GIFT-Eval Pretraining Time Series Collection with 4.5 Million Series

GIFT-Eval Pre-training Datasets contain 4.5 million univariate and multivariate time series totaling 230 billion data points, spanning seven domains and 13 frequencies. The collection, created by Salesforce, is designed for pretraining foundation models and is explicitly aligned with the GIFT-Eval benchmark to avoid data leakage between training and testing splits.

Use Cases

• Pretrain foundation models on 4.5 million univariate time series for forecasting tasks aligned with the GIFT-Eval benchmark.
• Develop models for multivariate analysis using the 17 multivariate datasets across the seven specified domains.
• Benchmark time series models across the 13 different frequencies present in the collection to evaluate generalization.
• Leverage the data's explicit split design to pretrain models without leakage issues for fair evaluation on GIFT-Eval.

Strengths

• Contains 4.5 million individual time series, providing a large-scale resource for model pretraining.
• Encompasses 230 billion total data points across univariate and multivariate series.
• Spans seven distinct domains and 13 different frequencies, offering broad coverage of time series characteristics.
• Explicitly designed with no data leakage between training and testing splits, ensuring fair evaluation on the GIFT-Eval benchmark.

Limitations

• Specific column definitions, sample data, and file formats are not provided, complicating initial data exploration.
• The raw description lacks details on data provenance, collection methodology, and specific domain definitions.
• Without known row counts per dataset or detailed metadata, assessing individual dataset scale and balance is difficult.

Provenance

Source

Salesforce

Collection Method

Aligned with the GIFT-Eval benchmark; specific gathering methodology not detailed.

Freshness

Last updated on January 21, 2025.