Time-HD High-Dimensional Time Series Forecasting Benchmark
Source
- Dataset metadata snapshot: source.md
- Metadata JSON: metadata.json
- Official Hugging Face: https://huggingface.co/datasets/Time-HD-Anonymous/High_Dimensional_Time_Series
- Official code: https://github.com/UnifiedTSAI/Time-HD-Lib
- Introducing paper: U-Cast
Core Claim
Time-HD is the benchmark-side contribution of U-Cast: a suite of high-dimensional multivariate time-series forecasting datasets intended to make channel count, cross-channel dependency, memory, and scalability visible in evaluation.
Dataset Notes
- The U-Cast paper reports 16 datasets with 1,105 to 20,000 channels.
- Domains include neural science, energy, cloud, weather, traffic, environment, epidemiology, finance, sales, web, and social behavior.
- The benchmark uses frequency-specific prediction lengths rather than one fixed horizon.
- Time-HD is a passive forecasting benchmark; it does not include actions, control inputs, interventions, or counterfactual rollout targets.
Why It Matters
Time-HD gives High-Dimensional Time Series Forecasting a concrete evaluation surface. It is stronger than ordinary low-channel forecasting benchmarks for testing whether a model can exploit non-redundant cross-channel information without being overwhelmed by channel-channel compute or global-trend collapse.
Limitations
- The official Hugging Face dataset card is sparse relative to the paper, so provenance and terms should be checked against the paper and source datasets before operational use.
- It is not an observability or telecom world-model benchmark because topology, event streams, and logged actions/interventions are not first-class fields.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Native multivariate encoding and high-channel scaling | partially closes | Time-HD collects 16 high-dimensional datasets with 1,105 to 20,000 channels. | It is a benchmark, not an encoding method, and lacks explicit channel topology. |
| Benchmark level | partially closes | The benchmark makes channel count, cross-channel dependency, memory, and scalability visible in forecasting evaluation. | It does not test actions, interventions, topology-aware reasoning, or latent-state maintenance. |
| Causal structure, counterfactuals, and control | insufficient evidence | The dataset notes identify it as passive forecasting. | Needs action, control input, intervention, or counterfactual rollout fields. |