MUSEval

We present MUSEval, a comprehensive benchmark for evaluating foundation models on multivariate time series forecasting tasks. MUSEval spans 16 multivariate time series domains and introduces novel synthetic data techniques, comprising 19B data points across 825K time series.

Dataset Structure

Data is structured into 4 main categories, comprising the following domains:

• Real-World: Energy, Public Info, Health, Sales, Climate & Environment, Engineering, Finance, Web
• Collections: Stock, Wikipedia
• Synthetic: Dynamic, Causal Model
• Derived: Images, Text, Video, Scientific

Dataset Information

The dataset preview summarizes metadata and key statistics for each time series dataset included in MUSEval. In particular,

• number_of_time_series: The total number of distinct, individual time series contained within each dataset.
• start_timestamp: The earliest timestamp in the dataset, indicating when the time series data begins.
• frequency: The temporal resolution of the time series (hourly, daily, weekly, etc.)
• total_data_points: The total number of data points across all time series and all variates in the dataset.
• total_timestamps: The total number of unique timestamps across all time series in the dataset.
• target_columns and covariate_columns: The names of the target variable being forecasted and variables used as covariates, respectively.
• target_values: A preview of target column values.

Usage Instructions

To submit your results to our leaderboard, please follow the instructions detailed in our GitHub repository.

Citation Information

If you use MUSEval in your research, please cite the original paper:

@article{mused-fm2025,
    title={MUSED-FM: A Multivariate Time Series Evaluation Dataset for Foundation Models},
    author={Synthefy Research Team},
    journal={arXiv preprint},
    year={2025}
}