datasets Submodule

LTSF Datasets

LTSF is a collection of time series forecasting datasets that are commonly used in benchmarking forecasting algorithms. Typically, the performance is reported as the mean squared error and mean absolute error over multiple forecasting horizons: 96, 192, 336, and 720 time steps.

class torchcast.datasets.ElectricityLoadDataset(path: str | None = None, split: str = 'all', download: str | bool = True, scale: bool = True, columns_as_channels: bool = True, transform: Callable | None = None, input_margin: int | None = 336, return_length: int | None = None)

Electricity Load dataset, obtained from:

https://github.com/laiguokun/multivariate-time-series-data

This is derived from:

https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014

But the data has been subsetted and pre-processed. It is sometimes abbreviated as the ECL dataset. The dataset consists of two series, the first consisting of timestamps and the second of the data itself.

class torchcast.datasets.ElectricityTransformerDataset(path: str | None = None, task: str = '15min', split: str = 'all', download: bool | str = True, scale: bool = True, columns_as_channels: bool = True, transform: Callable | None = None, input_margin: int | None = 336, return_length: int | None = None)

This is the Zhou et al. 2021 electricity transformer dataset, obtained from:

https://github.com/zhouhaoyi/ETDataset

This is sometimes abbreviated as the ETT dataset.

The dataset contains three series: a timestamp, the time series itself, and a series of covariate variables.

class torchcast.datasets.ExchangeRateDataset(path: str | None = None, split: str = 'all', download: bool | str = True, scale: bool = True, columns_as_channels: bool = True, transform: Callable | None = None, input_margin: int | None = None, return_length: int | None = None)

This is the Lai et al. 2017 dataset of currency exchange rates, taken from:

https://github.com/laiguokun/multivariate-time-series-data

The dataset consists of two series, the first consisting of timestamps and the second of the data itself.

class torchcast.datasets.GermanWeatherDataset(path: str | None = None, year: int | Iterable[int] = 2020, site: str | Iterable[str] = 'beutenberg', split: str = 'all', download: bool | str = True, scale: bool = True, columns_as_channels: bool = True, transform: Callable | None = None, input_margin: int | None = 336, return_length: int | None = None)

This is a dataset of weather data from Germany, obtained from:

https://www.bgc-jena.mpg.de/wetter/weather_data.html

This dataset was first used in time series forecasting in Zeng et al. 2022, which used only the data from Beutenberg in 2020.

The dataset contains two series, the first being a timestamp and the second the multivariable series to be forecast.

class torchcast.datasets.ILIDataset(path: str, split: str = 'all', scale: bool = True, columns_as_channels: bool = True, transform: Callable | None = None, input_margin: int | None = 336, return_length: int | None = None)

This dataset describes both the raw number of patients with influenza-like symptoms and the ratio of those patients to the total number of patients in the US, obtained from the CDC. This must be manually downloaded from:

https://gis.cdc.gov/grasp/fluview/fluportaldashboard.html

To download this dataset, click “Download Data”. Unselect “WHO/NREVSS” and select the desired seasons, then click “Download Data”.

The dataset consists of two series, the first consisting of timestamps and the second of the data itself.

class torchcast.datasets.SanFranciscoTrafficDataset(path: str | None = None, split: str = 'all', download: str | bool = True, scale: bool = True, columns_as_channels: bool = True, transform: Callable | None = None, input_margin: int | None = None, return_length: int | None = None)

This is the San Francisco traffic dataset from Lai et al. 2017 <https://arxiv.org/abs/1703.07015>, obtained from:

https://pems.dot.ca.gov

The dataset contains two series, the first being a timestamp and the second the multivariable series to be forecast.

Monash Archive Datasets

The Monash archive is a collection of time series forecasting datasets in a standard format.

class torchcast.datasets.MonashArchiveDataset(task: str, split: str = 'train', path: str | None = None, download: str | bool = True, transform: Callable | None = None, return_length: int | None = None)

This provides access to all Monash forecasting archive datasets, as discussed in Godahewa et al. 2021.

Depending on the task, the dataset consists of either one or two series. The first, optional series consists of timestamps, and the second of the data itself.

UCR/UEA Datasets

The UCR/UEA archive is a collection of time series classification datasets in a standard format. The UCR archive provides univariate time series, while the UEA archive provides multivariate time series.

class torchcast.datasets.UCRDataset(task: str, split: str = 'train', path: str | None = None, download: bool | str = True, transform: Callable | None = None, return_length: int | None = None)

This is the UCR dataset for univariate time series classification, found at:

https://www.timeseriesclassification.com/

The dataset consists of two series, the first consisting of the time series itself and the second of the label to be predicted.

class torchcast.datasets.UEADataset(task: str, split: str = 'train', path: str | None = None, download: bool | str = True, transform: Callable | None = None, return_length: int | None = None)

This is the UEA dataset for multivariate time series classification, found at:

https://www.timeseriesclassification.com/

This includes the additional datasets introduced by:

Middlehurst et al. 2026. ``The Multiverse of Time Series Machine Learning: an Archive for Multivariate Time Series Classification.’’ arXiv: 2603.20352.

The dataset consists of two series, the first consisting of the time series itself and the second of the label to be predicted.

Other Datasets

class torchcast.datasets.AirQualityDataset(path: str | None = None, download: bool | str = True, transform: Callable | None = None, return_length: int | None = None)

This is the De Vito et al. 2008 air quality dataset.

The dataset contains two series, the first being a timestamp and the second the multivariable series to be forecast.

class torchcast.datasets.BOOMDataset(task: int | str, path: str | None = None, download: bool | str = True, transform: Callable | None = None, return_length: int | None = None)

This is the Benchmark Of Observability Metrics (BOOM) dataset, a dataset of time series collected by DataDog from monitoring of pre-production environments. More information can be found at Cohen et al. 2025.

The dataset contains two series, the first being a timestamp and the second the multivariable series to be forecast.

class torchcast.datasets.LargeSTDataset(root_path: str, years: Iterable[int] | int, task: str = 'ca', split: str = 'all', scale: bool = True, resample_time: bool = True, transform: Callable | None = None, return_length: int | None = None)

This dataset is the LargeST traffic dataset from Liu et al. 2023. It must be downloaded manually from:

https://www.kaggle.com/datasets/liuxu77/largest

And unzipped.

The dataset consists of two series, the first consisting of timestamps and the second of the data itself.

class torchcast.datasets.MonsterDataset(task: str, split: str = 'train', fold: int = 0, path: str | None = None, download: bool | str = True, transform: Callable | None = None, return_length: int | None = None)

This is Monster time series classification dataset, first published in Dempster et al. 2025.

The dataset consists of two series, the first consisting of the time series and the second of the class labels.

class torchcast.datasets.STEADDataset(paths: str | Iterable[str], transform: Callable | None = None, return_length: int | None = None)

This is the STanford EArthquake Dataset, consisting of seismic signals to be classified as earthquake or not-earthquake. More information can be found at:

Mousavi et al. 2019. ``STanford EArthquake Dataset (STEAD): A Global Data Set of Seismic Signals for AI.’’ IEEE Access, vol. 7, pp. 179464-179476. Link

The dataset must be downloaded manually from:

https://github.com/smousavi05/STEAD

The data consists of two series: the waveform itself and the class label.

class torchcast.datasets.TailedTSDataset(path: str | None = None, download: bool | str = False, transform: Callable | None = None, return_length: int | None = None)

class torchcast.datasets.TempusDataset(task: str, path: str | None = None, download: bool = True, transform: Callable | None = None, return_length: int | None = None)

This is the Tempus dataset for benchmarking time series foundation models, found `here<https://github.com/Smlcrm/TempusBench>`__ and documented by Goktas et al. 2026.

The number of series in the dataset is variable. It may consist of:

• A single series containing the data.

• A pair of series, the first containing the timestamp and the second the data.

• A trio of series: the timestamp, the data, and covariates that are to be used in forecasting the data.

class torchcast.datasets.TFBDataset(task: str, split: str = 'all', path: str | None = None, download: bool = True, transform: Callable | None = None, return_length: int | None = None)

This dataset provides the Qiu et al. 2024 TFB Time Series Forecasting Benchmark collection, obtained from:

https://github.com/decisionintelligence/TFB

This collection includes the LTSF collection as a subset, but the data is pre-processed and split in different ways.

Depending on the task, the data consists of either a single time series, or the timestamp and the time series.

class torchcast.datasets.TIMEDataset(task: str, freq: str | None = None, path: str | None = None, download: bool | str = True, transform: Callable | None = None, return_length: int | None = None)

This is TIME time series forecasting dataset, first published in Qiao et al. 2026. It is intended for evaluating time series foundation models, and so has no train split.

The dataset contains two series, the first being a timestamp and the second the multivariable series to be forecast.

class torchcast.datasets.UTSDDataset(task: str, split: str = 'default', path: str | None = None, download: bool | str = True, transform: Callable | None = None, return_length: int | None = None)

This is the UTSD pre-training dataset, first published in Liu et al. 2024.

The dataset consists of a single series containing the time series.