\section{Conclusion}
\label{sec:conclusion}

At a 20-trading-day horizon, on a 2013--2026 panel of U.S. listed firms with a strictly point-in-time information set,
one-minute OHLCV features change the out-of-sample log loss of forecasts of an adverse operating-income surprise by
\dlogBFivevsBTwo{} [\lologBFivevsBTwo, \hilogBFivevsBTwo] (logistic) and \dligLBFivevsLBTwo{} [\loligLBFivevsLBTwo,
\hiligLBFivevsLBTwo] (gradient-boosted trees) relative to a daily-market-and-accounting benchmark whose own increment
over accounting data is \dligLBTwovsLBOne. The gain from one-minute sampling is zero within intervals that are narrower than the pre-registered
minimum detectable difference for the boosted family and comparable to it for the logistic family; it is zero at every horizon from forty days to one day, for every
accounting target, in every subgroup and refit year, and on unseen firms. The small incremental content that the
one-minute signal carries is inconsistent with the frictions mechanism (it lives in large, levered firms), is already
impounded in prices by the cutoff, and has no economic value after factor adjustment and costs. The bracketed fields
of the question we set out to answer are therefore filled with a measured null.

Two limits bound the claim. The accounting panel contains only firms listed at the end of the sample, so the result
speaks to survivors; a delisted-fundamentals backfill, which the vendor appears to support, is the first extension.
Point-in-time availability is the vendor's filing-date-plus-one-day policy on current (possibly restated) values; a
filing-level reconstruction from SEC data would allow an as-originally-reported specification. Neither limit affects
the cross-frequency comparison, which is the object of the paper, but both affect the level of predictability and the
base rate of severe deterioration.

Three directions follow. First, the daily increment is large and its composition (returns, volatilities, volume
ratios, illiquidity) deserves the same frequency-controlled scrutiny at the weekly and monthly scale: where between the
quarter and the day does market anticipation of fundamentals actually form? Second, quote data would allow the
mechanisms that bars cannot see, order imbalance and spreads, to be tested with the same point-in-time design; the
present result predicts that they too would be summarised by the daily close at a 20-day horizon, a prediction worth
falsifying. Third, the pipeline, the frozen design and the leakage battery are reusable for any forecasting question
that joins market and accounting clocks; we release them with the hashed manifests needed to reproduce every exhibit
from the pinned dataset versions.