Celsia/HPEC2026

Are Multidimensional Models Worth Their Computational Cost in Demand Forecasting?

Cheng-Jui Fan · Nikolay Aristov · Elenna R. Dugundji
IEEE High Performance Extreme Computing Conference (HPEC), 2026 — Paper 372

Dataset: HS6 U.S. Trade Demand-Forecasting Lattice

Dataset for the IEEE HPEC 2026 paper "Are Multidimensional Models Worth Their Computational Cost in Demand Forecasting?" (Paper ID 372). Built from the U.S. Census Bureau's bulk port-level HS6 merchandise trade files (public domain), it provides a dense commodity × state × flow lattice for benchmarking flat vs. explicitly multidimensional sequence models (RNNs, Transformers, S4/S4ND, Mamba-2/3, factorized attention).

Code and paper overview · Models and checkpoints · Archived results · Reproduction guide

Checkpoints and saved results for the 477-run main matrix live in the model repository, Celsia/HPEC2026 (repo_type="model"). Use repo_type="dataset" for this repository.

Contents

raw/                                384 Census bulk ZIPs — port-level HS6 exports + imports, 2010–2025
raw/reference/                      Census codebooks and fixed-width record layouts
processed/census_lattice_9ch.npz    benchmark lattice the archived checkpoints were trained on (9 channels, ~231 MB)
processed/census_lattice_9ch.json   its sidecar metadata (REQUIRED — see below)
processed/legacy_precorrection/     the earlier 28,292-series build — do not train on it (see "Versions")
processed/census_lattice.npz        an even earlier 2-channel development build (~64 MB), with its sidecar

Processed lattice (processed/census_lattice_9ch.npz)

Download the exact training pair: NPZ lattice and required JSON sidecar. The release provenance records this pinned revision and both checksums.

• Grid: 1,343 HS6 commodities × 14 port-states × 2 flows (export/import), 192 monthly steps (2010-01 … 2025-12). Cells kept at ≥80% grid density → 30,087 active series (14,873 export, 15,214 import).
• Cohort selection uses training months only. A series is retained when its aggregate value is non-zero in ≥95% of the 144 training months (2010-01 … 2021-12); state and commodity ranking likewise sees only training months. Validation and test months (2022–2025) play no part in membership.
• Channels (9): aggregate + air/containerized/bulk/land value, and aggregate + air/containerized/bulk weight; all log1p, then per-series MinMax fitted on the training split only with the range floored at 1.0 (raw = expm1(norm * norm_range + norm_min)).
• Splits by target-month index: train [0, 144), validation [144, 168), test [168, 192).
• Sidecar JSON is load-bearing: the .npz carries no metadata. Target channels ([0, 5] = aggregate value, aggregate weight), axis labels, split boundaries and the cohort-selection contract live in the matching .json — always load it alongside. The benchmark loader refuses a lattice whose sidecar lacks the cohort_selection contract.
• Built by scripts/build_census_lattice.py from raw/ with the Schedule D reference in raw/reference/ (HS4 dev variant via --level 4).

Arrays in the .npz:

array	shape	dtype	meaning
panel_norm	(30087, 192, 9)	float32	normalized model input/target panel
panel_raw	(30087, 192, 9)	float32	raw values/weights before log1p/MinMax
aggregate_raw	(192, 2)	float32	national aggregate value/weight, summed over all 282,848 raw series before cohort selection
series_idx	(30087, 3)	int32	(commodity, state, flow) index of each series
norm_min, norm_max, norm_range	(30087, 9)	float32	per-series, per-channel normalization constants in log space
mask	(1343, 14, 2)	bool	which lattice cells hold a retained series

SHA-256 of the current pair:

ed768219506b3fbbda5aa1eb598affde8cf127f58982e295ad0ff3ebd13fba6e  processed/census_lattice_9ch.npz
baac89c01fe73c9cdb509e720711358210d7af7cd5fc7fb94d7875a7c9ca26d8  processed/census_lattice_9ch.json

Versions

build	series	commodities	export / import	location
current (corrected) — what the archived checkpoints were trained on	30,087	1,343	14,873 / 15,214	processed/census_lattice_9ch.*
pre-correction (legacy)	28,292	1,263	13,883 / 14,409	processed/legacy_precorrection/, and revision 47dd2a5f6be759562fa9c1e257bca50854e0fd6f

The pre-correction build selected the cohort with a leak: a series was kept if it was non-zero in ≥95% of all 192 months, so membership depended on trade activity during the 2022–2025 validation and test period. The corrected build ranks and filters on the 144 training months only. Applying the old rule to freshly downloaded archives reproduces the legacy cohort exactly (28,292 series, 1,263 commodities, 13,883 / 14,409), so the entire difference is the leakage fix and none of it is Census revising the source; the 14 states are identical and in the same order. Tensor preprocessing (log1p, training-only per-series MinMax floored at 1.0, the 7 raw mode fields summed over port and country into 9 channels) is unchanged. The legacy sidecar has no cohort_selection contract and lacks norm_range and aggregate_raw; the loader refuses it. It is kept only so the earlier artifact stays inspectable.

The current pair is the exact file instance the sweep trained on (copied from the training machine, not rebuilt); the archived GRU checkpoint reproduces its logged test MSE on it.

Raw files

230-character fixed-width records: HS6 (Schedule B) + country (Schedule C) + district/port (Schedule D → state) + YYYYMM + value/weight fields. Schedule B / HS codes were re-versioned in 2012, 2017, and 2022 — match each year to its vintage code list (layouts and vintage code tables in raw/reference/). Census revises and re-issues monthly archives, so a rebuild from newer downloads is not guaranteed bit-identical to this lattice.

License & provenance

Source data is produced by the U.S. Census Bureau and is in the public domain (17 U.S.C. § 105). This repackaging adds no restrictions.

Citation

@inproceedings{fan2026multidim,
  title     = {Are Multidimensional Models Worth Their Computational Cost in Demand Forecasting?},
  author    = {Fan, Cheng-Jui and Aristov, Nikolay and Dugundji, Elenna R.},
  booktitle = {IEEE High Performance Extreme Computing Conference (HPEC)},
  year      = {2026}
}