Unified Time Series Dataset with Hierarchical Volumes

UTSD aggregates time series data across 7 distinct domains, containing up to 1 billion individual time points. The dataset is structured into four hierarchical volumes to support research on large models and pre-training. It was created by the thuml organization and updated in 2024 and 2025.

Use Cases

• Benchmarking pre-training algorithms on hierarchical time series volumes.
• Analyzing cross-domain patterns using data from 7 different fields.
• Developing time series forecasting models on datasets scaling to 1 billion points.

Strengths

• Contains up to 1 billion time points.
• Covers data from 7 different domains.
• Offers a hierarchical four-volume structure for research.

Limitations

• Specific column names, data formats, and exact row counts are not provided.
• The sample size for individual domains or series is unknown.
• Geographic and temporal coverage details are unspecified.

Provenance

Source

thuml organization on Hugging Face.

Collection Method

Curated from multiple domains; specific gathering method not detailed.

Freshness

Updated with a new numpy format in October 2024; last platform update was June 2025.