MOMENT is an initiative focused on developing open-source foundation models for general-purpose time series analysis. The paper highlights significant challenges in pre-training due to a lack of extensive public dataset repositories, diverse characteristics of time series data, and limited benchmarks for performance evaluation. To overcome these hurdles, the research compiles the Time Series Pile, a comprehensive dataset for multi-dataset pre-training. The authors also propose a benchmark for evaluating the performance of time series models under constrained conditions, demonstrating the effectiveness of their pre-trained models in resource-limited scenarios.
MOMENT addresses the challenges of time series analysis by leveraging a large, diverse dataset called the Time Series Pile for effective model training and evaluation.
The introduction of the Time Series Pile allows for large-scale pre-training of models, tackling challenges related to the lack of cohesive public datasets and resource limitations.
Our experiments confirm that pre-trained models can achieve effective performance with minimal data and supervision, highlighting the potential of foundation models in time series applications.
We establish a benchmark to systematically evaluate time series foundation models across diverse tasks, addressing the nascent stage of model evaluation in this area.
Collection
[
|
...
]