\section{Introduction}
\label{sec:introduction}

When one security suffers a violent price move within a trading minute, does
the shock stay local or does it spread? The answer shapes how fast prices
aggregate information, how much liquidity providers must charge for standing
in the market, how execution algorithms should schedule parent orders, and
whether an idiosyncratic event can grow into a market-wide dislocation.
Regulators confront the same question from the other side: May~6, 2010 showed
that modern equity markets can move from normal functioning to synchronized
collapse within minutes, faster than any human intervention
\citep{kirilenko2017flash,menkveld2019flash}. Yet the empirical record on
intraday shock transmission is built almost entirely from short samples,
small panels, or single episodes. Whether the propagation mechanism itself
changed over the three decades in which trading moved from exchange floors to
fragmented, algorithmic, ETF-intermediated electronic markets is, to our
knowledge, undocumented.

Theory predicts change in both directions. Faster, cheaper, more competitive
markets should absorb ordinary idiosyncratic shocks more efficiently:
liquidity provision reacts within seconds, arbitrage capital is abundant in
normal times, and information is incorporated nearly continuously
\citep{hendershott2011algorithmic,menkveld2013high,brogaard2014high}. But the
same institutions that supply this efficiency---high-speed market making with
tight inventory limits, cross-asset arbitrage, index and basket
intermediation---withdraw or transmit pressure precisely when many securities
move together \citep{brogaard2018high,cespa2014illiquidity,bendavid2018etfs}.
Modern markets may therefore absorb ordinary shocks more efficiently while
synchronizing more tightly on common ones. The title asks whether U.S.\
equity markets became faster, wider, and more fragile at the intraday horizon.
The answer this paper documents is: faster, yes; wider, no---transmission
became sharper rather than broader; more fragile, only in a specific and
bounded sense that we make precise.

Our laboratory is a panel of \DataRows{} one-minute bars covering
\DataTickers{} tickers, including delisted securities, from January 1992
through March 2026. The data are audited exhaustively before use: observed
trading dates match the official exchange calendar exactly over thirty-four
years, timestamp conventions are verified on both sides of every
daylight-saving transition, and integrity violations are negligible
(\DataOHLCViol{} impossible OHLC bars in \DataRowsBn{} billion). Because the
source provides bars only for minutes with trades, identifies securities by
ticker alone, and does not document its venue scope, the empirical design
works with liquidity-screened, lagged-rank universes, ticker spells rather
than tickers, and detected corporate-action windows; Section~\ref{sec:data}
documents every screen and every limitation we could not remove.

Three findings organize the paper. First, \emph{isolated idiosyncratic shocks
do not raise market-wide tail activity by more than a few percent of its base
rate}. We detect \NEventsRealMM{} million one-minute events in which a
top-500 stock moves by at least six robust standard deviations of its own
residual return while the market factor is calm and fewer than 2\% of other
names move simultaneously. Relative to matched within-stock placebos, the
difference-in-differences estimate of the share of other securities with a
standardized residual beyond two in absolute value is between $-1$ and
$-3\times10^{-4}$ at horizons of 1 to 120 minutes, against a base rate of
about \BaseBreadthShare{}, with day-block bootstrap standard errors near
$1\times10^{-4}$. We also show, however, that real events occur in minutes
of elevated ambient cross-sectional activity (pre-event breadth is
\PreBreadthDiff$\times10^{-4}$ higher than for placebos), so that the
difference-in-differences and a pre-level-stratified comparison bracket the
market-wide response between zero and roughly \StratShareOfBase\% of the base
rate. Both bounds are economically small, and the pattern holds in every
decade and for positive and negative shocks alike.

Second, \emph{what transmission exists is concentrated and, within the limits
of our measurement, has intensified}. The shocked stock's twenty closest
lagged-correlation neighbors are \NetDiffFive$\times10^{-4}$ more likely than
other receivers to show a tail exceedance five minutes later, net of the same
contrast for placebos---a response of about six percent of the neighbor base
rate, estimated with a standard error of \NetDiffFiveSE$\times10^{-4}$
(\NetDiffFiveTickerSE$\times10^{-4}$ with ticker clustering). The
differential is \NetDiffFiveNineties$\times10^{-4}$ in the 1990s,
\NetDiffFiveOughts{} in the 2000s, \NetDiffFiveTens{} in the 2010s, and
\NetDiffFiveTwenties{} in the 2020s. We present this increase with an
explicit caveat: neighbor sets are identified from one-minute residual
correlations in the preceding month, and the paper's own Epps evidence shows
that those correlations were close to zero in the 1990s, so part of the
secular rise reflects sharper identification of the network rather than
stronger transmission along it.

Third, \emph{common information is impounded across the cross-section far
faster than it was, and tail risk within the post-decimalization era is
concentrated in market-wide minutes}. In 1996 the mean pairwise correlation
of one-minute returns among the hundred largest stocks was
\EppsNinetySixOne{} against \EppsNinetySixThirty{} at thirty minutes---the
classical Epps effect---whereas in 2015 it was \EppsFifteenOne{} against
\EppsFifteenThirty{}. The frequency of one-minute returns exceeding 1\% fell
from roughly \ExtremeGtOneNineties{} per 10,000 in the 1990s to
\ExtremeGtOneTens{} per 10,000 in the 2010s; part of that decline is the
mechanical disappearance of eighth and sixteenth ticks, and the decline
measured from 2002 onward is roughly tenfold. Minutes in which more than 10\%
of the universe moves beyond four robust standard deviations
simultaneously---severe co-jump minutes---are roughly an order of magnitude
more frequent in 2008 and 2020 than in surrounding calm years, and the
frequency of $|r|>5\%$ minute returns rose from \ExtremeGtFivePreCovid{} per
10,000 in 2017--2019 to \ExtremeGtFiveCovid{} in 2020. We do not claim that
such synchronization is secularly more frequent than in the 1990s: under the
same standardization the wide-tick 1990s produce severe-minute counts that
exceed 2008, which we attribute to discreteness rather than fragility and
which precludes a cross-regime comparison.

Identification receives explicit attention throughout. The central threat is
that securities respond to common information rather than to each other. The
shock definition residualizes against a leave-one-out market factor with
strictly lagged loadings and scales; isolation filters require a calm market
factor and fewer than 2\% simultaneous extremes at the event minute; every
event is compared to a placebo in the same stock at the same minute of day on
a calm day; and the pre-event level of every outcome is reported rather than
assumed equal across arms. The minutes that the isolation filters remove are
not discarded but studied separately as the systemic co-jump sample. What the
design identifies is the predictive association between an isolated
idiosyncratic extreme and subsequent cross-sectional tail activity; it does
not separate fundamental information transmission from trading-channel
transmission, and we phrase every claim accordingly.

The paper contributes to four literatures. To the contagion literature, which
measures transmission at daily or lower frequency for small panels, we
contribute event-level evidence from the full U.S.\ cross-section at the
minute scale over thirty-four years. To the market-microstructure and
high-frequency-trading debate, which contrasts normal-times liquidity gains
with episodic fragility, we contribute a uniform time series of propagation
and synchronization measured across the entire structural transformation of
U.S.\ trading. To the co-jump literature, we contribute a long-sample account
of how the frequency and severity of synchronized extreme minutes evolved,
with the comparability limits of such a series made explicit. And to the
measurement literature, we document what OHLCV bars can and cannot identify
about transmission, with every limitation quantified.

Section~\ref{sec:literature} positions the paper. Section~\ref{sec:data}
describes and audits the data. Section~\ref{sec:measurement} defines shocks,
outcomes, and networks. Section~\ref{sec:design} presents the empirical
design and the inference. Section~\ref{sec:results} reports the descriptive
transformation of the market and the baseline propagation results, including
the design diagnostics. Section~\ref{sec:mechanisms} examines networks,
cumulative responses, and co-jumps. Section~\ref{sec:robustness} summarizes
robustness, and Section~\ref{sec:conclusion} concludes.