\section{Data}
\label{sec:data}

% Every statement below is verified in docs/DATA_AUDIT.md (Stage A pilot
% audit and Stage C full-sample audit, 2026-07-13).

\subsection{Source and structure}

Our data are one-minute OHLCV bars for U.S.-listed securities originally
sourced from Finnhub.io (\url{https://finnhub.io}), a market data provider,
and aggregated from the provider's monthly archives into 411 monthly Parquet
files covering January 1992 through March 2026 (87.7~GB compressed) with no
missing months. We access the aggregation through a public archive whose exact
byte-level snapshot is pinned and checksum-verified in the replication
package, which fixes the data used throughout. The archive carries no explicit
licence; we use it for research with attribution and do not redistribute the
minute data, only the aggregates, event-level estimates, and code. Each row
records the open, high, low, and close price (U.S.\ dollars, unadjusted for
corporate actions), share volume, and ticker for one minute of trading in one
security. Timestamps are UTC and denote the start of the bar minute; we verify
the convention empirically by mapping bar times to the Eastern-time session in
months on both sides of daylight-saving transitions.

A bar exists only for minutes with at least one reported trade: there are no
zero-volume rows in the source. Missing minutes are therefore informative about
trading activity, and we treat them explicitly rather than filling a regular
grid. Two properties of the source cannot be established from the data and are
carried as limitations. The archive does not document whether bars aggregate
consolidated-tape trades or a single venue; the presence of sparse low-volume
prints after the 13:00 close on early-closing days suggests off-exchange
trades are included, but volumes could not be reconciled against a
consolidated benchmark for more than a few securities. And securities are
identified by ticker alone, without a permanent identifier or security-type
label, so that ETFs, ADRs, preferred shares, and warrants are present and can
be screened only heuristically.

\subsection{Full-sample audit}

The archive contains \DataRows{} bars. Observed trading dates match the
official XNYS calendar exactly---zero missing sessions and zero dates outside
the calendar over thirty-four years---and the regular-hours share of bars is
flat across calendar months, which rules out any daylight-saving mishandling
(an offset error would displace roughly 15\% of session rows in March and
November). Early-closing sessions are truncated at the true exchange close.
Integrity failures are negligible: \DataOHLCViol{} impossible OHLC bars
(concentrated in 1993--1995, flagged and excluded) and \DataDupKeys{}
duplicate ticker--minute keys (1.3 per million), which we merge under a logged
deterministic rule. No bar has a non-positive price or a zero or negative
volume. A pilot audit of three months (June 1992, June 2009, December 2025)
conducted before the full download, summarized in the Online Appendix,
established the same facts on 58.8~million bars and verified volume units
against public daily volumes for benchmark securities.

The panel spans \DataTickers{} distinct tickers forming \DataSpells{}
continuous spells, a spell being a run of trading days without a gap longer
than thirty calendar days; \DataMultiSpellTickers{} tickers have multiple
spells, so spells, not tickers, are the identity unit of every statistic
that follows a security through time. Entries and exits occur continuously,
with exits running between 800 and 1,600 per year: delisted securities are
present throughout, and only 22--26\% of the tickers active in June 1992
appear in 2009 or 2025. The data are therefore not a backfilled survivor
universe, although we cannot verify that every delisting is captured to its
final trading day. There is one discrete source-coverage expansion in 2020
(about 11,000 entries, largely over-the-counter names) that fixed-rule lagged
universes absorb (Figure~\ref{fig:coverage}). A conservative overnight-ratio
screen (canonical split ratios within $\pm1.5\%$, both-day dollar volume of
at least \$50,000, prior close of at least \$1) detects \NSplitEvents{}
corporate-action events (\NSplitFwd{} splits, \NSplitRev{} reverse splits);
event days $\pm1$ trading day are excluded from all return-based analyses.
Dividends and other distributions are not detected; their overnight effect
never enters an intraday return because returns are computed within the
trading day only. Table~\ref{tab:sample_construction} tracks the sample from
raw bars to the regular-hours analysis panel.

\input{tables/table1_sample_construction}

\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figE1_coverage.pdf}
\caption{\textbf{Coverage of the full panel.} (a) Distinct symbols per year
and the screened (lagged top-3000-eligible) universe; (b) regular-hours bars
per year; (c) median and 10th-percentile bars per stock-day on the lagged
top-1000 universe.}
\label{fig:coverage}
\end{figure}

\subsection{Universes}

Coverage widens from roughly 8,300 symbols per month in 1992 to about 23,100
in late 2025, and much of the recent expansion reflects over-the-counter
instruments rather than listed common equity. All analyses therefore operate
on universes defined by strictly lagged rules: a security is eligible in
month $m$ if in month $m{-}1$ it traded on at least ten days with a median
close of at least \$5 and a median of at least 100 regular-hours bars per
day, and the top-500, top-1000, and top-3000 universes rank eligible
securities by month-$(m{-}1)$ regular-hours dollar volume. Membership never
uses information from month $m$. A ticker-shape heuristic (fifth-letter
suffixes and warrant, unit, and preferred suffixes) flags likely non-common
instruments; it does not identify ETFs, which trade under ordinary tickers.

\subsection{Known limitations}

The archive contains no quotes, spreads, depth, or trade direction; all
statements about liquidity in this paper rest on dollar volume and the
high--low range. Security types are unlabeled, prices are unadjusted, venue
scope is undocumented, and the licence is unspecified. Pre-2001 minute
returns are discrete at eighths and sixteenths, and roughly half of the
one-minute returns of top-1000 securities in the mid-1990s are exactly zero
(\ZeroRetNinetySix\% in 1996 against \ZeroRetTwentyFour\% in 2024,
Online Appendix); any second-moment or tail statistic compared across the
decimalization boundary inherits this discreteness. We flag the tick-regime
caveat wherever it binds and anchor every secular comparison on a
within-era placebo contrast.