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

# Buy the Rumor, Sell the News: When Is News Priced In?
Source: [https://arxiv.org/html/2608.14014](https://arxiv.org/html/2608.14014)
## Buy the Rumor, Sell the News: When Is News Priced In?CCS:Applied computing EconomicsCCS:Computing methodologies Natural language processing

Alireza KargarzadehNote:Corresponding author\.Affiliation:Tailstate Intelligence Ltd,London,United Kingdomemail:[alireza\.kargarzadeh@tailstate\.ai](mailto:[email protected])Nariman KhaledianAffiliation:Independent Researcher,Antibes,Franceemail:[khaledian\.nariman@gmail\.com](mailto:[email protected]),Navid ParviniAffiliation:Zanista AI Ltd,London,United Kingdomemail:[navid\.parvini@zanista\.ai](mailto:[email protected]),Sid GhatakAffiliation:Increase Alpha, LLC,Miami,United Statesemail:[s\.ghatak@increasealpha\.com](mailto:[email protected])andArman KhaledianAffiliation:Zanista AI Ltd,London,United Kingdomemail:[arman\.khaledian@zanista\.ai](mailto:[email protected])

© none

###### 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\.8 times its value twenty 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\.

###### Keywords:

Financial news, Event study, LLM distillation, Market efficiency

## 1\.Introduction

Every trading day, thousands of news articles are written about individual stocks, and a growing share of quantitative systems read them\. Two old sayings surround this flow: that news is already priced in by the time it is published, and that the rumor is bought while the news is sold\. Taken literally, both locate the price move connected to a news event before and at publication rather than after it\. Watching a live news feed makes the question concrete: often enough, a stock receives plainly good news and its price falls anyway, or drifts down for weeks after upbeat coverage\. Seeing this repeatedly was the direct motivation for this study\. Whether that is accurate, for which kinds of news, and by how much are basic questions about how fast markets absorb public information, and their answers determine what news can and cannot contribute to forecasting and trading systems\.

These questions have stayed open at the scale of the full news flow because three layers of measurement infrastructure were missing\. The first is tagged events: an earnings report, a lawsuit, and a promotional listicle are different economic objects that move prices differently, so each article must be labeled with the kind of event it reports, at high accuracy and for millions of articles\. The second is story structure: the same event is syndicated and followed up for days, so repeated coverage must be grouped into one story before events can be counted\. The third is a benchmark for publicity itself: separating the effect of a news direction from the effect of merely being in the news requires knowing what happens to a stock that receives coverage with no direction at all\. The recent literature that applies large language models to financial news documents genuine return predictability, but it concentrates on prediction and builds none of these measurement layers\.

This paper builds the three layers and runs the measurement\. We introduce a taxonomy of 17 news tags and five attributes, designed in iterations: an earlier system in which roughly half a million articles carried 63 loosely defined tags showed, through the behavior of prices around each tag, which distinctions matter and which do not\. A two\-step pipeline tags the full corpus: 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 covering roughly 3,000 US stocks \(2023 to 2026\) at 95 percent fidelity to the teacher and negligible cost\. A clustering step then groups articles into stories, so first reports separate from follow\-up coverage\. On this foundation we measure, quantitatively: when news is priced in and to what extent the two sayings hold; what pure publicity does to prices, as distinct from positive or negative coverage; how news moves volatility; and how story\-driven news differs from fundamental news\. Returns are measured as beta\-adjusted abnormal returns around 1\.68 million \(stock, day, tag\) events, with significance from a bootstrap that resamples trading dates, and with 364,405 neutral\-sentiment events serving as the built\-in placebo for publicity\.

The two expressions hold, and the placebo reveals a third fact that reinterprets much of the raw picture\. The news\-aligned move sits overwhelmingly before and at publication, and for rumor\-flagged events the rumor day captures everything\. A drift that exists with or without news, the residual of the return benchmark over this sample, is behind what otherwise looks like widespread post\-news reversal; the placebo design removes it\. Net of it, markets underreact to quantified fundamental news and overreact to soft stories\. Section[5](https://arxiv.org/html/2608.14014#S5)develops each finding with the evidence\.

Contributions\.\(1\) A reproducible recipe for labeling multi\-million 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\. \(2\) The largest tagged, placebo\-controlled measurement of where news\-aligned price moves sit in event time that we are aware of\. \(3\) The background\-drift correction, which reinterprets post\-news reversal, the apparent asymmetry between good and bad news, and a class of news\-fading strategies\. \(4\) A per\-tag, per\-attribute, per\-size drift table usable as a ranking and importance\-sampling prior for news\-conditioned forecasting models\.

## 2\.Related work

Recent work on LLMs and equity news splits into three strands\. The first asks whether LLM\-extracted sentiment predicts returns:[Lopez\-Lira and Tang 2024](https://arxiv.org/html/2608.14014#bib.bib18)show ChatGPT headline scores predict next\-day returns,[Chen et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib3)extend the exercise to market and macro prediction,[Kirtac and Germano 2024](https://arxiv.org/html/2608.14014#bib.bib15)trade LLM sentiment signals, and[Iacovides et al\. 2024](https://arxiv.org/html/2608.14014#bib.bib10)fine\-tune open models for the same task;[Kim et al\. 2024](https://arxiv.org/html/2608.14014#bib.bib14)apply LLMs to fundamental analysis and[Papasotiriou et al\. 2024](https://arxiv.org/html/2608.14014#bib.bib20)to stock ratings, and[Ghafouri et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib7)probe the behavioral signatures of the models themselves\.[Kargarzadeh 2024](https://arxiv.org/html/2608.14014#bib.bib12)and[Ghatak et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib8)combine LLM news signals with macroeconomic and technical indicators in trading frameworks\.[Chen et al\. 2024](https://arxiv.org/html/2608.14014#bib.bib4)extract news embeddings across 16 markets and find short\-horizon news momentum that persists for days in small stocks, a finding our adjusted earnings and analyst drifts echo from the event side\.[Jadhav and Mirza 2025](https://arxiv.org/html/2608.14014#bib.bib11)review LLM applications in equity markets and find the literature concentrated on sentiment extraction and return prediction;[Cao et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib2)draw the same picture for quantitative investment more broadly, with news entering models as a predictive feature rather than as an object of measurement\. Our question is different: we do not build a predictor, we measure where in event time the news\-aligned move sits, and our placebo design shows that a naive reading of post\-news drift would mislead exactly the strategies this strand builds\.

The second strand structures news into events rather than sentiment scalars\.[Li et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib16)learn structured event representations for return prediction,[Wang et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib21)maintain an event memory for forecasting, and[Li et al\. 2026](https://arxiv.org/html/2608.14014#bib.bib17)train end\-to\-end event\-driven trading policies\. These systems need to know which event tags matter and how fast their information decays; our measured drift table is exactly that prior, estimated on 1\.68 million events rather than learned end to end\.

The third strand is methodological\.[Pangakis and Wolken 2024](https://arxiv.org/html/2608.14014#bib.bib19)and[Xia et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib22)show that distilling LLM\-generated labels into small supervised classifiers matches human annotation quality at a fraction of the cost; our tagging pipeline is an industrial\-scale instance with per\-tag validation, and[Khaledian et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib13)cut the cost of retrieval over financial text along similar lines\.[Gao et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib6)and[Chen and Pu 2026](https://arxiv.org/html/2608.14014#bib.bib5)document look\-ahead risks when LLMs generate forecasts from text they may have memorized, and[He et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib9)train chronologically consistent models to avoid them; our design sidesteps the problem, since the LLM only assigns event tags and attributes, never forecasts, and the corpus is produced daily in live operation\.

## 3\.Data

The corpus comes from NewsWitch, a commercial financial\-news product that has run in daily production since the beginning of 2024\.111For access to NewsWitch data, contact info@zanista\.ai\.Each day it processes over a quarter million news URLs concerning the roughly 3,000 most\-covered US\-listed stocks, drawn from a crawl of more than half a million sources; over the sample this amounts to more than 70 million raw articles\. The 4\.57 million retained here are the important ones: deduplicated, financial in nature, and directly related to the covered stock universe\. For every crawled article, the title, body, and metadata are read by an LLM that decides whether the article is genuinely related to the stock and important enough to keep, assigns a sentiment direction on five levels from strongly negative to strongly positive, and writes a summary; retained articles carry these fields along with the publication timestamp and source domain\. The corpus has supported earlier studies of LLM\-driven trading and retrieval\([Kargarzadeh 2024](https://arxiv.org/html/2608.14014#bib.bib12);[Ghatak et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib8);[Khaledian et al\. 2025](https://arxiv.org/html/2608.14014#bib.bib13)\)\. Sentiment and summaries in the live period are produced day by day, with no access to subsequent returns\. Our event tags were assigned retrospectively, by a classifier trained later; they are descriptive labels of what an article reports, not forecasts, and the teacher\-only cut in Section[5\.3](https://arxiv.org/html/2608.14014#S5.SS3)shows the findings do not depend on the classifier or its training vintage\. The 2023 portion was backfilled in bulk, is under 3% of signed events, and shows the same patterns as the live years\. Prices are daily adjusted closes for 2,591 tradeable symbols; 3\.4% of articles reference symbols without price history \(delistings skew small\), and 9\.7% of events are skipped for missing price windows\.

## 4\.Method

Our method has three stages: tag every article with a distilled LLM tagger, group articles into stories and events, and measure abnormal returns around those events against a placebo\.

### 4\.1\.Distilled LLM tagging

A frozen prompt presents 17 event\-tag definitions, five binary attribute definitions \(scheduled, forward looking, primary source, quantified, rumor\), and four example articles with their completed labels to a small GPT teacher \(gpt\-5\-mini, hosted on Azure\) with a strict JSON schema, at about $0\.0002 per article\. The taxonomy covers real corporate events \(earnings, guidance, analyst actions, launches, M&A, legal and regulatory, and so on\) and deliberately includes two junk categories, price commentary and promotional content, so that articles describing price action or promotion have somewhere to go other than a real event tag\. The teacher labeled 29,472 random articles; a distilroberta\-base student \(82M parameters, seven heads: primary tag, secondary tag, five attributes\) trained on these labeled 600,000 fresh articles; the 100,000 lowest\-confidence articles went back to the teacher \(uncertainty sampling, with a quota for rare tags\); and the student was retrained on the merged 129,463 labels\. On a held\-out 10,000\-article sample labeled independently by both, the student agrees with the teacher on 87\.5% of primary tags overall and 94\.0% at confidence≥0\.8\\geq 0\.8\(85% of articles\); attribute agreement is 92–99%\. Deployment uses the classifier everywhere, routes the 15% low\-confidence tail back to the teacher, and keeps teacher labels where they exist, for an estimated 95% corpus\-wide teacher fidelity at a total cost of $29 for teacher labels within an under\-$200 budget including validation and arbitration\. Every article records its label source, which powers a robustness check below\. Inference runs at 125–160 articles per second on a MacBook Pro laptop \(M4 Pro, 24 GB RAM\)\.

### 4\.2\.From articles to events

Bundling repeated coverage into distinct events is a hard problem in its own right, and we solve it with a two\-step, embedding\-based clustering\. Within each stock\-day, articles are clustered on the cosine similarity of their title\-and\-summary embeddings \(threshold 0\.80\), restricted to the same primary tag; each day\-cluster is then matched against the stock’s running stories, so an event continues across days when later coverage stays close to the story’s frozen centroid\. The clustering is what turns articles into events for the event study, and it also quantifies how much news repeats itself: it reveals that 55% of all articles are follow\-up coverage of a story already running, and it flags each article as the first report \(NEW\) or follow\-up\. How long a story may stay open depends on its tag: we calibrated per\-tag lifetimes by tracking how the similarity of later same\-story coverage to the first\-day centroid decays with story age, and capped each tag where continued matches become rare\. M&A and legal sagas stay open for up to 90 days, leadership and operations stories for weeks, commentary for two days; measured persistence per tag is reported in Table[1](https://arxiv.org/html/2608.14014#S4.T1)\. For the event study, an event is a \(stock, trading day, tag\) aggregate: articles published after 16:00 New York time roll to the next trading day, and the article count is retained as an attention measure\. The event’s impact comes from the vendor sentiment: each article carries one of five levels, and the event’s impact averages its articles\. An event is neutral when every one of its articles is Neutral, and signed otherwise\. The drift tables use only the sign of the impact\. This yields 1,862,297 events, of which 1,681,657 are scored against prices: 1,317,252 signed \(73% positive\) and 364,405 neutral\.

Table 1\.Story persistence by tag, measured from the clustering: number of stories and mean lifetime in trading days\.
### 4\.3\.Measurement design

Letrtr\_\{t\}be the stock’s daily return andrm,tr\_\{m,t\}the daily return of the S&P 500 index, measured through SPY, the exchange\-traded fund that tracks it\. The abnormal return isA​Rt=rt−βt​rm,tAR\_\{t\}=r\_\{t\}\-\\beta\_\{t\}\\,r\_\{m,t\}withβt\\beta\_\{t\}a rolling 252\-day OLS beta of the stock against the S&P 500, as of daytt\(minimum 126 observations, missing betas set to 1\)\. Around each event day we cumulateA​RARover four windows: days−5\.\.0\-5\.\.0, day 0, days\+1\.\.\+5\+1\.\.\+5, and days\+6\.\.\+20\+6\.\.\+20\. Signed events are stacked by sentiment signs∈\{−1,\+1\}s\\in\\\{\-1,\+1\\\}, so a move in the direction of the news counts as positive whether the news was good or bad\. Significance comes from a cluster bootstrap that resamples trading dates \(5,000 draws\), because events on the same day are cross\-sectionally correlated\. For size buckets, stocks are split each year into three equal\-sized groups \(small, mid, large\) by dollar volume\.

The placebo and the adjusted estimator\.A single\-beta benchmark is deliberately simple: transparent, replicable, and standard\. No fixed benchmark tracks every stock perfectly, and whatever residual drift a benchmark leaves behind flows into every event window measured against it\. The design therefore includes a placebo rather than a bet on the benchmark\. Neutral events carry coverage without direction, so their drift measures what happens to comparable stocks of the same size around news days in general\. Call that the baselinebk,wb\_\{k,w\}, the mean abnormal return of neutral events in size bucketkkover windowww\. When we want the effect of the news direction itself, we subtract the baseline from the event’s abnormal return before applying the sign:s⁡\(A​Rw−bk,w\)s\\,\(AR\_\{w\}\-b\_\{k,w\}\)\. Throughout the paper,*adjusted*means exactly this, the move in the news direction in excess of what comparable stocks drifted anyway, and*raw*means no subtraction\. Because the baseline is measured against the same benchmark as the events, any benchmark misfit appears on both sides of the subtraction and cancels: a richer factor model would shrink the baseline itself but leave the adjusted estimates essentially unchanged\. Raw and adjusted columns are reported side by side\.

Two implementation details matter for inference\. First, the baseline is pooled across event tags within a size bucket, which assumes coverage drift is tag\-independent\. A tag\-matched variant, which subtracts each tag’s own neutral baseline per bucket, leaves the earnings continuation and the macro, leadership, and competition reversals essentially unchanged, while the capital\-returns estimate shrinks and the launch estimate crosses zero; neutral coverage of those two tags selects unusual stocks \(dividend payers, promoted small caps\), so their magnitudes are more model\-dependent than their signs\. Second, the baseline is estimated once, on the full neutral sample; a bootstrap that re\-estimates it inside every draw gives the same or slightly stronger significance for every headline tag, so ignoring baseline noise in the reported p\-values is conservative\.

What this design can and cannot claim\.Sentiment is assigned to articles, and articles are written about moves as well as before them, so the pre\-publication drift mixes genuine anticipation with reporting on moves that already happened\. We quantify this with the NEW\-only rerun: restricting to first reports shrinks the earnings pre\-window from\+1\.41%\+1\.41\\%to\+1\.02%\+1\.02\\%, so follow\-up coverage inflates measured anticipation by roughly a third, and all post\-publication conclusions are unchanged\. The study is a descriptive account of where the news\-aligned move sits in event time, not a causal claim about news moving prices\.

## 5\.Results

Our main findings are as follows\. First, the news\-aligned move sits before and at publication, and rumor\-flagged events show the buy the rumor, sell the news pattern literally \(Section[5\.1](https://arxiv.org/html/2608.14014#S5.SS1)\)\. Second, a background drift that exists with or without news explains most of the raw post\-news picture \(Section[5\.2](https://arxiv.org/html/2608.14014#S5.SS2)\)\. Third, net of the baseline, quantified fundamental news drifts and soft news reverses, and the pattern survives every robustness cut \(Section[5\.3](https://arxiv.org/html/2608.14014#S5.SS3)\)\. The remaining sections measure source differences, volatility, and economic significance\.

### 5\.1\.The move sits before and at publication

Table[2](https://arxiv.org/html/2608.14014#S5.T2)is the master table; Figure[1](https://arxiv.org/html/2608.14014#S5.F1)\(left\) shows the event\-time picture\. Almost every pre and day\-0 column in the raw table is positive: whatever the market does around news, it agrees in direction with the news before the news is out\. For the quantified fundamental tags the pattern is extreme: earnings\+1\.41%\+1\.41\\%pre and\+0\.42%\+0\.42\\%on the day against−0\.05%\-0\.05\\%afterwards; analyst actions\+1\.12%\+1\.12\\%and\+0\.22%\+0\.22\\%against−0\.13%\-0\.13\\%; price commentary, guidance, and competition all show the same shape\. A useful summary comes from the running total plotted in Figure[1](https://arxiv.org/html/2608.14014#S5.F1): starting five days before publication, add up each day’s average abnormal return in the news direction\. Pooled across all signed events, this running total reaches\+0\.58%\+0\.58\\%by the close of publication day, then falls back and ends day\+20\+20at\+0\.20%\+0\.20\\%\. The ratio of the two levels is2\.82\.8: by the closing bell of publication day the market had already moved almost three times as far as where it would stand a month later, and the extra ground was given back over the following weeks\. The ratio measures how complete the move was, not how large: earnings events have the largest moves but almost no give\-back, so their ratio is 1\.06, against 1\.20 for analyst actions and 1\.08 for guidance; at or above one throughout, the move was finished, or more than finished, when publication day ended\. The pooled ratio is higher than any single tag’s because the pooled after\-window also inherits the background drift examined in Section[5\.2](https://arxiv.org/html/2608.14014#S5.SS2)\.

Table 2\.Signed abnormal drift by event tag, sorted by adjusted post\-event drift\. All columns are percent, in the direction of the news\. “%pos” is the share of positive\-sentiment events\. “adj\.” subtracts the baseline of the event’s size bucket before signing \(Section[4\.3](https://arxiv.org/html/2608.14014#S4.SS3)\);padjp\_\{\\text\{adj\}\}is the date\-bootstrap p\-value of the adjusted days\+6\.\.\+20\+6\.\.\+20drift\.Figure 1\.Cumulative abnormal return in the news direction around publication \(day 0\), five event tags plus the neutral placebo\. Left: raw\. Right: net of the background drift\. Bands are±2\\pm 2date\-clustered standard errors\. Raw curves show the priced\-in signature \(the move precedes publication\) and near\-universal post\-news decay tracking the placebo\. Adjusted curves separate genuine continuation \(earnings, analyst\) from genuine reversal \(launches, macro\), and flatten legal and regulatory news to the baseline\.The rumor attribute makes the expression measurable \(Table[3](https://arxiv.org/html/2608.14014#S5.T3)\)\. Among 18,618 rumor\-flagged signed events, 94% are followed by a non\-rumor event with the same tag within 60 trading days \(median gap 6 days\)\. The rumor day delivers\+0\.36%\+0\.36\\%in the rumor’s direction\. Between rumor and confirmation the stock adds nothing \(−0\.09%\-0\.09\\%\)\. The confirmation day itself is worth\+0\.01%\+0\.01\\%, and days \+6 to \+20 after confirmation give back−0\.06%\-0\.06\\%\. M&A rumors are the sharpest case:\+0\.24%\+0\.24\\%on rumor day, then−0\.32%\-0\.32\\%into the confirmation,−0\.10%\-0\.10\\%on the news, and−0\.37%\-0\.37\\%after it\. Whoever traded the rumor captured the entire move; whoever bought the confirmation bought the top\. Consistent with this, events preceded by a same\-tag rumor in the prior month show no additional drift after the news\.

Table 3\.Buy the rumor, sell the news\. Mean abnormal return in the rumor’s direction \(percent\) at each stage, for rumor\-flagged events with a same\-tag non\-rumor event within 60 trading days\.
### 5\.2\.The background drift and the placebo

The placebo was designed as a calibration check and became a finding\. Neutral events measure what happens to a stock around news days that carry no direction; quiet stock\-days, defined as days with no event for the stock and none in the prior five trading days, measure the same windows with no news at all\. Both drift down\. Over the following month, stocks with neutral coverage trail their beta benchmark by−0\.92%\-0\.92\\%among small caps,−0\.58%\-0\.58\\%among mid caps, and−0\.34%\-0\.34\\%among large caps; quiet days drift−0\.74%\-0\.74\\%,−0\.88%\-0\.88\\%, and−0\.59%\-0\.59\\%in the same buckets \(Figure[2](https://arxiv.org/html/2608.14014#S5.F2)\)\. The two sets of numbers tell one story\. The drift is not caused by publicity: it is the residual of the single\-beta benchmark over this sample, present with news and without, of the same order across size groups, with covered small caps adding a modest0\.2%0\.2\\%per month on top and covered mid and large caps being negligible\. This is exactly why the design subtracts a placebo instead of trusting any benchmark: news studies inherit a drift that exists with or without news, and any measurement or strategy that scores news direction without a placebo of comparable stocks will misread that drift as directional information\.

Figure 2\.The background drift: abnormal returns of stocks receiving neutral news, by window and size group\. Quiet stock\-days with no recent news drift similarly \(Section[5\.2](https://arxiv.org/html/2608.14014#S5.SS2)\): the drift is the residual of the single\-beta benchmark over this sample, not an effect of publicity, and the placebo adjustment removes it either way\.To see how this distorts directional measurements, take an average positive\-news event in a small cap\. Its raw post\-event drift is the sum of two parts: the market’s reaction to the news direction, plus the−0\.9%\-0\.9\\%that any covered small cap drifts over the following month anyway\. Measured in the direction of the news, that second part shows up as*reversal*of good news\. Now take a negative\-news event in the same stock: the same−0\.9%\-0\.9\\%drift now points in the same direction as the news, so it shows up as*continuation*of bad news\. One underlying drift, two apparent behaviors\. Because 73% of signed events are positive, the pooled average is dominated by the first case, which is why the raw table looks like widespread reversal\.

The adjusted estimator of Section[4\.3](https://arxiv.org/html/2608.14014#S4.SS3)removes exactly this confound, and three apparent anomalies disappear at once \(Figure[3](https://arxiv.org/html/2608.14014#S5.F3)\)\. The apparent asymmetry between good and bad news: raw drift over days \+6 to \+20 is−0\.62%\-0\.62\\%after positive news \(looks like overreaction to good news\) and\+0\.63%\+0\.63\\%after negative news \(looks like underreaction to bad news\); after the baseline is subtracted, both become0\.00%0\.00\\%\(p≈0\.9p\\approx 0\.9\)\. The size gradient: small caps appear to reverse 2\.4 times more than large caps in the raw data \(−0\.41%\-0\.41\\%against−0\.17%\-0\.17\\%per month\), and adjusted, the gradient is gone\. And the one apparent continuation winner in the raw table, legal and regulatory news \(\+0\.29%\+0\.29\\%,p<0\.001p<0\.001\): legal news is the only tag where most events are negative \(83%\), so for this tag the baseline reads as continuation rather than reversal; adjusted, legal news drifts−0\.08%\-0\.08\\%\(p=0\.22p=0\.22\), indistinguishable from any other covered stock\.

Figure 3\.The asymmetry illusion\. Raw post\-news drift suggests good news reverses and bad news continues\. Once the baseline is subtracted, both collapse to zero: the apparent asymmetry is the baseline interacting with the 73% positive share of news sentiment\.
### 5\.3\.What survives: the tag map

Figure[4](https://arxiv.org/html/2608.14014#S5.F4)and the adjusted columns of Table[2](https://arxiv.org/html/2608.14014#S5.T2)give the residual structure, and it is orderly\. On the continuation side sit the quantified fundamental tags: capital returns\+0\.35%\+0\.35\\%, earnings\+0\.22%\+0\.22\\%, guidance\+0\.13%\+0\.13\\%, analyst actions\+0\.10%\+0\.10\\%over days \+6 to \+20, all significant; the earnings figure is the familiar post\-earnings\-announcement drift recovered at corpus scale from tagged news alone\. On the reversal side sit the soft, attention\-driven tags: macro read\-throughs−0\.34%\-0\.34\\%, product launches−0\.18%\-0\.18\\%, leadership stories−0\.18%\-0\.18\\%, competition−0\.17%\-0\.17\\%\. Launch and partnership coverage is 96–97% positive, the most promotional corner of the corpus, and it is precisely where prices overshoot\.

Figure 4\.The tag map: adjusted drift over days \+6 to \+20 in the news direction \(bars; pale bars not significant at 5%\) against raw drift \(circles\)\. Adjustment moves every tag toward zero and reorders them: quantified fundamentals drift, soft news reverses\.Table[4](https://arxiv.org/html/2608.14014#S5.T4)stress\-tests the map with three cuts\. The first keeps only NEW articles \(1\.09M events\), removing all follow\-up coverage\. The second addresses simultaneous stories: a covered stock\-day typically carries more than one event tag at once \(the median is two\), so drift attributed to a launch could in principle belong to a co\-occurring earnings story; restricting to stock\-days with exactly one event tag removes that contamination\. The third addresses classifier error: recomputing only on events whose articles were labeled directly by the GPT teacher, bypassing the distilled classifier entirely, tests whether tagging mistakes create the patterns\. Across the three cuts the headline tags keep their signs and similar magnitudes, with one exception: analyst actions flip to−0\.05%\-0\.05\\%on single\-tag days\. The exception has a natural reading: analyst notes rarely appear without other same\-day coverage \(fewer than seven percent of analyst events sit on single\-tag days\), so that cell is a small and unusual subsample\. Single\-tag days otherwise strengthen the survivors \(capital returns\+0\.47%\+0\.47\\%, launches−0\.48%\-0\.48\\%, macro−0\.52%\-0\.52\\%\)\. By year, the raw pooled reversal halves from 2024 \(−0\.49%\-0\.49\\%\) to 2026 \(−0\.18%\-0\.18\\%\), and the adjusted pooled drift is zero from 2025 on, consistent with a market that learns\. Finally, because seventeen tags are tested at once, we apply Benjamini\-Hochberg control to the adjusted p\-values of Table[2](https://arxiv.org/html/2608.14014#S5.T2): at a 5 percent false discovery rate the macro reversal and the capital\-returns and earnings continuations survive, the analyst, leadership, and launch results sit at the boundary \(qqbetween 0\.05 and 0\.07\), and the guidance continuation is suggestive rather than decisive\.

Table 4\.Robustness of the adjusted days\+6\.\.\+20\+6\.\.\+20drift \(percent\)\. Columns: all events; NEW articles only; stock\-days with one event tag; events labeled directly by the GPT teacher\.Attributes slice the same way \(not tabulated for space\): scheduled events show more anticipation and slightly positive adjusted drift \(\+0\.08%\+0\.08\\%,p=0\.03p=0\.03\), unscheduled events none; quantified articles drift, unquantified ones reverse \(−0\.12%\-0\.12\\%,p=0\.01p=0\.01\); rumor\-flagged events show the largest pre\-windows\. Publication timing matters for the day itself: articles published during market hours land on a day that moves 1\.18 times the stock’s normal absolute move, against 1\.05 for overnight publications\.

### 5\.4\.Sources are not interchangeable

Grouping events by the kind of outlet that wrote them \(Table[5](https://arxiv.org/html/2608.14014#S5.T5)\) separates anticipation from information\. Retail\-analysis sites, portals, mainstream media, and aggregators all show\+0\.6%\+0\.6\\%to\+1\.5%\+1\.5\\%of pre\-publication drift: they write about moves in progress\. Press wires, which carry company and regulator releases, show none at all \(−0\.06%\-0\.06\\%\): the release is the event, not commentary on one\. Wire\-sourced positive tilt also reverses hardest \(−0\.19%\-0\.19\\%adjusted, marginal\), consistent with promotional press releases overselling\. At the level of individual outlets, which we leave unnamed, the two with the most promotional catalogs show economically large negative adjusted drift \(−0\.4%\-0\.4\\%to−0\.8%\-0\.8\\%per month\), suggesting a source\-quality signal that survives the baseline; we leave a full source\-reputation study to future work\.

Table 5\.News source groups\. NEW share is the fraction of articles that are the first report of their story; day\-0\|A​R\|\|AR\|ratio is the publication\-day absolute move as a multiple of the stock’s normal day\. Events here are \(stock, day, tag, source group\) aggregates, so a story covered by several source groups appears once per group; counts therefore exceed the number of signed events\.
### 5\.5\.Second moments: news carries width

Neutral news is directionally empty but not informationless\. Neutral guidance events come with publication\-day moves 1\.29 times the stock’s normal day and 1\.23 times over the following week; neutral earnings 1\.11 times\. Attention scales the effect: days with ten or more articles move 1\.36 times normal against 1\.05 for single\-article days\. All of these are ratios of a stock’s absolute move to its own typical day, so the background drift of Section[5\.2](https://arxiv.org/html/2608.14014#S5.SS2), tiny at the daily scale, does not enter them\. After events, realized volatility runs about 0\.86 of the EWMA forecast \(0\.87 for neutral, 0\.86 for signed events\), so news resolves uncertainty: variance is elevated into the event and compresses after it, the second\-moment mirror of “priced in\.” A distributional forecaster should therefore read event tags as width signals, widening on scheduled disclosure events and narrowing after them, independent of direction\.

### 5\.6\.Economic significance

Table[6](https://arxiv.org/html/2608.14014#S5.T6)prices the two raw patterns a practitioner would be tempted to trade\. Both are calendar\-time portfolios built the same way: a position opens at the close of day \+5 after a qualifying event and closes at the close of day \+20 \(the window where the drift lives\), all open positions are equally weighted and rebalanced daily, and returns are measured on beta\-hedged abnormal returns, so the portfolios are market\-neutral overlays\. Concretely, each position is a stock leg plus an offsetting beta\-sized position in the index, so the returns and Sharpe ratios reported here are the realized returns of that hedged implementation\. The first strategy*fades*small\-cap launch and partnership news: after a positive\-sentiment launch or partnership event in a small cap it shorts the stock, after a negative\-sentiment one it buys, betting that the sentiment move reverses\. It earns 15\.8% annualized gross with a Sharpe ratio of 1\.35 and survives 20 bps of cost per side \(Sharpe 0\.77\)\. The benchmark strategy ignores sentiment entirely: it shorts*every*small\-cap stock that appeared in the news at all, neutral or signed, over the same windows\. It earns 15\.9% with the same Sharpe of 1\.35, and the two cumulative curves lie on top of each other \(Figure[5](https://arxiv.org/html/2608.14014#S5.F5)\)\. The fading strategy’s entire edge is the background drift; the news direction contributes nothing\. Following legal and regulatory news direction, the raw table’s continuation candidate, earns a Sharpe of 0\.66 gross that dies at 10 bps, as the composition argument predicts\. Both headline strategies are predominantly short books, and the cost model charges only a flat fee per side: borrow fees and locate availability are not modeled, and for small caps they would consume a further slice of the returns\. The residual tag\-map alphas \(earnings continuation at\+0\.22%\+0\.22\\%per event over 15 days\) are real but thin relative to plausible costs\. At universe scale, the exploitable object in news flow is presence and width, not direction\.

Table 6\.Calendar\-time portfolios, days \+6 to \+20, abnormal returns, 2023 to 2026\. Net Sharpe ratios charge the stated cost per side\.Figure 5\.Cumulative returns of the calendar\-time portfolios, each hedged with a beta\-sized index position, next to an S&P 500 buy and hold\. The sentiment\-fading strategy \(orange\) is indistinguishable from shorting every covered small cap \(dashed\): the news direction adds nothing\.

## 6\.Implications for news\-conditioned forecasting

These measurements bear directly on the design of news\-conditioned forecasting systems, in three ways\. First, ranking: when a day’s news must be compressed into a short context, event lines should be ordered by the measured per\-tag, per\-attribute, per\-size drift priors rather than by source reputation heuristics\. Second, representation: since direction is mostly priced in at publication while width effects persist, the text should be encoded as tagged, flagged event lines \(tag, scheduled, rumor, primary source, NEW or follow\-up age\) rather than as sentiment scalars; the tag and flags carry the durable information\. Third, evaluation: any claimed news alpha should be benchmarked against a coverage\-presence baseline of the kind measured here, or it will rediscover the baseline and call it signal\.

## 7\.Limitations

Five limitations frame how far these results reach, and each points at follow\-up work\. The first concerns labels\. Sentiment and summaries come from the corpus vendor’s LLM, and our event tags are layered on top by a distilled classifier that disagrees with its teacher on about one article in eight\. Those disagreements concentrate in the two junk categories, price commentary and promotional content, where a wrong label matters least, and the teacher\-only check in Table[4](https://arxiv.org/html/2608.14014#S5.T4)shows the conclusions do not depend on them\. Still, a mislabeled article can end up attached to the wrong story, so cleaner labels would sharpen the clustering as well\. The second concerns the return benchmark\. Each stock is hedged with a single market beta, so style effects such as size are not explicitly modeled; the placebo adjustment absorbs much of this, because the baseline is measured within size buckets, but a formal multi\-factor treatment is a natural next step\. Third, timing\. Daily closes cannot separate the trading\-hours reaction from the overnight gap, and the sample covers 2023 to 2026, a single market regime whose year\-by\-year decay suggests the residual patterns are already fading; the 2023 slice is backfilled rather than live\. Fourth, the corpus is public news, the news available on the open internet, which is what most market participants actually read\. It is in the nature of public information that by the time an event is written up in an article anyone can access, some of it has already moved through filings, terminals, and professional channels; this holds for any collection of public news, and part of what we measure as anticipation is this property of public information itself\. Public news is largely priced by the time it appears, but it is not empty:[Kargarzadeh 2024](https://arxiv.org/html/2608.14014#bib.bib12)builds trading strategies on public news of this kind, combined with macroeconomic indicators in a momentum\-style framework, and reports good results, and our tag map shows where such residual information lives\. Finally, the economic significance exercise is deliberately simple: fixed windows, equal weights, flat costs\. It is built to price the measured patterns, not to be a trading system\. Richer strategies can certainly be developed on this data; here the focus is on how news is priced, and we leave strategy design to future work\.

## 8\.Conclusion

In this study, we asked when news is priced in, and we answered it at the scale of the full news flow\. Both market sayings survive the test\. Across 1\.68 million tagged events, the price move connected to a news event lives before and at publication: anticipation builds over the preceding days, publication day absorbs what remains, and the weeks after add nothing that was not already there\. The rumor version holds in its most literal form: by the time a rumor is confirmed, the market has finished trading it, and whoever waited for certainty bought the top\.

The measurement also shows how easily news gets more credit than it deserves\. Stocks drifted below a standard beta benchmark across this sample whether they were in the news or not, and without a placebo of comparable stocks that drift reads as reversal of good news, continuation of bad news, and profitable\-looking strategies that fade news sentiment\. Subtracting the placebo makes these illusions disappear together, and what remains is our simplest lesson about how news is priced: markets underreact to numbers and overreact to stories\. Hard, quantified disclosures keep drifting in their own direction for weeks after publication; soft narrative coverage gives back its move; and volatility falls once the news is out, because publication resolves uncertainty rather than creating it\.

For anyone building forecasting or trading systems on news, the durable information is the kind of event, the fact and intensity of coverage, and its width; the direction is spent by the closing bell\.

###### Acknowledgements\.

This work was carried out at TailState Intelligence Ltd\. The NewsWitch news corpus and its data infrastructure were provided by Zanista AI Ltd\. We thank the Zanista AI team for data access and support\.

## References

- \(1\)
- Cao et al\.\(2025\)Bokai Cao, Saizhuo Wang, Xinyi Lin, Xiaojun Wu, Haohan Zhang, Lionel M\. Ni, and Jian Guo\. 2025\.From Deep Learning to LLMs: A Survey of AI in Quantitative Investment\.*arXiv preprint arXiv:2503\.21422*\(2025\)\.
- Chen et al\.\(2025\)Jian Chen, Guohao Tang, Guofu Zhou, and Wu Zhu\. 2025\.ChatGPT and DeepSeek: Can They Predict the Stock Market and Macroeconomy?*arXiv preprint arXiv:2502\.10008*\(2025\)\.
- Chen et al\.\(2024\)Yifei Chen, Bryan T\. Kelly, and Dacheng Xiu\. 2024\.Expected Returns and Large Language Models\.SSRN working paper 4416687, revised\.
- Chen and Pu \(2026\)Zefeng Chen and Darcy Pu\. 2026\.Autonomous Market Intelligence: Agentic AI Nowcasting Predicts Stock Returns\.*arXiv preprint arXiv:2601\.11958*\(2026\)\.
- Gao et al\.\(2025\)Zhenyu Gao, Wenxi Jiang, and Yutong Yan\. 2025\.Detecting Lookahead Bias in LLM Forecasts\.*arXiv preprint arXiv:2512\.23847*\(2025\)\.
- Ghafouri et al\.\(2025\)M\. Ghafouri, N\. Yousefi, A\. Akbari, R\. Tavakoli, F\. Davoodi, G\. Aminian, et al\.2025\.Risk, Ambiguity, and Infinity: Behavioral Signatures of Modern Large Language Models\. In*Proceedings of the ACM International Conference on AI in Finance \(ICAIF\)*\.
- Ghatak et al\.\(2025\)S\. Ghatak, Arman Khaledian, N\. Parvini, and Nariman Khaledian\. 2025\.Increase Alpha: Performance and Risk of an AI\-Driven Trading Framework\.*arXiv preprint arXiv:2509\.16707*\(2025\)\.
- He et al\.\(2025\)Songrun He, Linying Lv, Asaf Manela, and Jimmy Wu\. 2025\.Chronologically Consistent Large Language Models\.*arXiv preprint arXiv:2502\.21206*\(2025\)\.
- Iacovides et al\.\(2024\)Giorgos Iacovides et al\.2024\.FinLlama: LLM\-Based Financial Sentiment Analysis for Algorithmic Trading\. In*Proceedings of the 5th ACM International Conference on AI in Finance \(ICAIF\)*\.
- Jadhav and Mirza \(2025\)Anand Jadhav and Vaqar Mirza\. 2025\.Large Language Models in Equity Markets: Applications, Techniques, and Insights\.*Frontiers in Artificial Intelligence*8 \(2025\), 1608365\.
- Kargarzadeh \(2024\)Alireza Kargarzadeh\. 2024\.*Developing and Backtesting a Trading Strategy Using Large Language Models, Macroeconomic and Technical Indicators*\.Master’s thesis\. Imperial College London\.
- Khaledian et al\.\(2025\)Arman Khaledian, A\. Ghadiridehkordi, and Nariman Khaledian\. 2025\.PCA\-RAG: Principal Component Analysis for Efficient Retrieval\-Augmented Generation\.*arXiv preprint arXiv:2504\.08386*\(2025\)\.
- Kim et al\.\(2024\)Alex Kim, Maximilian Muhn, and Valeri Nikolaev\. 2024\.Financial Statement Analysis with Large Language Models\.*arXiv preprint arXiv:2407\.17866*\(2024\)\.
- Kirtac and Germano \(2024\)Kemal Kirtac and Guido Germano\. 2024\.Sentiment Trading with Large Language Models\.*Finance Research Letters*62 \(2024\)\.
- Li et al\.\(2025\)Gang Li, Dandan Qiao, and Mingxuan Zheng\. 2025\.Structured Event Representation and Stock Return Predictability\.*arXiv preprint arXiv:2512\.19484*\(2025\)\.
- Li et al\.\(2026\)Xiang Li, Zikai Wei, Yiyan Qi, Wanyun Zhou, Xiang Liu, Penglei Sun, Jian Guo, Yongqi Zhang, and Xiaowen Chu\. 2026\.Janus\-Q: End\-to\-End Event\-Driven Trading via Hierarchical\-Gated Reward Modeling\.*arXiv preprint arXiv:2602\.19919*\(2026\)\.
- Lopez\-Lira and Tang \(2024\)Alejandro Lopez\-Lira and Yuehua Tang\. 2024\.Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models\.*Journal of Financial Economics*\(2024\)\.
- Pangakis and Wolken \(2024\)Nicholas Pangakis and Sam Wolken\. 2024\.Knowledge Distillation in Automated Annotation: Supervised Text Classification with LLM\-Generated Training Labels\. In*Proceedings of the Sixth Workshop on Natural Language Processing and Computational Social Science \(NLP\+CSS\)*\.
- Papasotiriou et al\.\(2024\)Kassiani Papasotiriou, Srijan Sood, Shayleen Reynolds, and Tucker Balch\. 2024\.AI in Investment Analysis: LLMs for Equity Stock Ratings\. In*Proceedings of the 5th ACM International Conference on AI in Finance \(ICAIF\)*\.
- Wang et al\.\(2025\)He Wang, Wenyilin Xiao, Songqiao Han, and Hailiang Huang\. 2025\.StockMem: An Event\-Reflection Memory Framework for Stock Forecasting\.*arXiv preprint arXiv:2512\.02720*\(2025\)\.
- Xia et al\.\(2025\)Mingxuan Xia, Haobo Wang, Yixuan Li, Zewei Yu, Jindong Wang, Junbo Zhao, and Runze Wu\. 2025\.Prompt Candidates, then Distill: A Teacher\-Student Framework for LLM\-driven Data Annotation\. In*Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics*\.