1\section{Research question and falsifiable hypotheses}2\label{sec:hypotheses}34The question is narrow on purpose. At a decision time $t$ inside the regular session, with every one-minute bar up to5$t$ complete, can a compact model that reads the recent minute-level history \emph{and} coarser past-only aggregates6produce better-calibrated and sharper predictive distributions for the next $H\in\{5,15,30,60\}$ minutes than models7that only read the minute stream, or than well-specified non-neural volatility models? ``Better'' is defined by8strictly proper scoring rules \citep{gneiting2007strictly, matheson1976scoring}, never by price-level errors: on a9one-minute grid the random walk is a formidable point forecaster, and a model can look impressive on price-level10mean squared error while learning nothing \citep{cont2001empirical}. We therefore train and score on return-based11targets and evaluate full predictive distributions.1213Let $r_{t+h}$ denote the log return of the bar opening $h$ minutes after $t$ and $R_{t,H}=\sum_{h\le H} r_{t+h}$14the cumulative return. The model emits, for each future minute, a Student-$t$ marginal with location, scale and15degrees of freedom, and (in a second stage) a dependence structure across the $H$ minutes. Five hypotheses were16written before any model was tuned and are tested as stated.1718\begin{hypothesis}[Multi-resolution history]\label{h1}19At equal parameter count and training budget, a model whose encoder reads the last 390 one-minute bars, the last five20days of 5-minute aggregates and the last twenty days of 30-minute aggregates achieves a lower $\CRPS$ (and $\WIS$) for21$R_{t,H}$ than the same architecture reading the one-minute bars only.22\end{hypothesis}2324\begin{hypothesis}[Transfer to unseen instruments]\label{h2}25A model trained on a pool of instruments, without any ticker identity, scores better on instruments it has never26seen than non-neural baselines fitted with the same information set, and its loss on unseen instruments is not27materially worse than on pooled instruments.28\end{hypothesis}2930\begin{hypothesis}[History-only regime conditioning]\label{h3}31Conditioning the decoder on history-only regime statistics (trailing volatility and relative volume) improves32calibration---measured by coverage gaps and the probability-integral-transform (PIT) dispersion---on days whose33cross-sectional volatility is in the top decile, relative to the same model without regime conditioning.34\end{hypothesis}3536\begin{hypothesis}[Coherent joint paths]\label{h4}37A dependence-aware decoder (a Gaussian copula with a low-rank factor structure over the $H$ future minutes, with the38Stage-A marginals held fixed) achieves lower energy and variogram scores for the joint path, and lower $\CRPS$ for39path functionals (maximum drawdown, realised variance), than sampling the same marginals independently.40\end{hypothesis}4142\begin{hypothesis}[Economic content, exploratory]\label{h5}43Signals derived from the predictive median contain incremental information after an assumed round-trip cost; this44hypothesis is exploratory because no quote data are available and costs are assumptions.45\end{hypothesis}4647Each hypothesis has a pre-specified test: a Diebold--Mariano statistic on the loss differential with an48overlap-aware heteroskedasticity- and autocorrelation-consistent variance \citep{diebold1995comparing,49harvey1997testing, newey1987simple}, a trading-day cluster bootstrap for interval estimates (all instruments of one50day form one cluster, because millions of overlapping windows are not independent observations), and Holm and51Benjamini--Hochberg adjustments across horizons and comparisons \citep{holm1979simple, benjamini1995controlling}.52Null findings are reported as such.53