\section{Related work}
\label{sec:literature}

\paragraph{Intraday volatility and probabilistic evaluation.}
The classical description of intraday returns combines a deterministic periodic component with persistent
conditional heteroskedasticity \citep{andersen1997intraday, bollerslev1986generalized, nelson1991conditional,
engle2001garch}. Realised measures built from intraday returns turned volatility into an (almost) observable
quantity and showed that simple heterogeneous autoregressions forecast it well \citep{andersen2003modeling,
barndorff2002econometric, corsi2009simple, bollerslev2016exploiting}; \citet{hansen2005forecast} found that little
beats a GARCH(1,1) once the comparison is fair, and \citet{patton2011volatility} showed that only a few loss
functions (QLIKE, MSE) rank variance forecasts consistently under noisy proxies. These results frame our baselines:
seasonal profiles, EWMA, HAR-RV, GARCH-$t$. For distributional forecasts we rely on the theory of proper scoring rules
\citep{matheson1976scoring, gneiting2007strictly}, the calibration--sharpness paradigm and PIT diagnostics
\citep{gneiting2007probabilistic}, the weighted interval score \citep{bracher2021evaluating}, closed-form CRPS for
the Student-$t$ family as implemented in \texttt{scoringRules} \citep{jordan2019evaluating, taillardat2016calibrated},
threshold-weighted variants \citep{gneiting2011comparing}, and multivariate energy and variogram scores
\citep{scheuerer2015variogram}. Inference on loss differentials follows \citet{diebold1995comparing},
\citet{harvey1997testing}, \citet{giacomini2006tests} and \citet{hansen2011model}.

\paragraph{Deep learning for financial microstructure and time series.}
\citet{sirignano2019universal} documented that a single neural model trained on pooled limit-order-book data transfers
across stocks---the closest antecedent of our ``universal'' question, although with order-book inputs that we do not
have. DeepLOB \citep{zhang2019deeplob} and deep order-flow imbalance models \citep{kolm2023deep} also use quote-level
data. With bars only, the relevant architectures are the generic sequence models: temporal convolutions
\citep{bai2018tcn}, DeepAR-style autoregressive likelihood models \citep{salinas2020deepar}, N-BEATS
\citep{oreshkin2020nbeats}, and Transformer variants \citep{vaswani2017attention, wen2023transformers}. Patch
tokenisation \citep{nie2023patchtst} and variable-as-token inversion \citep{liu2024itransformer} are the
Transformer designs we adapt; \citet{zeng2023transformers} is the reminder that linear baselines are competitive.
Pretrained foundation models for time series \citep{ansari2024chronos, das2024timesfm} promise zero-shot
forecasting but assume a single real-valued series without missingness masks or multi-resolution context; we discuss
in Section~\ref{sec:design} why they are not evaluated here. Diffusion-based multivariate forecasters
\citep{rasul2021autoregressive} are a possible Stage-B alternative that we do not pursue given their compute cost.
Surveys of financial machine learning \citep{gu2020empirical, kelly2023financial, lim2021time} stress the
evaluation pitfalls we try to avoid: overlapping samples, survivorship, look-ahead, and multiple testing
\citep{lopez2018advances, bailey2014deflated}.

\paragraph{Dependence across horizons.}
Coherent joint paths require a dependence model. Copulas separate marginals from dependence
\citep{sklar1959, patton2006modelling}; our Stage~B uses a Gaussian copula with a low-rank-plus-diagonal
correlation over the sixty future minutes, trained on the probability-integral transforms of the observed
targets with the marginals frozen.