1\section{Related literature}2\label{sec:literature}34\paragraph{Prices lead earnings.} Since \citet{ball1968}, accounting research has documented that stock prices5anticipate accounting income: annual returns forecast earnings changes \citep{beaver1980}, leading returns raise6earnings-response coefficients \citep{kothari1992}, and the low contemporaneous return--earnings association reflects7prices' timeliness rather than noise \citep{collins1994}. The post-earnings-announcement drift \citep{bernard1989,8bernard1990} and the accrual anomaly \citep{sloan1996, richardson2005} show that the incorporation of accounting news9is incomplete, while fundamental analysis \citep{ou1989, piotroski2000} extracts forecasts of future earnings from10statements alone. Recent work replaces hand-crafted ratios by machine learning on detailed statement items11\citep{chen2022, easton2024, cao2024}. All of these studies operate at daily or lower frequency and treat the market side12as a return; our contribution is to fix the accounting side and vary the \emph{resolution} of the market side.1314\paragraph{Information in intraday activity.} Market microstructure theory ties the composition of trading to private15information \citep{kyle1985, easley1987}, and realized-measure econometrics separates continuous variation, jumps and16their signs from high-frequency bars \citep{andersen1998, andersen2003, barndorff2004, barndorff2010semivariance,17patton2015, lee2008}. Signed jump variation and downside semivariance price the cross-section18\citep{bollerslev2020good, bollerslev2016beta}, jumps cluster around news \citep{lee2012, bajgrowicz2016}, and19announcement days change betas and the volume--volatility relation \citep{patton2012, bollerslev2018volume,20savor2016}. The question of the best sampling frequency has been settled for \emph{variance measurement} -- five-minute21realized variance is hard to beat \citep{liu2015, bollerslev2016errors, bollerslev2018risk} -- but not for the22\emph{forecasting of fundamentals}, which is our object. \citet{zhang2023volatility} show that pooling intraday23commonality improves machine-learning volatility forecasts; we include volatility families at every frequency and use a24symmetric placebo target to separate a volatility channel from a signed-information channel.2526\paragraph{Pre-announcement trading and price discovery.} Short sellers \citep{christophe2004}, individual investors27\citep{kaniel2012}, institutions \citep{hendershott2015} and option traders trade informatively before earnings28announcements; differences of opinion shape announcement returns \citep{berkman2009}; algorithmic trading may reduce29pre-announcement information acquisition \citep{weller2018}; and the announcement reaction itself occurs largely in30quotes rather than trades \citep{gregoire2022, brogaard2019}. These studies use trader-type or quote data that bars do31not contain. We ask a complementary question with anonymous OHLCV only: how much of the forthcoming accounting outcome32is visible in the shape of pre-release trading, and does that visibility depend on resolution?3334\paragraph{Attention and processing.} Limited attention \citep{hirshleifer2003, hirshleifer2009, dellavigna2009,35aboody2010} and media coverage \citep{fang2009} govern how fast public information is processed; the timing and36heterogeneity hypotheses H3--H4 are motivated by these channels and by distress-risk measurement \citep{campbell2008,37beneish1999}.3839\paragraph{Forecast evaluation and machine learning in finance.} We follow the evaluation standards of40\citet{diebold1995}, \citet{giacomini2006}, \citet{hansen2005} and \citet{hansen2011mcs}, with multiple-testing control41\citep{holm1979, benjamini1995}, cluster-robust inference \citep{petersen2009, cameron2011}, block bootstrap42\citep{kunsch1989, politis1994} and purged, embargoed time splits \citep{lopezdeprado2018}. Equal-budget comparisons of43information sets are inspired by the asset-pricing machine-learning literature \citep{gu2020, freyberger2020, kozak2020,44chen2024deep, avramov2023} and its warnings about false discoveries \citep{harvey2016, chordia2020}. Gradient-boosted45trees \citep{ke2017lightgbm} provide the nonlinear benchmark. The companion UQO working paper on intraday shock46propagation \citep{boucher2026wp15} uses the same vendor data for a different question; no result is reused here.47