Time-Series-Library Benchmark Collection
Source
- Dataset metadata snapshot: source.md
- Metadata JSON: metadata.json
- Official Hugging Face: https://huggingface.co/datasets/thuml/Time-Series-Library
- Official code: https://github.com/thuml/Time-Series-Library
- Paper: https://arxiv.org/abs/2407.13278
Core Claim
Time-Series-Library is a broad benchmark collection for deep time-series analysis. In this repository it is the durable handle for the LSF/LTSF-style long-term forecasting datasets used in the Toto paper.
Dataset Notes
- The long-term forecasting subset includes ETT, Electricity, Traffic, Weather, Exchange, and ILI.
- Toto’s LSF evaluation uses ETTh1, ETTh2, ETTm1, ETTm2, Electricity, and Weather.
- The collection also includes short-term forecasting, imputation, anomaly detection, and classification subsets.
Why It Matters
LSF/LTSF datasets remain a common continuity benchmark for time-series forecasting papers. They are useful for comparison with older literature, but should not be confused with high-dimensional or observability-specific benchmarks.
Limitations
- These datasets are low-dimensional relative to BOOM and Time-HD.
- They are often treated as saturated or too narrow to establish general foundation-model transfer.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Benchmarks | warning | Collects common forecasting, imputation, anomaly-detection, and classification datasets used for continuity comparisons. | Low-dimensional passive tasks cannot establish high-channel latent state, context use, or control utility. |
| Native multivariate encoding | insufficient evidence | Some tasks are multivariate, but the collection is often used as LSF/LTSF forecasting continuity data. | Needs high-dimensional channel interactions, metadata, topology, and action/event channels. |