UQO Working Paper No. 15· v2.0· November 2026· 77 pages

Faster, Wider, More Fragile? The Evolution of Intraday Shock Propagation in U.S. Equity Markets, 1992–2026

Finance & Economics

Isolated six-sigma moves are not followed by market-wide tail activity, and 98.1 of the +105.0 ×10⁻⁴ neighbour contrast is already there five minutes before the event.

Across 693,250 isolated one-minute events (effectively 1997–2026) matched to within-stock placebos, the difference-in-differences in market-wide breadth is zero or negative from 1 to 120 minutes. The pre-differenced response of the twenty closest correlation neighbours is +6.8 (2.4) ×10⁻⁴ at five minutes and negative at fifteen and thirty. The estimates are predictive associations and exclude by construction any co-movement completed within the event minute.

Simon-Pierre Boucher — contact@spboucher.ai

Paper PDF (revised)Read online Code & data repository

−0.65 (1.31) ×10⁻⁴

difference-in-differences in market-wide breadth five minutes after an isolated six-sigma move (ticker-cluster standard error), against a base rate of about 0.10; −4.78 (1.57) at thirty minutes and negative at every horizon from 1 to 120 minutes

≈ 3%

upper bound on the market-wide response as a share of the base rate (31 ×10⁻⁴, pre-level-stratified estimator, biased upward by regression to the mean); no more than 4.9% in any period

98.1 of 105.0

part of the five-minute neighbour contrast (×10⁻⁴, twenty closest lagged-correlation neighbours versus other receivers, net of placebos) that is already present five minutes before the event

+6.8 (2.4) ×10⁻⁴

pre-differenced neighbour response at five minutes (Romano–Wolf p = 0.058 over the period × horizon family); 162.1 at the event minute, 29.6 and 11.9 at one and two minutes, −13.7 and −21.7 at fifteen and thirty; 1.6 (2.4) once receivers already active at the event minute are excluded

−14.7 (12.5) ×10⁻⁴

late-minus-early contrast of the pre-differenced neighbour response: no rise over time, although the undifferenced contrast goes from 22.6 in 1997–1999 to 194.9 in the 2020s while the month-to-month overlap of neighbour sets goes from 0.04 to 0.52

3.6–10.9×

frequency of severe co-jump minutes (more than a tenth of observable names beyond four robust standard deviations) in 2008 and 2020 relative to 2013–2019, depending on the volatility scale; outside crises the series declines after 2002 (−1.37 per 1,000 per year, SE 0.42)

0.004 → 0.30

mean pairwise one-minute correlation among the hundred largest stocks in 1996 and in 2015 (0.16 and 0.39 at thirty minutes); based on four Epps cross-sections without confidence bands

Abstract

Using 7.69 billion one-minute bars on the full listed and delisted U.S. cross-section, 1992–2026, we test whether isolated six-sigma moves in single stocks are followed by abnormal activity elsewhere, comparing 0.69 million events (effectively 1997–2026) with matched within-stock placebos. Market-wide tail activity does not rise: the difference-in-differences is zero or negative at 1 to 120 minutes and a stratified upper bound is near 3% of the base rate. The shocked stock’s twenty closest correlation neighbors are more active than other names after the event (+105.0 × 10⁻⁴), but 98.1 of that exists five minutes earlier; the pre-differenced response is +6.8 (2.4) at five minutes and negative at fifteen: joint activity, not propagation. Its post-2010 step tracks neighbor-identification precision. Severe co-jump minutes are 3.6–10.9 times more frequent in 2008 and 2020 than in 2013–2019, depending on the volatility scale, and decline in calm years after 2002.

JEL: G10, G12, G14

Key Results

A uniform minute-level measure over thirty-four years

7,686,275,524 one-minute bars on 81,306 tickers forming 221,957 continuous spells, delisted securities included, from January 1992 through March 2026, audited and screened with strictly lagged universes, loadings and scales.

Isolated extremes are not followed by market-wide tail activity

The market-wide response is bracketed between a zero-or-negative difference-in-differences and a stratified upper bound of about 3% of the base rate; the bracket holds in each of the four periods, for positive and negative shocks, and in a known-null calibration.

The neighbour contrast is joint activity, not a propagation channel

A pre-event benchmark shows that the excess activity of correlated names exists before the event; what follows is a burst at the event minute and the minute after, +6.8 ×10⁻⁴ at five minutes, and nothing that lasts. A permuted network of random names from the same dollar-volume decile produces neither a pre-event gap nor a response.

The post-2010 step is read with identification precision

The undifferenced neighbour contrast steps up between the 2000s and the 2010s, but the pre-differenced response does not; the step coincides with a sharp rise in how precisely one-minute correlations identify neighbours (overlap 0.04 in 1997–1999, 0.32 in the 2000s, 0.52 in the 2010s).

Fragility only in a bounded, scale-dependent sense

One-minute returns above 1% fell from about 518 to 4.8 per 10,000 between the 1990s and the 2010s, severe co-jump minutes spike in 2008 and 2020 by a factor reported as a range, and their frequency declines in calm years, which qualifies rather than confirms the rising-synchronization reading of shorter samples.

Decision rules and diagnostics that qualify the numbers

Pre-event levels of every outcome in both arms, a known-null calibration of the bracket, inference under five clustering units with Romano–Wolf family-wise adjustment, and decision rules fixed before the corrected sample was estimated.

Quality control

This paper went through a multi-stage quality control: 6 quality-control reports and 84 points checked. The version presented on this page is the revised manuscript. Scores are on a 0–100 scale and are a diagnostic of the manuscript, not an acceptance probability.

Score after revision

65.8/100

56.8 before quality control · +9.0

Quality control dated November 4, 2026

Quality-control decision: major revision
Scientific integrity: PASS
Reproducibility: R1reproduced from documented analysis data
Maturity: Bone revision cycle from submission
Quality-control reports: 6
Points checked: 84

Scores by dimension

Before quality controlAfter revision

Originality & contribution62.2 → 60.0
Scientific rigor55.0 → 67.0
Identification & methodology53.0 → 62.0
Data & reproducibility50.4 → 68.0
Robustness & validation55.8 → 66.0
Writing70.5 → 70.0
Figures & tables59.0 → 76.0
Literature positioning63.2 → 78.0

Revised manuscript77 pages16 figures27 tables12 equations140 references32,641 words

What quality control changed

Quality control found two defects in the event-detection stage and a missing pre-event benchmark for the network statistic; the sample was re-drawn and the findings restated. The market-wide null survived and sharpened, but the claim of a concentrated transmission channel that tripled over time was withdrawn: the neighbour contrast is joint activity already present before the event.

Event sample: 1.85 million isolated six-sigma events → 693,250 events (680,176 placebos, match rate 98.1%) after correcting the simultaneity cap and the staleness screen and adding a raw-return floor; the propagation sample is now effectively 1997–2026.
Network finding: a +64.0 ×10⁻⁴ five-minute neighbour response read as transmission → a +105.0 ×10⁻⁴ contrast of which 98.1 exists five minutes earlier; pre-differenced response +6.8 (2.4) at five minutes, negative at fifteen and thirty.
Time trend: network channel ‘roughly tripled’ between the 1990s and the 2020s → withdrawn for the response (late-minus-early −14.7, SE 12.5); kept only as a step in the pre-event joint activity of correlated names that coincides with better neighbour identification (overlap 0.04 → 0.52).
Market-wide spillover: ‘between zero and about 2% of the base rate’ → difference-in-differences zero or negative at 1 to 120 minutes, stratified upper bound about 3% of the base rate and at most 4.9% in any period.
Severe co-jump minutes: ‘roughly ten times more frequent’ in 2008 and 2020 than in calm years → a range of 3.6–10.9 times relative to 2013–2019 depending on the volatility scale; the apparent 1997/98 break disappears under the observable-name denominator (4.3 per 1,000 in 1997, 3.2 in 1998).
Added: a limitations section with eleven items, a data and code availability statement, decision rules fixed before re-estimation, and family-wise adjusted inference under five clustering units.

Limitations stated in the paper

Co-movement completed within the event minute is excluded by construction, so ‘wider: no’ is a statement about the minutes after an isolated event; ordering moves within the minute would need transaction-level timestamps.
The propagation sample is 1997–2026, neighbour identification is imprecise before 2005, and the collapse of the simultaneity-filter pass rate in 2001–2002 is unexplained.
Bars carry no quotes, depth, order flow or security type and the venue scope is undocumented: fundamental and trading-channel readings cannot be separated, residual erroneous prints survive the screens, and no reconciliation against CRSP or a consolidated-tape source was executed.
The ‘faster’ conclusion rests on four Epps cross-sections without confidence bands, the co-jump factor depends on the volatility scale and the calm benchmark, and the theory tests are weak in crises (minimum detectable market-wide response of 2.3% of the base rate in 2008–2009 and 2020).

Read online (interactive reader)Revised PDF Revised branch on spbgit

Data

Minute-level OHLCV panel

Public Hugging Face dataset mito0o852/OHLCV-1m (upstream source Finnhub.io according to the dataset card), pinned at a fixed revision: 411 monthly Parquet files, 87.7 GB compressed, January 1992 to March 2026, 7.69 billion bars, 81,306 tickers. Bars exist only for minutes with trades; securities are identified by ticker alone and the venue scope is undocumented. A split screen detects 6,944 corporate-action events.

Shock-event sample

693,250 isolated one-minute events in the lagged top-500 dollar-volume universe: at least six robust standard deviations of the stock's residual return and at least two ticks or 25 basis points, with a calm market factor and fewer than 2% of observable names moving simultaneously. 680,176 matched within-stock placebos (match rate 98.1%); 3,592 source tickers. The propagation sample is effectively 1997–2026.

Co-jump and Epps series

Yearly counts of severe co-jump minutes under an observable-name denominator and two volatility scales (89.5 per 1,000 in 2008 and 83.1 in 2020 against a mean of 8.23 over 2013–2019 under the prior-month scale; 4.3 in 1997), and Epps curves for four cross-sections of the hundred largest stocks.

Methodology

Shock detection with strict lag discipline

Returns are residualized against a leave-one-out market factor with strictly lagged loadings and scales; events must pass a raw-return floor, a staleness screen and a simultaneity cap computed over names with an observable return.

Matched placebos and difference-in-differences

Each event is compared with a placebo in the same stock at the same minute of day on a calm day of the same month; outcomes are differenced post minus pre and real minus placebo, and complemented by a pre-level-stratified estimator that bounds the response from above, calibrated on a known null.

Pre-differenced neighbour response

The contrast between the twenty closest lagged-correlation neighbours and other receivers is measured five minutes before the event, at the event minute and after it, by period and under alternative designs: receivers active at the event minute excluded, 30-minute and daily correlation neighbours, and a permuted network.

Co-jump minutes and Epps curves

Severe minutes are those in which more than a tenth of observable names exceed four robust standard deviations, counted under prior-month and within-month scales and benchmarked on rule-based calm windows; pairwise correlations are computed by sampling interval on previous-tick grids.

Inference

Bootstrap standard errors with 2,000 replications clustered by source ticker, repeated under month, day and paired event-day resampling and two-way clustering; Romano–Wolf adjusted p-values over families of cells; minimum detectable effects reported for the weak crisis tests.

Reproducibility

The repository contains the complete pipeline (download, audit, cleaning, universes, shock detection, placebo construction, network measurement, inference, tables, figures and macros; Python 3.12, dependencies locked).

Every table and figure is generated by a numbered script from machine-readable intermediate results, and every numerical claim in the text resolves to a generated macro.

A full rebuild from the raw snapshot takes about 2 hours, download included, on a 16-core Apple-silicon workstation; the propagation exhibits rebuild from the stored-estimate archive (1.4 GB, checksums committed) in about 45 minutes, and a reduced six-month end-to-end check runs in about five minutes.

A regression test regenerates the headline tables, the decision rules and the macro files and compares them with the committed versions.

The raw bars, the clean minute panel and derived minute-level series are not redistributed because the dataset card specifies no licence; the SHA-256 checksum of every raw file is committed. A public deposit of the stored-estimate archive with a persistent identifier remains to be made; until then it is available from the author on request.

Cite this paper

Boucher, S.-P. (2026). Faster, Wider, More Fragile? The Evolution of Intraday Shock Propagation in U.S. Equity Markets, 1992–2026. UQO Working Paper No. 15, Université du Québec en Outaouais.

@techreport{boucher2026wp15,
  author      = {Simon-Pierre Boucher},
  title       = {Faster, Wider, More Fragile? The Evolution of Intraday Shock Propagation in U.S. Equity Markets, 1992–2026},
  institution = {Université du Québec en Outaouais, Département des sciences administratives},
  number      = {15},
  type        = {UQO Working Paper},
  year        = {2026},
  month       = {nov},
  url         = {https://www.spboucher.ai/research/wp15},
}