\section{Introduction}
\label{sec:intro}

Every quarter, the accounting system reveals with a lag what the firm has already lived through: revenues that
disappointed, margins that compressed, cash that did not come in. Market prices are widely believed to anticipate part
of this revelation, and a long literature shows that they do at daily and annual horizons \citep{ball1968, beaver1980,
kothari1992, collins1994}. A newer belief, common among practitioners and increasingly in the academic literature on
high-frequency data, is that the \emph{texture} of trading, the minute-by-minute composition of volume and returns,
contains information that the daily close does not: informed traders split orders and trade when liquidity is deep
\citep{kyle1985, easley1987}, jumps and signed variation mark the arrival of firm-specific news \citep{lee2012,
patton2015, bollerslev2020good}, and attention-constrained investors process public information gradually
\citep{hirshleifer2003, dellavigna2009}. If this belief is right, one-minute bars should forecast forthcoming accounting
deterioration better than daily bars, holding everything else fixed. This paper tests that proposition.

The test is designed around one distinction, anticipation versus revelation, and one object, the marginal information
value of sampling frequency. Anticipation, the fact that prices move before statements, is not at issue; it is
expected under semi-strong efficiency \citep{fama1970, grossman1980} and we reproduce it. What is at issue is whether
observing the market at a finer resolution adds forecasting content about the \emph{accounting} outcome, over and above
(i) point-in-time fundamentals, (ii) daily OHLCV, (iii) coarser aggregates of the same bars and (iv) size, industry and
calendar controls. To answer it we build a nested ladder of information sets with the same universe, the same calendar
coverage, the same cutoff and horizon, the same feature \emph{families} computed from 30-minute, 5-minute and 1-minute
bars, and the same tuning budget per set, so that any difference is attributable to resolution rather than to
dimensionality or capacity.

The data are the owner's FinData OS lakehouse of FirstRate Data: 6.7 billion raw 1-minute bars for 16,195 U.S. stock
symbols (2000--2026, delisted names included), quarterly fundamentals with filing dates for \nEntities{} entities, and
373,326 earnings-release events with their timing. Point-in-time discipline is strict: the forecaster at cutoff $c$ sees
bars up to the close of $c$ and only those accounting facts whose vendor availability instant (filing date plus one
day) precedes $c$; the reveal date is the earnings press release, not the filing; and twelve leakage guards are unit
tests in the replication package, including a live equivalence test of our as-of join against the API. Targets are
accounting surprises against a seasonal random walk with drift \citep{foster1977}, never prices.

Three findings emerge. First, the ladder works as the literature predicts: point-in-time fundamentals lower the
out-of-sample log loss from \llBZero{} to \llLBOne{} (AUC \aucLBOne) and daily market data to \llLBTwo{} (AUC
\aucLBTwo), a daily increment of \dligLBTwovsLBOne{} that is unambiguous. Second, one-minute data add nothing. On an
untouched 2021--2026 test period of \nlogBFivevsBTwo{} firm-quarters, the paired change in log loss from adding the
1-minute families is \dlogBFivevsBTwo{} [\lologBFivevsBTwo, \hilogBFivevsBTwo] with logistic regression and
\dligLBFivevsLBTwo{} [\loligLBFivevsLBTwo, \hiligLBFivevsLBTwo] with gradient-boosted trees; 30-minute and 5-minute
families behave identically; the tree intervals exclude any gain at all and the logistic intervals exclude gains larger
than a fifth of the daily increment. The pre-registered minimum detectable difference (\mde) is wider than the tree
intervals and comparable to the logistic ones, so this is a precise null rather than an underpowered one. The null survives five horizons from forty days to one day before the release, five
accounting targets, a symmetric volatility placebo, firms never seen in training, every annual refit, and every
pre-specified subgroup. Third, the little information that the incremental one-minute signal does carry sits in large
and levered firms rather than in small, illiquid or neglected ones, is already in prices twenty days before the release
(no announcement-window abnormal return), and has no economic value in factor-adjusted decile portfolios.

The contribution is therefore a measurement: on a historically correct universe with a strictly point-in-time
information set, the marginal value of one-minute sampling for forecasting accounting deterioration is zero within
tight bounds, while the value of daily sampling is large. The result parallels, for fundamentals, what \citet{liu2015}
established for variance measurement, that the finest grid is not the most informative, and it bounds the mechanisms of
the high-frequency literature: whatever informed trading, jump arrival or gradual processing is visible in one-minute
bars twenty days before a release is summarised by the daily close. We also deliver a reproducible point-in-time
pipeline, frozen hypotheses and decision rules, and an honest statement of two structural limits: the accounting panel
is survivor-conditioned (firms listed in November 2026) and the vendor's point-in-time stamps are a documented proxy.

The paper proceeds as follows. Section~\ref{sec:theory} sets out the framework and the pre-registered hypotheses,
Section~\ref{sec:literature} positions the question, Section~\ref{sec:data} describes the data, the universe and the
point-in-time protocol, Section~\ref{sec:method} the design, Section~\ref{sec:results} the forecasting results,
Section~\ref{sec:mechanisms} the heterogeneity and pricing tests, Section~\ref{sec:robustness} the negative controls,
and Section~\ref{sec:conclusion} concludes with limitations and the research the null result motivates.