autogluon/fev_datasets

Forecast evaluation datasets

This repository contains time series datasets that can be used for evaluation of univariate & multivariate forecasting models.

The main focus of this repository is on datasets that reflect real-world forecasting scenarios, such as those involving covariates, missing values, and other practical complexities.

The datasets follow a format that is compatible with the `fev` package.

Data format and usage

Each dataset satisfies the following schema:

• —each dataset entry (=row) represents a single univariate or multivariate time series
• —each entry contains
• —1/ a field of type Sequence(timestamp) that contains the timestamps of observations
• —2/ at least one field of type Sequence(float) that can be used as the target time series or dynamic covariates
• —3/ a field of type string that contains the unique ID of each time series
• —all fields of type Sequence have the same length

Datasets can be loaded using the 🤗 `datasets` library.

import datasets

ds = datasets.load_dataset("autogluon/fev_datasets", "epf_de", split="train")
ds.set_format("numpy")  # sequences returned as numpy arrays

Example entry in the epf_de dataset

>>> ds[0]
{'id': 'DE',
 'timestamp': array(['2012-01-09T00:00:00.000000', '2012-01-09T01:00:00.000000',
        '2012-01-09T02:00:00.000000', ..., '2017-12-31T21:00:00.000000',
        '2017-12-31T22:00:00.000000', '2017-12-31T23:00:00.000000'],
       dtype='datetime64[us]'),
 'target': array([34.97, 33.43, 32.74, ...,  5.3 ,  1.86, -0.92], dtype=float32),
 'Ampirion Load Forecast': array([16382. , 15410.5, 15595. , ..., 15715. , 15876. , 15130. ],
       dtype=float32),
 'PV+Wind Forecast': array([ 3569.5276,  3315.275 ,  3107.3076, ..., 29653.008 , 29520.33  ,
        29466.408 ], dtype=float32)}

For more details about the dataset format and usage, check out the `fev` documentation on GitHub.

Dataset statistics

Disclaimer: These datasets have been converted into a unified format from external sources. Please refer to the original sources for licensing and citation terms. We do not claim any rights to the original data. Unless otherwise specified, the datasets are provided only for research purposes.

Citation

If you find these datasets useful in your work, please cite the following paper

@article{shchur2025fev,
  title={{fev-bench}: A Realistic Benchmark for Time Series Forecasting},
  author={Shchur, Oleksandr and Ansari, Abdul Fatir and Turkmen, Caner and Stella, Lorenzo and Erickson, Nick and Guerron, Pablo and Bohlke-Schneider, Michael and Wang, Yuyang},
  year={2025},
  eprint={2509.26468},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}

Publications using these datasets

• —"ChronosX: Adapting Pretrained Time Series Models with Exogenous Variables"