TeX 53.6%
Python 46.2%
1\section{Related literature}2\label{sec:literature}34% Citations verified at DOI level via OpenAlex/Crossref on 2026-07-13 and5% re-verified against Crossref on 2026-10-17 (docs/CHANGELOG-REVISION.md).67The question sits at the intersection of several literatures: theories of8cross-asset shock transmission; empirical work on fire sales, flows, and9ownership linkages; the measurement of comovement and liquidity commonality;10lead--lag effects and the Epps effect; network approaches to contagion; the11transformation of U.S.\ market structure and the high-frequency trading12debate; and the econometrics of jumps, co-jumps, and scheduled announcements.13We take each in turn and close with the paper's positioning.1415\subsection{Theories of cross-asset shock transmission}1617Theory identifies two broad families of channels through which a shock to one18security moves others. The first is informational. \citet{king1990transmission}19model agents who infer information from other markets' price changes, so that20an idiosyncratic ``mistake'' in one market transmits broadly, and21\citet{kodres2002rational} show that cross-market rebalancing of shared macro22exposures transmits idiosyncratic shocks even without common fundamentals.23The second family operates through balance sheets and funding.24\citet{kyle2001contagion} generate contagion as a wealth effect of convergence25traders; \citet{yuan2005asymmetric} shows that borrowing constraints make26contagion and asymmetric price movements likelier after losses; and27\citet{brunnermeier2009market} link market liquidity to funding liquidity in28margin spirals that explain sudden, correlated liquidity dry-ups. Most29directly related to our setting, \citet{cespa2014illiquidity} show that when30liquidity providers learn from other securities' prices, illiquidity becomes31contagious and the market admits liquidity-crash equilibria, a mechanism they32connect to May~6, 2010. These theories jointly predict that transmission33should be state-dependent, asymmetric, and increasing in the degree to which34pricing and liquidity provision are cross-security activities. Our measures of35propagation breadth, network concentration, sign asymmetry, and market-state36dependence are designed to test these comparative statics at the minute scale.3738\subsection{Fire sales, flows, and ownership networks}3940A complementary empirical literature documents non-fundamental transmission41through investor portfolios: forced sales by mutual funds facing outflows42generate price pressure and reversals \citep{coval2007asset}, predictable43flows generate predictable pressure \citep{lou2012flow}, and uninformative44imbalances in one stock spill over to correlated stocks through risk-averse45liquidity providers \citep{andrade2008trading}. Ownership structure predicts46fragility \citep{greenwood2011stock}, common active ownership forecasts47correlation beyond fundamentals \citep{anton2014connected}, and overlapping48portfolios aggregate into measurable systemic vulnerability49\citep{greenwood2015vulnerable,duarte2021fire}. This work establishes that50idiosyncratic demand shocks do propagate, but at monthly or quarterly51resolution. Whether the same propagation operates within the hour, and52whether its speed changed as intermediation became algorithmic, is the margin53our data address.5455\subsection{Comovement, commonality, and liquidity dynamics}5657Liquidity has a systematic component: spreads and depths co-move with market-58and industry-wide liquidity \citep{chordia2000commonality}, common factors in59order flow explain much of the commonality in returns60\citep{hasbrouck2001common}, and commonality is supplied in part by common61intermediaries \citep{coughenour2004common}. Liquidity falls after market62declines, especially when funding is tight \citep{hameed2010stock}, specialist63losses forecast spread widening \citep{comertonforde2010time}, and returns to64liquidity provision spike as liquidity ``evaporates'' in turmoil65\citep{nagel2012evaporating}; aggregate liquidity risk is priced66\citep{pastor2003liquidity,acharya2005asset} and commonality strengthens in67downturns across markets \citep{karolyi2012understanding}. Relative to this68literature we do not measure liquidity levels: bars contain no quotes, and the69only liquidity information we use is prior-month dollar volume as a ranking70device, in the spirit of \citet{amihud2002illiquidity}, and the high--low71range as a descriptive volatility proxy \citep{parkinson1980extreme,72garman1980estimation}. Such OHLC-based proxies have a long pedigree73\citep{roll1984simple,corwin2012simple,goyenko2009liquidity}, but they do not74identify the quote-level mechanisms that the commonality literature studies.7576\subsection{Lead--lag effects, asynchronous trading, and the Epps effect}7778Cross-autocorrelation in returns is among the oldest facts in empirical asset79pricing: large stocks lead small stocks \citep{lo1990contrarian}, predominantly80within industries \citep{hou2007industry}, along customer--supplier links81\citep{cohen2008economic,menzly2010market}, and, at tick resolution, from82liquid to illiquid assets with intraday-seasonal structure83\citep{huth2014high}; \citet{chinco2019sparse} show that one-minute-ahead84cross-stock predictability is sparse and short-lived. At the shortest85horizons measured comovement is attenuated mechanically. \citet{epps1979comovements}86first documented that correlations among stocks fall toward zero as the87sampling interval shrinks. Part of this ``Epps effect'' is a pure artifact of88nonsynchronous trading and price discreteness89\citep{scholes1977estimating,lo1990nonsynchronous,harris1991stock}, which90\citet{reno2003closer} shows can reproduce the decay by simulation and which91estimators built on actual transaction times remove92\citep{hayashi2005covariance,zhang2011estimating}; microstructure noise93further contaminates high-frequency second moments94\citep{hansen2006realized}. Part is genuine delay in the incorporation of95common information. \citet{toth2009epps} document that the Epps curve among96NYSE stocks flattened steadily between 1993 and 2003 and argue that denser97trading explains only part of the change. Our Epps evidence extends this98series to 2024 on previous-tick synchronized grids and must be read with99both mechanisms in mind; we return to the point in100Section~\ref{sec:results}. The same mechanics matter for our design, because101neighbor sets are identified from one-minute residual correlations and are102therefore identified less precisely in eras where the Epps effect was strong.103104\subsection{Networks, connectedness, and contagion measurement}105106Network theory shows when idiosyncratic shocks fail to diversify away107\citep{acemoglu2012network,elliott2014financial}; \citet{glasserman2016contagion}108survey the contagion-network literature; real-economy links propagate109firm-level shocks measurably \citep{barrot2016input} and network positions110carry asset-pricing implications \citep{herskovic2018networks}. On the111measurement side, connectedness is estimated from variance decompositions112\citep{diebold2012better,diebold2014network}, Granger-causality graphs113\citep{billio2012econometric}, sparse VARs \citep{barigozzi2019nets}, and114tail-risk networks \citep{hautsch2015financial}. The contagion-measurement115tradition warns that unconditional correlation comparisons are biased in116volatile regimes \citep{forbes2002no} and defines contagion as comovement117beyond fundamentals \citep{bekaert2014global}; classic volatility-spillover118evidence \citep{engle1990meteor,hamao1990correlations} anticipates our design119at coarser resolution and for a handful of markets. These networks are120estimated at daily or lower frequency on small panels. We estimate pre-shock121correlation neighborhoods on minute data for the five hundred most active122securities each month and use them as lagged conditioning information, with123the same neighborhood contrast computed on placebo events as the benchmark.124125\subsection{Market-structure evolution and the high-frequency trading debate}126127Our sample spans the modern transformation of U.S.\ equity trading.128Decimalization in 2001 cut spreads \citep{bessembinder2003trade}; liquidity129promotes efficiency \citep{chordia2008liquidity}; turnover exploded in the1302000s alongside algorithmic trading \citep{chordia2011recent}; NYSE automation131made prices more efficient while raising the cost of immediacy132\citep{hendershott2011automation}; and fragmentation did not measurably harm133quality \citep{ohara2011fragmentation}. The normal-times evidence on134algorithmic and high-frequency trading is broadly positive135\citep{hendershott2011algorithmic,brogaard2014high,menkveld2013high,136hasbrouck2013low,chaboud2014rise,boehmer2021algorithmic}, with137\citet{menkveld2016economics} surveying the channel-by-channel evidence, and138markets absorb even large predictable trades with limited lasting impact139\citep{bessembinder2016liquidity}. Theory tempers the optimism: speed140investment can be socially excessive \citep{biais2015equilibrium}, fast news141trading concentrates volume and short-run correlation142\citep{foucault2016news}, toxic arbitrage degrades liquidity143\citep{foucault2017toxic}, the continuous limit order book hard-wires a144latency arms race \citep{budish2015high,aquilina2022quantifying}, and speed145can crowd out information acquisition \citep{weller2018does,baldauf2020high},146with fast traders eventually trading with informed institutional orders147\citep{vankervel2019high}; \citet{ohara2015high} argues the machine age148changes what microstructure concepts mean. This literature motivates our149central hypothesis but measures liquidity levels, price discovery, or single150events. We contribute the missing time series: propagation breadth and151network concentration of extreme shocks measured uniformly across three152decades of structural change, with year-by-year estimates rather than153imposed trend breaks.154155\subsection{Fragility and flash events}156157Episodic evidence shows that modern liquidity can vanish abruptly. On May~6,1582010, high-frequency traders did not trigger the crash but amplified selling159once it began \citep{kirilenko2017flash}; cross-market arbitrage broke down160\citep{menkveld2019flash}; and fragmentation shaped which ETFs were hit161\citep{madhavan2012exchange}. Sub-second ``fractures'' in individual stocks162proliferated as machine ecologies moved beyond human response times163\citep{johnson2013abrupt}. More systematically, high-frequency traders supply164liquidity in single-stock extreme moves but their provision weakens when many165stocks move together \citep{brogaard2018high}---precisely the margin our166breadth measure tracks. Proposed early-warning metrics remain contested: VPIN167signaled deteriorating conditions before the Flash Crash168\citep{easley2012flow}, but its power appears mechanically tied to volatility169\citep{andersen2014vpin}. This debate cautions against fragility indicators170that repackage volatility; our propagation measures are therefore defined171relative to lagged conditional distributions and benchmarked against matched172placebo minutes.173174\subsection{ETFs, indexing, and passive intermediation}175176A structural change specific to the later sample is the growth of index-linked177intermediation. ETF ownership raises non-fundamental volatility of underlying178stocks through the arbitrage channel \citep{bendavid2018etfs} and contributes179to return comovement \citep{da2018exchange}; index-product arbitrage flipped180index return serial dependence from positive to negative across twenty markets181\citep{baltussen2019indexing}. Effects on efficiency are contested182\citep{israeli2017dark,glosten2021etf,coles2022index}. For us, the ETF layer183is a candidate mechanism for the sharpening of propagation along correlation184neighborhoods, since basket arbitrage converts idiosyncratic shocks into185correlated flows. We cannot test it directly: the data carry no security-type186labels, so ETFs are neither identified nor excluded, a limitation we state187rather than paper over.188189\subsection{Jumps, co-jumps, and scheduled announcements}190191Our shock measurement builds on high-frequency econometrics. Intraday192volatility periodicity must be filtered before anything else193\citep{andersen1997intraday}, and robust periodicity estimators are needed194because the periodicity estimate is itself contaminated by the jumps one195wishes to detect \citep{boudt2011robust}; our minute-of-day factor is a196cross-sectional median for that reason. Bipower variation separates jumps from197diffusion \citep{barndorffnielsen2004power}, formal jump tests exist at the198observation level \citep{lee2008jumps,aitsahalia2009testing}, and jump tails199carry extreme cross-sectional dependence \citep{bollerslev2013jump}. Two200cautions discipline our thresholds: at ultra-high frequency much of what201coarser methods call jumps is friction \citep{christensen2014fact} or a202``drift burst'' extending over several minutes \citep{christensen2022drift},203and multiple-testing corrections eliminate a large share of spurious204detections \citep{bajgrowicz2016jumps}. Our events are threshold exceedances205of standardized residuals, not test-based jump detections, and we use the206term ``extreme move'' rather than ``jump'' wherever the distinction matters.207208The co-jump literature is the nearest empirical neighbor to our propagation209object. Index-level jumps are systematic co-jumps linked to macro news while210idiosyncratic jumps diversify away \citep{bollerslev2008risk}; formal tests211for common jump arrivals exist \citep{jacod2009testing}; co-jumps among212large U.S.\ stocks are frequent and concentrated around announcements213\citep{gilder2014cojumps}; simultaneous multivariate jumps are associated with214liquidity dry-ups and priced \citep{caporin2017systemic}; jumps coincide with215liquidity-demand shocks and news \citep{boudt2014intraday}; and simultaneous216jumps across countries carry portfolio consequences \citep{das2004systemic}.217Closest to our co-jump series, \citet{bormetti2015modelling} model systemic218price co-jumps among many stocks with Hawkes factor processes and219\citet{calcagnile2018collective} document that the number of stocks jumping220in the same minute is far larger than independence or a single factor would221imply, with collective synchronization events increasing in frequency in the222high-frequency era. Self- and cross-excitation of jumps has a canonical223language in Hawkes processes \citep{hawkes1971spectra,bacry2015hawkes},224estimated on international indices by \citet{aitsahalia2015modeling} and225tested at high frequency by \citet{dungey2018testing}. Scheduled macro226announcements provide the identified common shocks against which227idiosyncratic propagation must be benchmarked: announcement surprises produce228jump-like responses \citep{andersen2003micro,andersen2007real}, generate229co-jumps \citep{lahaye2011jumps}, structure volume and volatility230\citep{bollerslev2018volume}, and structure expected returns around231announcement times \citep{savor2013how,lucca2015pre}. Our robustness battery232accordingly excludes the main announcement windows. Where the co-jump233literature tests for simultaneity, our propagation measures trace the234sequence---breadth at one through 120 minutes---and our co-jump series235traces the secular evolution of simultaneity itself.236237\subsection{Factor structure at the minute scale}238239Separating idiosyncratic from systematic movement requires a factor240structure: common factors exist in high-frequency returns and order flow241\citep{hasbrouck2001common}, high-frequency principal components are242consistent with low-frequency covariance structure243\citep{aitsahalia2019principal}, latent continuous and jump factors are244estimable \citep{pelger2020understanding}, and betas decompose into245continuous, jump, and overnight components with distinct pricing246\citep{bollerslev2016roughing,bollerslev2020realized}; jump-tail risk commands247a large, time-varying premium \citep{bollerslev2011tails}. We deliberately use248a single, strictly lagged market loading per stock-month rather than these249richer structures, because any loading estimated on the event month would250overfit the very minutes it is used to classify; the cost, a coarser251residual, is absorbed by the placebo contrast.252253\subsection{Positioning and contribution}254255Each literature contributes a piece: theory predicts state-dependent,256asymmetric transmission; the flows literature demonstrates non-fundamental257propagation at monthly resolution; the network literature measures258connectedness on small panels at daily frequency; the market-structure259literature documents the transformation of trading but measures liquidity260levels and single events; the co-jump literature tests simultaneity at high261frequency over short samples. To our knowledge---based on a documented search262dated 2026-07-13 and repeated on 2026-10-17---no prior study combines263minute-level resolution, the full U.S.\ equity cross-section including264delisted securities, and a 1992--2026 sample spanning floor trading,265decimalization, Reg~NMS fragmentation, and ETF-intermediated algorithmic266markets. The joint analysis of market-wide and network-concentrated267propagation, its secular evolution, and the evolution of synchronized extreme268minutes at that resolution and length---with placebo-timing, placebo-network,269announcement-exclusion, and tick-regime falsifications---is the contribution.270