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

\subsection{Source and license}

All market and accounting data come from FinData OS, a private lakehouse of FirstRate Data operated by the author and
accessed through its versioned API (\texttt{findata.spboucher.ai/api/v1}). The lake holds 6.7 billion raw 1-minute
bars for 16,195 U.S. stock symbols (2000--2026, including 8,069 delisted symbols), quarterly fundamentals for
\nEntities{} entities (67 metrics, filing dates 2012--2026) and 373,326 earnings-release events. Bars are timestamped
at the bar open in UTC and cover 04:00--20:00 Eastern; zero-volume minutes are absent at the source and are never
imputed. FirstRate Data is licensed for private research use; no raw record is redistributed, and the replication
package contains code, hashed extraction manifests (dataset identifiers and versions, row counts, SHA-256 digests) and
derived aggregates only. Public augmentations are the SEC's \texttt{company\_tickers.json} and \texttt{submissions}
files (CIK, SIC industry), and the Kenneth French daily factor library.

\subsection{Universe and survivorship}

The accounting panel contains the \nEntities{} entities for which the vendor supplies quarterly statements; it has
\emph{no delisted entity}, because the vendor ingestion covered tickers active on 1 November 2026. The price side is
survivorship-free (delisted names are present in the bar lake), but the firm-quarter sample is conditioned on survival
to the end of the sample. This is a structural caveat: firms that deteriorated and delisted are absent, so adverse
base rates are understated and the most vulnerable firms are under-represented. We state it in the abstract and
treat it as the first limitation. Table~\ref{tab:universe} reports the construction of the universe: \nEligible{}
firms with at least eight quarterly periods, at least 250 sessions of 1-minute bars after 2012 and a non-governmental
SIC division; \nWithCik{} are matched to a CIK. Identifiers are vendor tickers; the vendor merges ticker changes under
the latest ticker, and we audit reuse with the SEC's former-name history.

\subsection{Reveal dates and point-in-time discipline}

Earnings-release dates are matched to fiscal quarters as the first release in $[E_q, F_q+10\text{ days}]$;
\releaseMatchedShare\% of firm-quarters are matched with a filing lag of at most 30 days (median lag between release
and filing: \filingLagMedian{} day). Release timing is known (before-open or after-close) for about two thirds of events.
Structured facts carry \texttt{filing\_date}, \texttt{available\_at} ($=$ filing date $+1$ day, a documented policy
rather than a historical vintage) and a revision identifier; because every fact was first ingested in November 2026 and
carries revision~1, restatements that occurred before ingestion are not reconstructible. Inputs therefore use prior
quarters only, outcomes use the vendor's current value of quarter $q$, and the stringent as-originally-reported
specification is left for a future SEC-filing-level reconstruction (Section~\ref{sec:limitations}).

\subsection{Sample}

The quarter panel has \nFirmQuarters{} firm-quarters for \nFirmsPanel{} firms. After the admissibility and exclusion
cascade of Table~\ref{tab:cascade} (market window of at least 15 valid sessions, raw close of at least one dollar,
available fundamentals, prediction of the next quarter, defined expectation, regular fiscal calendar), the event-time
sample has \nSampleEvent{} observations across the five horizons (\nSampleFirms{} firms, \nSampleQuarters{} firm-quarters)
and the primary $h=20$ cohort has \nObsPrimary{} observations with cutoffs from \firstCutoff{} to \lastCutoff; the
fixed-calendar sample has \nSampleCalendar{} firm-months. The adverse outcome occurs in \baseRatePrimary\% of
primary-cohort firm-quarters. Table~\ref{tab:summary} reports summary statistics and Figure~\ref{fig:coverage} the
coverage by year together with the adverse rate.

\begin{table}[H]
\centering\footnotesize
\caption{Construction of the universe and of the primary sample}
\label{tab:universe}
\begin{minipage}[t]{0.42\textwidth}\centering
\input{tables/tab02_universe}
\end{minipage}\hfill
\begin{minipage}[t]{0.56\textwidth}\centering
\label{tab:cascade}
\input{tables/tab02b_sample_cascade}
\end{minipage}
\begin{minipage}{\textwidth}\vspace{0.3em}\footnotesize
\emph{Notes.} Left: firms. Right: event-time observations (five horizons pooled); steps are applied in the frozen order
of \texttt{docs/PIT\_PROTOCOL.md}.
\end{minipage}
\end{table}

\begin{table}[H]
\centering\footnotesize
\caption{Summary statistics, primary cohort ($h=20$)}
\label{tab:summary}
\adjustbox{max width=\textwidth}{\input{tables/tab03_summary}}
\begin{minipage}{\textwidth}\vspace{0.3em}\footnotesize
\emph{Notes.} Intraday families are daily values averaged over the 20-session window ending at the cutoff; realized
variances are sums of squared bar log-returns within the regular session. Raw, unwinsorised values.
\end{minipage}
\end{table}

\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{figures/fig02_coverage.pdf}
\caption{Sample coverage and adverse-outcome rate by year of cutoff ($h=20$ cohort).}
\label{fig:coverage}
\end{figure}