# Time-HD High-Dimensional Time Series Forecasting Benchmark

Canonical source: https://huggingface.co/datasets/Time-HD-Anonymous/High_Dimensional_Time_Series

Official code: https://github.com/UnifiedTSAI/Time-HD-Lib

Introducing source: [U-Cast](../../wiki/sources/u-cast-2025.md)

## Dataset Type

High-dimensional multivariate time-series forecasting benchmark suite.

## Temporal Structure

Regularly sampled numeric multivariate time series with aligned channels inside each dataset. Frequencies range from 1 ms to 1 day, and the paper sets prediction length by frequency rather than using one fixed horizon everywhere.

## Actions Or Interventions

None. Time-HD is a passive forecasting benchmark. It does not include operator actions, control inputs, interventions, or counterfactual rollout targets.

## Reported Composition

The U-Cast paper reports 16 Time-HD datasets with 1,105 to 20,000 channels: Neurolib, Solar, Atec, Meter, Temp, Wind, Traffic-CA, Traffic-GLA, Traffic-GBA, Air Quality, SIRS, SP500, M5, Measles, Wiki-20k, and Mobility.

Domains include neural science, energy, cloud, weather, traffic, environment, epidemiology, finance, sales, web, and social behavior.

## Suitability Note

Time-HD is the strongest current anchor for HDTSF benchmark design in this repository. It is useful for testing channel count, cross-channel dependency, latent hierarchy, memory footprint, and scalability. It is not enough for action-conditioned observability or telecom world models because it does not join numeric observations with topology, event streams, or logged actions/interventions.

## Access And License Notes

The Hugging Face dataset page lists Apache-2.0 and exposes CSV-style high-dimensional subsets. The public dataset card is sparse, so provenance, terms, and limitations should be checked against the U-Cast paper and the original source datasets before operational use.