1\section{Research design}2\label{sec:method}34\subsection{Clocks, cutoffs and the information set}56Figure~\ref{fig:timeline} summarises the event-time design. For firm $i$ and fiscal quarter $q$ we observe the period end7$E_{i,q}$, the earnings-release date $A_{i,q}$ (the reveal date, with its before-open/after-close timing when known), the810-Q/10-K filing date $F_{i,q}$ and the vendor availability instant $V_{i,q}=F_{i,q}+1$ day. The prediction cutoff is9$c=A_{i,q}-h$ trading days for $h\in\{40,20,10,5,1\}$, with $h=20$ as the preregistered primary horizon. A cutoff is10admissible only if $c<\min(A_{i,q},F_{i,q})$ and $c>A_{i,q-1}$, so the feature window never contains the release of11quarter $q$ and never straddles two reveals. The information set is12\begin{equation}13 \Info(c)\;=\;\set{\text{bars}_{i,t}:\ \mathrm{session}(t)\le c}\;\cup\;\set{\text{facts}_{i,j}:\ V_{i,j}\le c},14 \qquad j^*(c)=\max\set{j: V_{i,j}\le c},15 \label{eq:infoset}16\end{equation}17regular-session bars up to the close of $c$ and accounting facts available by $c$ only; the latest such quarter is18$j^*(c)=q-1$ in the retained sample (we require that the forecaster predicts the quarter immediately following the19latest available one).20Twelve leakage guards (future facts, fiscal-end joins, post-cutoff bars, test-period transforms, daylight-saving errors,21identifier drift, split artefacts, release-day bars, overlapping windows across splits, survivorship, ingestion dates,22randomised placebos) are implemented as unit tests in the replication package; the local as-of join was verified live23against the FinData API's \texttt{as\_of} semantics on random entities and instants.2425\begin{figure}[H]26\centering27\includegraphics[width=\textwidth]{figures/fig01_timeline.pdf}28\caption{Event-time design and the no-leakage rule. Cutoffs $c=A_q-h$ precede the earnings release $A_q$; the 20-session29feature window ends at $c$; structured facts of quarter $q$ become available at $V_q=F_q+1$ day. Schematic of the30protocol, not data.}31\label{fig:timeline}32\end{figure}3334A fixed-calendar design (month-end cutoffs, target = the next release within 65 trading days) is used for the pricing35tests and as a robustness check because it does not condition on knowledge of the release date.3637\subsection{Targets}3839The primary outcome is the scaled operating-income surprise $S_{i,q}$ of equation~\eqref{eq:surprise}, with the40seasonal expectation computed from quarters available at $c$ and scaled by the latest available total assets. The binary41adverse outcome is $D_{i,q}=\ind{S_{i,q}\le-0.005}$, i.e., a shortfall of at least half a percent of assets (roughly a42fifth of a typical quarterly return on assets). Secondary43outcomes are the net-income analogue, a year-on-year ROA decline of at least one point, a negative operating-cash-flow44innovation of at least one percent of assets, a liquidity/coverage deterioration (current ratio down 10\% or interest45coverage falling below one), a symmetric placebo $|S_{i,q}|\ge0.5\%$ that a pure volatility channel would predict, and an46exploratory EPS miss against the vendor's estimate (available for \epsEstimateShare\% of releases; vintage unknown).47The continuous surprise, winsorised at the training-period 1st/99th percentiles, is modelled by ridge regression. No48target uses prices.4950\subsection{Features by sampling frequency}5152All market features are derived from the same raw 1-minute bars of the regular session (09:30 to the closing print),53aggregated to 5-minute and 30-minute buckets and to daily bars. Each intraday \emph{family} is computed per session54on bars of the given length and averaged (or summed) over the 20-session window ending at $c$. With $r_{t}$ the55within-session log return of bar $t$ and $n$ the number of bars present, the variance families are56\begin{equation}57 \RV=\sum_{t} r_t^{2},\qquad58 \BPV=\frac{\pi}{2}\sum_{t}\abs{r_t}\,\abs{r_{t-1}},\qquad59 \mathrm{RJ}=\frac{\max(\RV-\BPV,0)}{\RV},\qquad60 \RS^{\pm}=\sum_{t} r_t^{2}\,\ind{r_t\gtrless 0},61 \label{eq:rv}62\end{equation}63together with the negative-semivariance share $\RS^-/\RV$, the signed-jump variation $(\RS^+-\RS^-)/\RV$, realized64skewness $\sqrt{n}\sum_t r_t^{3}/\RV^{3/2}$ and kurtosis $n\sum_t r_t^{4}/\RV^{2}$, the concentration $\max_t r_t^{2}/\RV$,65and the first-order autocorrelation of $r_t$. The activity families are the intraday drift $\log(C/O)$, the opening and66closing 30-minute volume shares, the volume Gini across bars, the down-bar volume share67$\sum_t v_t\ind{r_t<0}/\sum_t v_t$, and the share of expected bars present. The co-movement families are the realized68beta to SPY, $\hat\beta=\sum_t r_t r^{\mathrm{SPY}}_t/\sum_t (r^{\mathrm{SPY}}_t)^{2}$ with sums taken over the window,69the share of $\RV$ accruing on market-down bars, the variance ratio $\overline{\RV}/\operatorname{Var}(r^{\mathrm{daily}})$,70and the volatility-of-volatility (dispersion of daily $\RV$ within the window). The 30-minute, 5-minute and 1-minute sets therefore have the \emph{same} \kIntra{} columns.71Daily features (\kDaily{} columns) are 20- and 60-day returns, volatilities, Parkinson range, skewness, kurtosis,72overnight and intraday return components, volume ratios, Amihud illiquidity, zero-return share, dollar volume, maximum73drawdown and bar activity. Point-in-time accounting features (\kFund{} columns) are margins, ROA, accruals, cash,74leverage, current ratio, coverage, capex, R\&D, issuance, revenue growth, size, the last available surprise, the trailing75mean surprise, the trailing adverse rate and the age of the information. Bar statistics are descriptive; nothing is76interpreted as order flow, spread or informed trading.7778\subsection{Baseline ladder and equal budgets}7980\begin{table}[H]81\centering\footnotesize82\caption{Nested information sets}83\label{tab:infosets}84\input{tables/tab04_info_sets}85\end{table}8687Table~\ref{tab:infosets} nests the information sets. B0 is the training base rate (and a shrunk firm-history rate);88B1--B5 are $\ell_2$-penalised logistic regressions whose penalty is tuned on the validation period only; LB1--LB5 are89gradient-boosted trees \citep{ke2017lightgbm} with an identical random-search budget (24 trials, early stopping on the90validation period) for every information set, so that the frequency comparison is a comparison of information sets,91not of capacity. Winsorisation bounds, imputation medians and standardisation constants are estimated on training rows92only. A supplementary set (B5all) concatenates all three intraday resolutions. Neural and text models are not used.9394\subsection{Validation and inference}9596The chronological split trains on cutoffs in 2012--2018, tunes on 2019--2020 and evaluates on 2021--2026, with purging97of training outcomes revealed after the next split's start minus a one-quarter embargo \citep{lopezdeprado2018}.98Expanding-window annual refits from 2016 provide out-of-fold predictions for the panel regressions and the timing and99pricing exercises; a 20\% firm hold-out evaluates transfer to unseen firms. The primary metric is log loss; Brier score,100AUC, PR-AUC and calibration are reported for every model. Paired differences (B5$-$B2, B5$-$B4, B5$-$B3) are assessed on the per-observation loss differential101$d_{i,q}=\ell^{\mathrm{A}}_{i,q}-\ell^{\mathrm{B}}_{i,q}$ with a block bootstrap over calendar quarters of the reveal102date \citep{kunsch1989, politis1994} and with the Giacomini--White statistic on quarterly means,103\begin{equation}104 t_{\mathrm{GW}}\;=\;\frac{\bar d}{\sqrt{\hat\sigma^{2}_{\mathrm{NW}}/T}},\qquad105 \bar d=\frac{1}{T}\sum_{\tau=1}^{T}\bar d_{\tau},106 \label{eq:gw}107\end{equation}108where $\bar d_\tau$ is the mean differential in calendar quarter $\tau$ and $\hat\sigma^{2}_{\mathrm{NW}}$ its109Newey--West long-run variance \citep{diebold1995, giacomini2006}; Holm's correction is applied within each hypothesis110family \citep{holm1979}. Panel regressions111\begin{equation}112 S_{i,q}\;=\;\alpha_i+\delta_{s(i)\times t(q)}+\beta\,\LFDS_{i,c}+\bgamma\tr\bX_{i,c}+\varepsilon_{i,q}113 \label{eq:panel}114\end{equation}115use the out-of-fold B5 probability as the Latent Fundamental Deterioration Signal (LFDS), firm and sector-by-quarter116fixed effects, and two-way clustering by firm and calendar quarter \citep{petersen2009, cameron2011}. Mechanism tests117(H3) interact the \emph{incremental} signal $\LFDS^{\mathrm{B5}}-\LFDS^{\mathrm{B2}}$ with ex-ante118terciles of size, illiquidity and leverage formed within each calendar quarter.119