\section{Data}
\label{sec:data}

\subsection{Source and connector}
All bars come from FinData~OS, a private lakehouse of FirstRate Data one-minute bars for U.S. stocks and ETFs
(6.70 and 0.91 billion rows respectively at audit time, 2000--2026). The connector discovers the catalog, the
published Arrow schema and the OpenAPI document, reads small verification pages through the paginated bars endpoint
and performs bulk retrieval through asynchronous Parquet exports pinned to a dataset version. Every call keeps the
request identifier, dataset identifier and version; every export file is schema-validated and hashed. The license
is private use: no raw bar, grid or per-origin return leaves the compute node, and the release contains code,
aggregate results and synthetic fixtures only.

\subsection{Conventions verified on real rows}
Timestamps are bar-\emph{open} instants in UTC: the bar stamped 09:30 ET covers 09:30:00--09:30:59 and is complete
at 09:31. A regular XNYS session is 09:30--16:00 ET, i.e.\ 391 bars on a normal day (slots $0,\dots,390$) and 211 on
a 13:00 early close; the last slot is the closing print, whose volume is one to two orders of magnitude larger than
a typical minute (6.7~million against a median of 0.12~million shares for AAPL on 2024-03-08). The closing print is
excluded from context windows and targets. Extended hours (04:00--20:00 ET) are present at the source and excluded
here. Zero-volume minutes are absent at the source and are never filled: missingness is a mask and a feature.
Daylight-saving transitions are handled through the exchange calendar (first regular bar at 14:30Z on 2024-03-08 and
13:30Z on 2024-03-11). Stored bars are raw; splits are applied explicitly per session from the instrument master,
which reproduces the server-side split-adjusted series exactly (AAPL 4:1 on 2020-08-31, ratio 4.000 on prices and
volume). Cash dividends are not applied: intraday log returns within a session are invariant to the dividend factor
and any dividend adjustment is window dependent; the overnight-gap feature across an ex-dividend date is left raw.
Mergers and spin-offs are unavailable; a delisted series simply ends.

\subsection{Point-in-time universe and coverage gate}
The selection date is 2017-12-29. Eligible instruments have a first bar at least six months before that date and a
last bar at least six months after the sample start (a documented survival requirement), an average daily dollar
volume of at least 5~million USD over December 2017 and at least fifteen daily bars in that window. From
\num{2147} candidates we draw 300 instruments proportionally over asset class (75\,\% stocks, 25\,\% ETFs),
dollar-volume tercile and daily-volatility tercile, with a round-robin over SIC divisions (stocks, from SEC
submissions) or ETF name categories inside each stratum. A first draw revealed regular-hours minute coverage
between 6\,\% and 100\,\% over 2018--2025 (median 82\,\%): a quarter of the instruments traded in fewer than 60\,\% of
regular minutes, which makes per-minute targets mostly masked. Before any model was tuned we audited the coverage of
all candidates and imposed an availability filter of at least 80\,\% coverage over the window (\num{1152} candidates
remain; Spearman correlation between dollar volume and coverage 0.76); the second draw keeps the same strata and
seed. Sixty instruments (20\,\%) are held out entirely for zero-shot evaluation. Only five delisted candidates passed
the filters because FinData's delisted one-minute archives are only partly ingested: the universe is therefore close
to a survivor sample, a limitation we carry through the paper.

\subsection{Sample, grid and targets}
The main sample spans 2018-01-02 to 2025-12-31 (\num{2011} sessions): \num{233958776} exported rows, of which
\num{220142835} fall in regular hours (93.7\,\% coverage on average, minimum 80.3\,\%; 13.8~million extended-hours
rows excluded; zero integrity rejections and duplicates; 242 vendor-flagged rows kept with their flags). Each
instrument is stored as a dense session~$\times$~391 grid of log open, high, low, close and log volume with a
presence mask, plus past-only 5-minute (78 per day) and 30-minute (13 per day) aggregates with their own masks and a
past-20-session seasonal profile of log volume and of squared returns. An origin $(s,m)$ is session $s$ at slot $m$,
with bars $0,\dots,m$ complete; it is eligible when $m\ge 30$, $m+60$ does not cross the last tradable slot (389, or
209 on an early close), bar $m$ is present and at least 80\,\% of the next 60 slots are present. Targets are
$y_h = \log C_{m+h}-\log C_{p(m+h-1)}$, where $p(\cdot)$ is the last present slot, with mask $\ind{\text{bar } m+h
\text{ present}}$, so that the masked sum over $h\le H$ is exactly the $H$-minute log return of the last observed
bar. Derived targets are the cumulative return $R_{t,H}$, the realised variance $\sum y_h^2$, high/low excursions,
and the tail event $\abs{R_{t,H}}>2.5\,\sigma_t\sqrt{H}$, where $\sigma_t$ is the trailing root-mean-square
one-minute return in the 390-bar window (floored at $10^{-5}$), computed from the past only.

\subsection{Features}
Every bar carries fifteen causal features: scaled return, range, upper and lower wicks, body, close location,
log volume, relative volume against the seasonal profile, missing flag, session-boundary flag, log temporal gap
(which encodes overnight and weekend gaps), minute fraction, time to close, weekday and early-close flag. Returns and
ranges are divided by $\sigma_t$ (times $\sqrt{5}$ and $\sqrt{30}$ for the aggregates), so inputs are scale free. Nine
global history-only statistics---$\log\sigma_t$, the scaled overnight gap, the calendar position, the ratios of
30-minute and intraday realised variance to $\sigma_t^2$, and the mean relative volume over 30 minutes---feed a
global token and the regime embedding. A leakage gate in the test suite perturbs every bar after the origin and
asserts that no input changes; a quarantined fixture that deliberately uses the next bar is rejected by the same test.

\subsection{Splits}
Sessions are split chronologically: training up to 2022-12-31, validation 2023, development test 2024 and an
untouched final test 2025 evaluated once, with a one-session embargo at the start of each evaluation block.
Held-out instruments are evaluated only with models trained on the pooled instruments. Training origins are two
random eligible slots per instrument-session (a 1.3-million-example epoch is not materialised: windows are indexed
into memory-mapped grids); evaluation origins are every ten minutes from 10:00, i.e.\ 36 per normal session.