UQO Working Papers · WP15 · v1.1
Faster, Wider, More Fragile? The Evolution of Intraday Shock Propagation in U.S. Equity Markets, 1992–2026
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
Modern markets absorb the idiosyncratic and synchronize on the systemic
Across 1.85 million isolated ±6σ minute events, market-wide spillover stays between zero and about 2% of its base rate; the twenty-neighbour network response rose from +28 to +90 ×10⁻⁴ between the 1990s and the 2020s.
Abstract
This paper studies how intraday shocks travel across U.S. equities using 7.69 billion one-minute bars covering the full listed and delisted cross-section from 1992 to 2026. Three findings emerge. First, isolated idiosyncratic shocks barely spread: across 1.85 million ±6σ one-minute events, the abnormal share of other securities with an extreme standardized return, measured against matched within-stock placebos, lies between zero and about 2% of its base rate at horizons of 1 to 120 minutes, in every decade.
Second, transmission is concentrated and has intensified: the shocked stock's twenty closest lagged-correlation neighbours respond (+64.0 ×10⁻⁴ in the probability of an extreme move at five minutes), and this network channel roughly tripled between the 1990s and the 2020s while market-wide spillover stayed bounded near zero — a rise that partly reflects sharper identification of the network as the Epps effect receded.
Third, common information is impounded far faster (one-minute pairwise correlation 0.004 in 1996 versus 0.30 in 2015), ordinary extreme minutes became an order of magnitude rarer after decimalization, and severe co-jump minutes — more than a tenth of the universe moving together — are roughly ten times more frequent in 2008 and 2020 than in calm years. Modern markets absorb the idiosyncratic, transmit it narrowly along statistical linkages, and synchronize on the systemic.
Key findings
7.69 bn
one-minute bars on 81,306 tickers, 1992–2026
1.85 M
isolated ±6σ minute events matched to within-stock placebos
≤ 2%
market-wide spillover as a share of its base rate
+64 ×10⁻⁴
twenty-neighbour extreme-move response at five minutes
+28 → +90
network channel (×10⁻⁴), 1990s to 2020s
0.004 → 0.30
one-minute pairwise correlation, 1996 versus 2015
Scientific contributions
- The longest minute-level panel of U.S. equities
- 7,686,275,524 bars, 81,306 tickers and 221,957 listing spells, including delisted securities, audited month by month and cleaned of splits, OHLC violations and duplicates before any variance-based measure.
- Idiosyncratic shocks barely spread
- The market-wide response to a six-sigma minute is bracketed between zero (difference-in-differences, −1 to −3 ×10⁻⁴) and about 2% of the base rate (+19 ×10⁻⁴ under a pre-level-stratified estimator), at 1–120 minutes and in every decade.
- Transmission became sharper, not wider
- The twenty closest lagged-correlation neighbours respond by +64.0 ×10⁻⁴ at five minutes (+51 in singleton minutes), and the channel rose from +28 in the 1990s to +90 ×10⁻⁴ in the 2020s; part of the rise reflects better identification of neighbours once the Epps effect receded.
- Common information is impounded within the minute
- One-minute pairwise correlation rose from 0.004 in 1996 to 0.30 in 2015 and 0.21 in 2024, whether because incorporation accelerated or because asynchronous-trading attenuation receded — most likely both.
- Fragility only in a bounded, post-decimalization sense
- Extreme minute returns above 1% fell from 518 to 4.8 per 10,000 between the 1990s and the 2010s, while severe co-jump minutes reach 91 and 84 per 1,000 in 2008 and 2020 against 9.5 in 2017–2019; the 1990s are not comparable because price discreteness manufactures synchronization.
- Diagnostics that qualify the headline numbers
- Real events arrive in minutes of elevated activity (pre-event breadth +40.3 ×10⁻⁴ above placebos), only 22% of events are alone in their minute, and ticker-cluster bootstrap standard errors (4.7 versus 1.8 ×10⁻⁴) are reported alongside the baseline inference.
Methodology
Methodology (4)
Shock detection and matched placebos
Standardized returns with intraday seasonal and robust scale factors; isolated ±6σ events compared with within-stock placebo minutes in a difference-in-differences design, complemented by a pre-level-stratified estimator that bounds the market-wide response from above.
Network response along lagged correlations
Each shocked stock's twenty closest lagged-correlation neighbours versus all other receivers, by decade; the network differential is decomposed into its real and placebo components to separate transmission from identification.
Epps curves and co-jump counts
Pairwise correlation as a function of sampling interval by year, and counts of securities with extreme standardized returns in the same minute, with a severity scale calibrated on calm, GFC and COVID periods.
Inference and robustness
Event-clustered, month-cluster and ticker-cluster bootstrap standard errors; eleven robustness designs (post-decimalization, mid-day, ex-announcement, calm versus crisis, |z| ≥ 8 and 12, ex-non-common, placebo network) plus design-diagnostic and network-component tables.
Research data
- Minute-level OHLCV panel
- 411 monthly Parquet files (87.7 GB) from the Hugging Face dataset mito0o852/OHLCV-1m, pinned at a fixed revision, January 1992 to March 2026: 7.69 billion bars, 81,306 tickers, 221,957 listing spells; delisted securities present (only 22–26% of 1992 tickers survive); exchange-calendar truncation, 6,944 split screens, OHLC and duplicate filters.
- Shock-event sample
- 1,854,384 isolated ±6σ one-minute events in the largest 500 securities, each matched to a within-stock placebo minute (1,766,058 placebos; match rate 95.2% overall, 71% in the 1990s), tracked at 1–120 minutes for breadth, intensity, signed response and network response.
Reproducibility
- Python package and 18 numbered scripts (download, audit, cleaning, shocks, networks, co-jumps, tables, figures), parameters in YAML, 7 pytest tests; download and cleaning run on the MacLustr cluster
- All 97 headline numbers generated as LaTeX macros from machine-readable estimate registries (scripts/15_build_results_macros.py)
- Figures as CVD-safe vector PDFs with source CSVs; UQO working-paper, IRFA journal and anonymous manuscripts plus an 8-page Internet Appendix built from the same sources
- Revision of 2026-10-17 documented in docs/CHANGELOG-REVISION.md with a frozen pre-revision PDF; 124 references verified by DOI on Crossref; analyses not executed listed explicitly
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.1current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 43 p. |