\section{Robustness and negative controls}
\label{sec:robustness}

\paragraph{Shuffled labels.} Permuting the adverse-outcome label within calendar quarter of the cutoff (training and
test rows, three draws) removes all predictability: the test AUC of the 1-minute model falls to \placeboAUC{} and the
B5$-$B2 differential is \placeboMeanD{} on average (Table~\ref{tab:placebo}). The pipeline does not manufacture
differences between information sets from the features alone, and the residual differential on shuffled labels gives
the noise floor against which the primary estimates should be read: the primary B5$-$B2 deteriorations
(Table~\ref{tab:contrasts}) are of the same order of magnitude as this floor, which is one more reason to describe the
1-minute increment as zero rather than as a harm.

\begin{table}[H]
\centering\footnotesize
\caption{Negative control: labels shuffled within calendar quarter}
\label{tab:placebo}
\input{tables/tab08_placebo}
\end{table}

\paragraph{Symmetric target.} A volatility channel would make 1-minute features informative about the \emph{magnitude}
of the surprise even if not about its sign. The symmetric target $|S_q|\ge0.5\%$ in Table~\ref{tab:targets} shows the
same null, so the result is not an artefact of a signed target.

\paragraph{Window contamination and sparse bars.} Excluding the 2,377 test observations whose 20-session window contains
the previous release, or restricting to firm-days with at least 98\% of expected bars, leaves the differential unchanged
(Table~\ref{tab:subgroups}). Sessions with fewer than 100 regular-session bars are already excluded from the window,
and the share of expected bars present is itself a feature, so illiquid names do not generate mechanical signals.

\paragraph{Release timing and sector.} The differential is the same for after-close and before-open releases, and for
financial and non-financial firms (the latter forecast much better, AUC 0.86, because their outcomes are smoother).

\paragraph{Model family and tuning.} The null holds for penalised logistic regression and for gradient-boosted trees with
identical tuning budgets; the sign of the LightGBM differential flips between the tuned configuration (\dligLBFivevsLBTwo)
and the fixed configuration used in the horizon ladder (Table~\ref{tab:horizons}, $h=20$), which is what one expects
from a true zero.

\paragraph{Firm hold-out and time.} The increment is zero on firms never seen in training
(\dligfirmholdheldout, CI [\loligfirmholdheldout, \hiligfirmholdheldout]) and in every refit year
(Figure~\ref{fig:rolling}); the only year with a visibly positive differential (2023 for the boosted family) is a
deterioration, not a gain.

\paragraph{What we could not test.} Delisted firms are absent from the accounting panel, so the stress test that would
matter most, forecasting the deterioration of firms that subsequently failed, cannot be run on this vendor's data; the
delisted-fundamentals backfill is documented as the first item of future work. Original-as-reported accounting values
are not available, so the stringent as-filed specification is deferred to a SEC-filing-level reconstruction. Both
limitations affect the level of predictability, not the comparison across sampling frequencies, which is the object of
the paper.