: §2 · p. 2

Methodological foundations

  • Pangakis and Wolken [18] and Xia et al. [21] show that distilled LLM labels match human annotation quality at a fraction of the cost.The study applies this approach at industrial scale with per-tag validation. Khaledian et al. [12] similarly reduce retrieval costs over financial text.Source: §2 · p. 2
  • The study sidesteps forecast-related look-ahead risks by restricting its LLM to event tags and attributes and producing its corpus daily in live operation.Gao et al. [5] and Chen and Pu [4] document risks from forecasting with potentially memorized text; He et al. [8] address them through chronologically consistent models.Source: §2 · p. 2

3 Data

The dataset contains 4.57 million filtered financial-news articles covering roughly 3,000 US-listed stocks, with daily adjusted closing prices for 2,591 tradeable symbols. Live sentiment is generated without subsequent returns, while event tags are assigned retrospectively.

3 Data

  • 4.57 million deduplicated, financially relevant articles were retained from more than 70 million raw articles covering roughly 3,000 US-listed stocks.NewsWitch has operated daily since early 2024, processing over a quarter million news URLs per day from more than half a million sources.Source: §3 · p. 2
  • An LLM assesses each article’s stock relevance and importance, assigns five-level sentiment, and writes a summary.Retained articles include these fields, publication timestamps, and source domains.Source: §3 · p. 2
  • Live sentiment and summaries use no subsequent returns; retrospectively assigned event tags describe reported events rather than forecast outcomes.The teacher-only analysis in Section 5.3 shows that the findings do not depend on the classifier or its training vintage.Source: §3 · p. 2
  • 2,591 tradeable symbols have daily adjusted closing prices; 3.4% of articles reference symbols without price histories, and 9.7% of events lack required price windows.Events with missing price windows are skipped; delistings disproportionately affect smaller stocks.Source: §3 · p. 2
  • Under 3% of signed events come from the bulk-backfilled 2023 portion, which shows the same patterns as the live years.Source: §3 · p. 2

4 Method

The method combines distilled LLM article tagging, story-aware event construction, and beta-adjusted returns evaluated against neutral-news placebos. It describes when news-aligned price moves occur rather than identifying causal effects of news.

4.1 Distilled LLM tagging

  • 95% estimated corpus-wide teacher fidelity comes from combining the distilled classifier with teacher review of the 15% low-confidence tail and retaining existing teacher labels.An 82M-parameter student learns 17 event tags and five binary attributes through uncertainty sampling, retraining on 129,463 teacher labels. Teacher labeling costs $29 within an under-$200 budget including validation and arbitration; dedicated price-commentary and promotional categories keep non-events separate.Source: §4.1 · pp. 2–3

4.2 From articles to events

  • 55% of articles are follow-up coverage, identified through embedding-based clustering that separates first reports from continuing stories.Same-tag stock-day clusters use cosine similarity threshold 0.80, then match frozen story centroids with tag-specific lifetimes; stock-day-tag events roll post-16:00 New York articles forward and average article sentiment. The 1,681,657 price-scored events comprise 1,317,252 signed and 364,405 neutral events, with neutrality requiring every constituent article to be Neutral.Source: §4.2 · p. 3

4.3 Measurement design

  • Abnormal returns subtract rolling-beta-adjusted SPY returns, with sentiment-aligned event windows and trading-date bootstrap inference.Betas use 252 days with at least 126 observations, defaulting missing betas to 1; return windows are −5..0, day 0, +1..+5, and +6..+20. Inference uses 5,000 date-cluster bootstrap draws, and annual size buckets are dollar-volume terciles.Source: §4.3 · p. 3
  • The adjusted estimator subtracts size-matched neutral-event drift from abnormal returns before applying sentiment sign, isolating direction-aligned moves beyond the placebo baseline.Adjusted returns are s(AR_w - b_k,w), while raw returns omit baseline subtraction. Using the same benchmark for events and placebos is intended to cancel benchmark misfit.Source: §4.3 · p. 3
  • Tag-matched placebo baselines preserve earnings continuation and macro, leadership, and competition reversals, but shrink capital-returns estimates and move launch estimates across zero.The main baseline pools tags within size buckets, assuming tag-independent coverage drift. Re-estimating the baseline within every bootstrap draw produces unchanged or stronger headline significance, making the reported treatment of baseline noise conservative.Source: §4.3 · p. 3
  • +1.02%, versus +1.41% with all coverage, is the earnings pre-window return using only first reports; post-publication conclusions remain unchanged.Pre-publication drift mixes genuine anticipation with reporting on moves that already happened, so follow-up coverage inflates measured anticipation by roughly a third. The study is descriptive, not causal.Source: §4.3 · pp. 3–4

5 Results

News-aligned price moves concentrate before and at publication, while background drift explains much apparent post-news reversal. After adjustment, quantified fundamentals continue and soft news reverses, but directional trading gains are thin and volatility falls after publication.

5.3 What survives: the tag map

  • +0.35% capital returns continuation and −0.34% macro reversal anchor the adjusted tag map: quantified fundamentals drift, while soft news reverses over days +6 to +20.Earnings (+0.22%), guidance (+0.13%), and analyst actions (+0.10%) continue; product launches (−0.18%), leadership stories (−0.18%), and competition (−0.17%) reverse.Source: §5.3 · pp. 4–5, §5 · p. 4, Table 2 · p. 5

5.1 The move sits before and at publication

  • 2.8 times the day +20 cumulative news-aligned move is already reached by publication-day close: +0.58% versus +0.20%, with rumor events showing negligible confirmation gains.Among 18,618 rumor-flagged signed events, 94% receive same-tag confirmation within 60 trading days, with a median six-day gap. Rumor day delivers +0.36%, followed by −0.09% before confirmation, +0.01% on confirmation, and −0.06% over post-confirmation days +6 to +20; prior-month same-tag rumors also eliminate additional post-news drift.Source: §5.1 · p. 4, Figure 1 · p. 5, Table 3 · p. 6

5.2 The background drift and the placebo

  • 0.00% adjusted drift after both positive and negative news replaces raw −0.62% and +0.63% over days +6 to +20, exposing baseline-driven asymmetry.Neutral-covered stocks drift −0.92% among small caps, −0.58% among mid caps, and −0.34% among large caps over the following month; quiet days drift −0.74%, −0.88%, and −0.59%, respectively, identifying a single-beta benchmark residual rather than a publicity effect that the 73% positive-news share turns into apparent pooled reversal. Adjustment also removes the 2.4-fold raw size gradient and reduces legal/regulatory continuation from +0.29% (p<0.001) to −0.08% (p=0.22).Source: §5.2 · p. 4, Figure 3 · p. 6, Figure 2 · p. 6

5.3 What survives: the tag map

  • Three robustness cuts—NEW articles only, single-tag stock-days, and GPT-teacher labels—largely preserve the tag map’s signs and magnitudes.Analyst actions flip to −0.05% on single-tag days, an unusual subsample containing fewer than seven percent of analyst events; raw pooled reversal shrinks from −0.49% in 2024 to −0.18% in 2026, while adjusted pooled drift is zero from 2025 onward. At 5% Benjamini-Hochberg false discovery rate, macro reversal and capital returns and earnings continuations survive; analyst, leadership, and launch results are borderline (q between 0.05 and 0.07), and guidance is suggestive.Source: §5.3 · pp. 5–6, Table 4 · p. 6
  • 1.18 times normal publication-day absolute movement accompanies market-hours articles, compared with 1.05 for overnight publications; event attributes also distinguish anticipation and subsequent drift.Scheduled events show more anticipation and +0.08% adjusted drift (p=0.03), whereas unscheduled events show no adjusted drift; rumor-flagged events have the largest pre-publication windows. Quantified articles drift, while unquantified articles reverse by −0.12% (p=0.01).Source: §5.3 · pp. 6–7

5.4 Sources are not interchangeable

  • +0.6% to +1.5% pre-publication drift appears for retail-analysis sites, portals, mainstream media, and aggregators, versus −0.06% for press wires carrying releases.Commentary outlets cover moves already underway, whereas company and regulator releases constitute the event; wire-sourced positive tilt reverses by −0.19% adjusted, with marginal significance. The two most promotional individual outlets show −0.4% to −0.8% adjusted monthly drift, suggesting a source-quality signal requiring further study.Source: §5.4 · p. 7, Table 5 · p. 7

5.5 Second moments: news carries width

  • 0.86 of the EWMA forecast describes post-event realized volatility: uncertainty compresses after publication, although neutral news and heavier coverage still amplify absolute moves.Neutral guidance moves 1.29 times normal on publication day and 1.23 over the following week; neutral earnings move 1.11 times, while ten-plus-article days move 1.36 times versus 1.05 for single-article days. Post-event volatility ratios are 0.87 for neutral and 0.86 for signed events, motivating wider forecasts around scheduled disclosures and narrower forecasts afterward, independently of direction.Source: §5.5 · p. 7

5.6 Economic significance

  • 15.8% annualized gross from fading small-cap launch/partnership sentiment matches 15.9% from shorting all news-covered small caps; both have Sharpe 1.35, leaving no directional edge.These equally weighted, daily-rebalanced, beta-hedged portfolios hold from day +5 close to day +20 close; fading retains Sharpe 0.77 at 20 bps per side, whereas legal-news following loses its gross Sharpe 0.66 edge at 10 bps. Borrow fees and locate availability are unmodeled, and residual earnings continuation of +0.22% over 15 days is thin relative to plausible costs.Source: Table 6 · p. 7, Figure 5 · p. 7, §5.6 · p. 7

6 Implications for news-conditioned forecasting

News-conditioned forecasting systems should rank news using measured drift priors, encode tagged and flagged events rather than sentiment alone, and benchmark claimed alpha against a coverage-presence baseline.

Ranking

  • When compressing daily news into a short context, rank event lines by measured per-tag, per-attribute, per-size drift priors rather than source reputation heuristics.This makes measured drift, rather than source reputation, the basis for prioritizing limited context.Source: §6 · p. 8

Representation

  • Encode news as tagged, flagged event lines rather than sentiment scalars because direction is mostly priced in at publication while width effects persist.Include the event tag, scheduled status, rumor flag, primary-source flag, and NEW status or follow-up age; these tags and flags carry durable information.Source: §6 · p. 8

Evaluation

  • Benchmark claimed news alpha against a coverage-presence baseline to avoid mistaking the baseline itself for predictive signal.The comparison should use a coverage-presence baseline of the kind measured in this study.Source: §6 · p. 8

7 Limitations

Five limitations qualify the findings: imperfect labels, a single-beta return benchmark, daily timing within one market regime, public-news coverage, and simplified economic-significance assumptions.

7 Limitations

  • 2023 to 2026 spans a single market regime, and daily closes cannot separate trading-hours reactions from overnight gaps.Year-by-year decay suggests the residual patterns are already fading. The 2023 slice is backfilled rather than live.Source: §7 · p. 8
  • Public-news anticipation partly reflects information already circulating through filings, terminals, and professional channels before publicly accessible articles appear.The tag map identifies residual information despite public news being largely priced at publication. Kargarzadeh reports good results combining public news with macroeconomic indicators in a momentum-style framework.Source: §7 · p. 8
  • The single-market-beta benchmark does not explicitly model style effects such as size, although placebo adjustment within size buckets absorbs much of this limitation.A formal multi-factor treatment remains a natural next step.Source: §7 · p. 8
  • About one article in eight receives an event label that disagrees with the classifier’s teacher, potentially affecting story clustering.Disagreements concentrate in less consequential price-commentary and promotional-content categories, and the teacher-only check indicates conclusions do not depend on them. Sentiment and summaries also rely on the corpus vendor’s LLM.Source: §7 · p. 8
  • The economic-significance exercise uses fixed windows, equal weights, and flat costs to price measured patterns, not to constitute a trading system.Richer strategy design is left to future work.Source: §7 · p. 8

8 Conclusion

News-related price moves concentrate before and at publication. Placebo-adjusted results distinguish underreaction to quantified disclosures from overreaction to narratives, while volatility falls as publication resolves uncertainty.

8 Conclusion

  • 1.68 million tagged events show that news-related price moves occur before and at publication, with subsequent weeks adding nothing beyond what was already there.By the time a rumor is confirmed, the market has finished trading it; waiting for certainty means buying the top.Source: §8 · p. 8
  • Markets underreact to numbers and overreact to stories: quantified disclosures drift in the news direction for weeks, while narrative coverage gives back its move.This distinction remains after subtracting the drift of comparable placebo stocks.Source: §8 · p. 8
  • Subtracting comparable placebo stocks removes apparent good-news reversals, bad-news continuation, and profitable-looking sentiment-fading strategies caused by sample-wide underperformance against a standard beta benchmark.Stocks drifted below that benchmark whether they appeared in the news or not.Source: §8 · p. 8
  • Volatility falls once news is published because publication resolves uncertainty rather than creating it.Source: §8 · p. 8
  • For news-based forecasting and trading, durable information lies in event type, coverage presence and intensity, and width; direction is spent by the closing bell.Source: §8 · p. 8
Loading 2608.14014v1…