After the Headline

Disclosure. Quantisnow operates the news feed studied here and sells access to it. No outside party funded or reviewed this work, and it has not been peer reviewed. Nothing here is investment advice.

Key findings

1. News comes outside trading hours. 59% of wire items arrive before the open and 28% after the close. Only 13% arrive during the regular session, and 45% land exactly on the hour or half hour.
2. Most movement has no visible cause. Days with a single-company wire item are 10% of stock-days and 23% of idiosyncratic variance. Among large moves, 43% have no item of any kind in our feed.
3. Headline type predicts size better than direction. Earnings days average +0.08% but move a mean 4.8% in absolute terms. Financing headlines are the clear exception at -4.26%.
4. More headlines, larger moves. Sessions with five or more wire items move 5% or more 48% of the time, against 20% for one item.
5. In liquid stocks, the move is the event. In the most liquid third, no group of large moves has a matched 5- or 20-session return reliably different from zero. Later losses concentrate in the least liquid third.
6. Large gaps up fade, outside the most liquid names. After a news-linked opening gap above 20%, the stock loses a further 5.5% on average by the close.
7. Several classic results do not show up here. Section 7 lists them. One year is short, so this is weak evidence against those results, but it should temper any trading conclusion drawn from them.

1. The question

Prices adjust to news. That has been the working view since Ball and Brown (1968) and Fama et al. (1969). Two follow-up questions remain open in practice. How much of a stock's movement can be tied to anything published at all? And once a large move has happened, is there anything left?

The answers changed as markets changed. Press releases now reach machines first, news analytics are sold as a product (von Beschwitz et al. (2020)), and language models read headlines at low cost (Lopez-Lira and Tang (2026)). Published effects also tend to shrink once they are known (McLean and Pontiff (2016)). We therefore look at recent data: a full year of timestamped items on US stocks, matched with daily prices for 5,834 common stocks.

2. Prior research

How much is news? Cutler et al. (1989) found that macro news and major events explain a modest part of aggregate moves, and Roll (1988) found that much firm-level movement is unrelated to public news. Better text tools raised the estimate: Boudoukh et al. (2019) attribute 49.6% of overnight idiosyncratic volatility to identified news, against 12.4% during trading hours.

What follows news? Drift after earnings is long documented (Bernard and Thomas (1989)). Chan (2003) finds drift after headlines, strongest after bad news, and reversal after large moves without news, mainly in smaller and less liquid stocks. Savor (2012) reaches a similar split using analyst reports as the measure of information. Jiang et al. (2021) use intraday data to separate news-driven returns and report widespread underreaction, and Heston and Sinha (2017) show that news sentiment predicts returns at horizons that depend on how the news is aggregated.

Why might prices underreact or overreact? Hong and Stein (1999) model slow diffusion among investors who see only part of the picture, and Hong et al. (2000) find momentum stronger where coverage is thin. Attention matters: Barber and Odean (2008), Da et al. (2011), DellaVigna and Pollet (2009) and Hirshleifer et al. (2009) all connect limited attention with trading or price response. Media content and reach matter too (Tetlock (2007), Tetlock et al. (2008), Engelberg and Parsons (2011)), and Tetlock (2011) finds reversals after stale stories. Engelberg et al. (2012) find that short sellers gain much of their edge by processing public news.

Speed. Busse and Green (2002) show prices responding within seconds to televised analyst reports, and Rogers et al. (2017) document that small timing advantages in filing distribution can be worth trading on. Daily data cannot address this directly, and we do not try.

Analysts, insiders and momentum. Womack (1996) finds price effects from recommendation changes that continue afterward, especially for sells. Lakonishok and Lee (2001) find that insider purchases carry information while sales carry little, and Cohen et al. (2012) separate opportunistic from routine trades. Jegadeesh and Titman (1993) is the reference for price momentum, and trading costs can absorb much of such returns (Korajczyk and Sadka (2004), Novy-Marx and Velikov (2016)).

3. Data and sample

News. Items come from the Quantisnow feed, which ingests press-release wires, analyst actions, regulatory notices and SEC filings and tags them to tickers. Monthly summary statistics from the same feed are public (Quantisnow (2026a)), as is a monthly ranking of the largest reactions (Quantisnow (2026b)). We use the time an item first entered the feed, which can lag true publication. A wire item is a non-SEC item tagged to exactly one company. Multi-company items and SEC filings count as other items. We drop 213 shareholder law-firm advertisements, which report on a fall and are not news about the company.

Prices. Daily consolidated (SIP) bars from a commercial market data vendor, adjusted for splits and dividends, from July 2025 to November 2, 2026. SPY is the market benchmark. We keep common stocks by name and ETF screens: 8,759 listed securities pass, 6,324 have bars and 5,834 survive the filters below.

Filters. A stock-session needs a prior close of at least $1, trailing 20-day median dollar volume of at least $100,000 and at least 20 prior sessions of returns. Both filters use only information available before the session. Daily moves above 300% are treated as errors.

Two repairs. Reading the largest events by hand revealed two errors, and both are fixed in everything below. First, 202 firms have more than one ticker on file, so an item could sit on an old ticker while prices sat on a new one. Items now follow the firm by SEC CIK. Second, spinoffs, ADR ratio changes and acquisition closings look like huge moves. We blank 167 stock-sessions with moves of 30% or more on under twice normal volume, plus three we identified by hand.

4. Method

Sessions. An item before 16:00 ET on a trading day belongs to that session. Anything later belongs to the next one. Several items on one stock in one session form a single observation, because they share one return. Counting each headline would overstate the sample.

Abnormal returns. The reaction is the stock's close-to-close return minus SPY's, assuming a beta of one. An item that arrives during the session may follow the move instead of causing it, so we report timing separately.

Large moves. A shock is an absolute abnormal return of at least 5% and at least three times the stock's own trailing 60-session volatility. We set these thresholds before reading any result and report a 10% variant.

Later returns. SPY is a poor benchmark for later returns here. Over the sample SPY returned +13.3% and the median stock with at least 200 sessions of data returned -4.1%, with 34% of stocks beating SPY, so nearly every group looks negative against it. Our main measure subtracts the average return of non-shock stocks on the same date, in the same liquidity and volatility terciles. Table 8 keeps the SPY version for comparison, and the difference is large.

Inference. Means are winsorized at 0.5% and 99.5%. Standard errors cluster by date and stock (Thompson (2011)). We also run calendar-time portfolios, which remove overlap between events. We show well over 100 statistics, so some will look significant by chance. Following Harvey et al. (2016), we treat |t| of 3 as the bar for a new claim, and tables bold those cells. Groups under 50 events are not reported.

5. Results

5.1 When news arrives

The feed follows a schedule. Of wire items on weekdays, 59.1% arrive before 09:30 ET, 13.3% during the session and 27.6% after 16:00. The busiest minutes are 08:00, 09:00, 16:05, 08:30 ET, and 44.7% of items land exactly on a :00 or :30 mark. SEC filings differ: 64.6% arrive after the close. Friday wire volume is about half of Thursday's (18,786 against 39,511), a pattern worth reading next to DellaVigna and Pollet (2009), although we do not test it.

Most information therefore reaches the market while the regular session is closed, and the first regular-session trade comes after an overnight or pre-market reaction. Section 5.7 shows that this reaction is usually not the whole move.

5.2 How much movement is tied to news

The median absolute move is 1.98% on wire days and 1.48% on days with no item. Moves of 5% or more happen on 21.0% of wire days and 11.2% of no-item days. Wire days are 9.7% of stock-days and 23.0% of variance, rising to 13.2% and 27.2% in the most liquid third. The ratio is close to the earlier literature in spirit: news days matter out of proportion, yet most variance sits elsewhere (Roll (1988), Boudoukh et al. (2019)).

Among shocks, 44.0% have a wire item and 43.2% have no item of any kind. That does not mean nothing happened. Our feed misses things, and some moves come from sector news. Adjusting returns for the sector median barely changes this, with 41.1% of 16,491 sector-adjusted shocks showing no item. We also checked filing dates: 4.0% of no-item shocks sit within a session of a periodic SEC report, against 2.0% for all no-item days. So missed earnings reports are real but are not the main explanation.

5.3 Headline type

The average earnings reaction is near zero with a 29% chance of a 5% move, which is what a fair bet on an uncertain event should look like. Clinical news is similar and larger. Financing news is different: offerings average -4.26% with volume at 3.2 times normal, and listing or compliance notices average -3.68%. M&A averages only +0.85% probably because the group includes acquirers and generic uses of the word.

5.4 News bursts

More items in a session go with larger moves and more volume. The median absolute move rises from 1.94% with one item to 2.32% with two, 2.62% with three or four, and 4.69% with five or more, while median relative volume goes from 1.14x to 2.06x. Causality is mixed: a large move draws analyst notes, so part of a burst is the reaction to the event. Among shocks, a larger burst does not predict continuation. Means are small and insignificant through three or four items, and the five-plus group is too small and noisy to read (n=65, t=1.0).

5.5 After a large move

Liquid stocks. In the top tercile every cell has |t| below 2. After a shock in a liquid stock, we find no dependable drift and no dependable reversal over a month, whether or not a headline came with it.

Illiquid stocks. Returns are negative across the bottom tercile, and news-linked falls are the strongest case: -3.71% after 5 sessions and -5.17% after 20. That fits bad news traveling slowly (Hong et al. (2000), Chan (2003)), but the up-move cells are negative too, so these stocks may simply do poorly after any shock.

News against no news. The clean version of Chan (2003) and Savor (2012) predicts drift after news and reversal without it. We get a partial match. After up moves, no-item stocks trail news-linked stocks by 1.8% over 20 sessions (t=2.6), in line with those papers but below our bar of 3. After down moves, no-item stocks keep falling instead of reversing, and the gap with news-linked stocks is insignificant (t=0.4). Inside the top tercile, the 20-session difference is insignificant in both directions.

A portfolio check. A calendar-time portfolio goes long up-shocks and short down-shocks with a wire item for five sessions, starting at the next session. Across all stocks it earns 74% annualized before costs (t=2.1). In the most liquid third it earns -6% (t=-0.2). The return is all in the least liquid third (t=3.0), and it comes mostly from the short side, in stocks that are often hard or impossible to borrow. These figures ignore borrow costs, spreads and price impact, which is the problem Korajczyk and Sadka (2004) and Novy-Marx and Velikov (2016) describe. We do not read this as an available return.

5.6 Do moves start before the news?

In the five sessions before a shock, matched returns are small for news-linked moves (+0.35%, t=1.8 before up moves; -0.21%, t=-1.0 before down moves). They are larger for no-item moves (+1.94%, t=5.6; -1.56%, t=-4.6). News-linked moves tend to start without a run-up, while no-item moves tend to extend one already under way. We read this as a description of two kinds of large move, not as evidence about leaks.

5.7 Opening gaps

Gaps up fade, and the fade grows with the gap. Gaps above 20% lose 5.5% with a wire item and 6.4% without one, and only 38% and 31% keep going. Small gaps up with no item also fade, at -1.71% and with 34% continuing, which suggests that a headline helps a gap hold. News-linked gaps down above 20% partly bounce (+2.45%, t=4.2).

The pattern does not extend to liquid names. In the top liquidity tercile, for gaps of at least 5%, the four cells range from -0.44% to -0.17%, with |t| near 1. Opening prints in thin stocks can also be noisy, which may overstate the fade.

Across all news sessions, the median absolute gap is smaller than the median absolute open-to-close move (for pre-market items, 0.96% against 1.83%). A large share of ordinary movement still happens during the session, and the gap points the same way as the day's full move 67% of the time.

5.8 Momentum

The past matters more than the type of shock. Prior losers do poorly after any large move, and the sharpest case is a no-item jump in a recent loser: -6.11% over 20 sessions (t=-6.9). Moves up with no news in beaten-down stocks are the group we would look at first for attention-driven buying (Barber and Odean (2008), Da et al. (2011)), though we cannot observe who traded. Prior winners hit by news-linked up moves earn +0.97% (t=1.5), which is consistent with continuation but not reliable.

Plain 20-session momentum is weak over this window. Matched returns by quintile of prior return run -1.28%, +0.37%, +0.18%, +0.06%, +0.63% from losers to winners, using 12 non-overlapping formation dates, far too few to test Jegadeesh and Titman (1993), whose horizons are also longer. Chan (2003) finds that news during formation matters. Splitting by at least two wire items in the prior 20 sessions, we see no such difference among prior losers (-1.45% with news, -1.10% without).

5.9 Analysts and insiders

Analysts. Upgrades average +2.72% and downgrades -3.13% on the day. Some of this is reverse causation, since analysts often respond to news that also moved the price. Womack (1996) finds that sell recommendations keep drifting lower. We do not: downgrades are at +0.89% after 20 sessions (t=2.8). Our sample is short, and downgrades here are all rating changes, not only moves to sell.

Insiders. Open-market buys are followed by +1.25% over 20 sessions (t=3.4), and +1.69% for buys by a CEO, CFO or president (t=3.1). Buyers typically purchase after a decline: the median prior 20-session return is -4.1%. Sales outside a trading plan show nothing (-0.04%), which matches Lakonishok and Lee (2001). One result calls for caution: sales under a trading plan are followed by +1.68% (t=4.2). An informative sale should not precede a gain, so this probably reflects which companies have plan sales, such as large recent winners, which our controls do not absorb. The same selection could inflate the buy result. Unlike Cohen et al. (2012), we do not separate routine from opportunistic buyers.

5.10 Headline words and repeated stories

Words that describe a specific action carry direction: raises goes up 61% of the time and offering 39%. Promotional words carry much less. “Positive” has a mean of +1.58% but a median of only +0.04% and rises 51% of the time, so a few large winners drive the mean. Section 6 has a case where “positive” came with a 48% fall. Companies write their own headlines, which is one reason dictionary methods built on journalists' text (Tetlock et al. (2008)) may not carry over, and why language models that read in context may do better (Lopez-Lira and Tang (2026)).

We tried a headline version of the staleness test in Tetlock (2011), measuring word overlap with the previous ten headlines on the stock. Near-repeat headlines move stocks more at the median (2.16% against 1.78%), because routine results share a template, and the interaction that would show reversal is insignificant (t=-0.6). Tetlock uses full story text. Headlines are probably too coarse for this test, so we count it as uninformative.

6. Case studies

We chose these cases after looking at the data, because they show how the numbers above can go right or wrong. They do not establish anything statistically. Moves are market-adjusted, and later returns are net of matched controls.

Revolution Medicines: clear news, priced at the open

At 08:00 ET on April 13, 2026, the company reported Phase 3 overall survival results for daraxonrasib in metastatic pancreatic cancer (feed item). The stock gained +40.4%, with +37.5% in the gap and +1.8% from open to close. It gained a further +7.3% the next day. The company then sold stock and convertible notes, raising about $2 billion at $142 per share (BioSpace). After 20 sessions the matched return was +3.5%. Nearly all of the first day's move was in the gap, and net of controls little more had been added 20 sessions later.

Erasca: a positive headline and a 48% fall

After the close on April 27, 2026, Erasca announced “Positive Preliminary Phase 1 Dose Escalation Data”. The next session moved -47.8%. Trade press reported that an analyst called the efficacy strong and pointed to a disclosed patient death as a possible reason for the fall (Fierce Biotech). A reader or model using the headline alone would get the direction wrong, because the information was in the body of the release.

POET Technologies: a neutral headline hiding bad news

On April 27, 2026, POET published a “Purchase Order Update”. It disclosed that Marvell, which had acquired the customer, cancelled all related purchase orders (Yahoo Finance). The stock moved -47.5% on 19 times normal volume. Twenty sessions later the matched return was +57.3%. One case proves nothing about reversal, and Table 3 shows no general reversal in liquid stocks, but it shows how wide the range of outcomes can be around a small average.

Enphase: a burst that is mostly reaction

Enphase reported results at 16:05 ET on February 3, 2026 (feed item), above estimates (Investing.com). Four analyst items followed by 10:18 ET the next morning, for 5 wire items in one session and a move of +39.1%. The burst here describes one event and the reaction to it, not five events. Our rule also labels this session “analyst”, because a pre-market item outranks an after-hours one, which shows the limit of labeling a session by a single headline. The matched return over the next 20 sessions was -20.4%.

Moderna and Fiserv: when the data get it wrong

On August 19, 2026, Moderna moved +177%. The cause was Phase 3 cancer vaccine data released jointly with Merck (feed item; Forbes). That release names two companies, so our single-company rule excludes it and the only wire items left are analyst upgrades. So the largest move in our data among stocks trading at least $50 million a day is filed under analyst actions. The Table 2 analyst row should be read with this in mind.

On October 29, 2025, Fiserv fell 44% after cutting guidance (feed item; BizTimes). In our first pass, this appeared as a large move with no news, because the items sat on the old ticker FI and prices sat on FISV. This error is what led us to link firms by CIK. Other mismatches of this kind may remain, and each would sit in the no-item group.

7. Robustness and what did not replicate

Signs are stable across most specifications. Size matters most: with at least $5 million of daily dollar volume and a $5 price, no cell at 20 sessions has |t| above 1.3. Attrition is low, with 0.5% of stock-sessions missing a 20-session return, and 1.1% for no-item falls. Since missing stocks may be failures, later returns in that group are, if anything, biased upward.

What did not replicate here:

• Drift after news, in liquid stocks (Chan (2003), Jiang et al. (2021)). Those papers use longer samples and, in the second case, intraday identification, so we do not claim a contradiction.
• Reversal after no-news falls (Chan (2003), Savor (2012)). No-item stocks kept falling.
• Continued drift after downgrades (Womack (1996)).
• Reversal after stale stories (Tetlock (2011)), although our headline measure is a weak test.

Decay after publication (McLean and Pontiff (2016)) is one possible reason, along with faster news processing (von Beschwitz et al. (2020)). A short sample is another, and we cannot separate them.

8. Limits

• One year, one regime. The sample has 261 sessions. Standard errors do not capture the risk that this year is unusual.
• “No item” means none in our feed. The feed has gaps, and the Fiserv and Moderna cases show how they appear.
• Coverage changes over time. Company links to wire items were 673,113 in 2022 and 134,916 in 2026 through October, as sources changed. The counts should not be compared across years, which is one reason we use a single recent year.
• Timestamps show when we saw an item. They are not official publication times, and an item can follow a move that it appears to explain.
• Labels come from rules applied to headlines. We did not hand-check a sample to estimate accuracy.
• Daily data. We cannot measure reaction speed or separate the news moment within a day. Opening prices in thin stocks can be noisy.
• Benchmarks. We assume a beta of one on the reaction day and match on two characteristics afterward. Other factor exposures are not removed.
• Survivorship. The universe comes from our current company table, so some delisted stocks may be missing. 748 stocks in the panel stopped trading before the end of the sample.
• Costs. All returns are before spreads, impact, fees and borrow. We tested no strategy that accounts for them.
• Multiple tests. We ran many cuts, and thresholds were set before results, but the extra tests and repairs followed a first look at the data, as described above.

9. Data, code and how to cite

Aggregate tables and the event panel are free under CC BY 4.0. The panel has no prices and no headline text, so it needs to be joined to your own return data, such as CRSP. The code shows every definition, although it reads files from our database and from a licensed price vendor that we cannot redistribute.

• News event panel: 144,659 items: id, ticker, UTC time, reaction session, timing, catalyst. No prices or headline text. Join to your own returns.
• Arrival clock: Wire and SEC items by half hour, Eastern Time (Figure 1).
• Catalyst reactions: Reaction-day statistics by headline type (Table 2).
• Headline bursts: Move size and volume by items per session (Figure 4).
• Post-shock returns: Every cut and robustness variant at 1, 5, 10 and 20 sessions (Tables 3 and 8).
• Post-shock paths: Session-by-session paths with standard errors (Figure 5).
• Gap fade: Open-to-close behavior after opening gaps (Figure 6, Table 4).
• Momentum interaction: Prior 20-day return crossed with shock type (Table 5).
• Analyst and insider events: Reaction and later returns (Table 6).
• Headline words: Reaction by word in the headline (Table 7).
• load.py: Builds the stock-session panel. Thresholds are set at the top.
• analyze.py: Runs every test in the paper.

Loading the panel and reproducing the session assignment:

import pandas as pd

url = "https://www.quantisnow.com/research/after-the-headline/news_event_panel.csv.gz"
events = pd.read_csv(url, parse_dates=["published_utc"])

# One observation per stock per session: headlines on the same stock and
# day share one return, so counting each would overstate the sample.
sessions = (events.groupby(["ticker", "reaction_session"])
                  .agg(items=("newsitem_id", "size"),
                       first_seen=("published_utc", "min"))
                  .reset_index())

# Join your own daily returns, then define a shock as in Section 4:
# |abnormal return| >= 5% and >= 3 x trailing 60-session volatility.

Suggested citation. Quantisnow Research (2026). After the Headline: News, Bursts and Momentum in US Stocks. Quantisnow Research Whitepaper 2026-1. https://www.quantisnow.com/research/after-the-headline

@techreport{quantisnow2026headline,
  title       = {After the Headline: News, Bursts and Momentum in US Stocks},
  author      = {{Quantisnow Research}},
  institution = {Quantisnow},
  type        = {Whitepaper},
  number      = {2026-1},
  year        = {2026},
  month       = nov,
  url         = {https://www.quantisnow.com/research/after-the-headline}
}

Found an error, or want to replicate this with your own return data? Write to contact@quantisnow.com. We will publish corrections in later versions of this page and keep earlier numbers available on request.

For information only. This is not investment advice or an offer of any security. Past returns do not predict future returns.

References