In this study, we address the scarcity of large public datasets for time series analysis compared to fields like NLP and computer vision. We introduce the Time Series Pile, a compilation of multiple datasets across four task-specific repositories, designed to provide a robust foundation for pre-training transformer models. This collection spans various domains and diverse characteristics, enabling improvements in time series forecasting and classification. Through this approach, we aim to enhance the performance of models in long-horizon forecasting tasks and related applications.
To bridge the gap in public time series datasets for pre-training models, we collate multiple task-specific datasets into the Time Series Pile, encompassing diverse domains.
Unlike NLP and computer vision, where large datasets are readily available for model training, public time series datasets are scattered and limited in size, necessitating our approach.
Collection
[
|
...
]