MOMENT is a family of open-source foundation models designed for general-purpose time-series analysis, targeting researchers and practitioners. It addresses challenges in large-scale time-series pre-training by introducing the Time-series Pile, a diverse dataset collection, and a benchmark for evaluating models under limited supervision, offering improved performance on imputation, anomaly detection, classification, and forecasting tasks.

How It Works

MOMENT employs a patching strategy, dividing time-series into fixed-length sub-sequences that are then mapped to patch embeddings. During pre-training, random patches are masked and the model learns to reconstruct the original time-series using a lightweight reconstruction head. This approach allows MOMENT to capture subtle temporal characteristics like trend, scale, frequency, and phase, and learn meaningful representations even without task-specific fine-tuning.

Quick Start & Requirements

• Install via pip: pip install momentfm or pip install git+https://github.com/moment-timeseries-foundation-model/moment.git
• Recommended Python version: 3.11
• Tutorials and experiments can be reproduced on a single NVIDIA A6000 GPU with 48 GiB RAM.
• Official documentation and tutorials: https://github.com/moment-timeseries-foundation-model/moment

Highlighted Details

• Outperforms statistical imputation baselines and is competitive with other methods in anomaly detection, classification, and short-horizon forecasting without fine-tuning.
• Achieves state-of-the-art results on anomaly detection and competitive performance in long-horizon forecasting via linear probing.
• Embeddings capture nuanced time-series properties like trend, scale, frequency, and phase.
• Demonstrates distinct class representations for datasets like ECG5000 without specific fine-tuning.

Maintenance & Community

• The project is associated with Auton Lab at Carnegie Mellon University.
• Active development is indicated by recent releases of small and base model versions and research code.
• Contributions are encouraged, with guidelines to be provided.
• Related work includes JoLT for multimodal time-series and text foundation models.

Licensing & Compatibility

• Licensed under the MIT License, permitting commercial use and closed-source linking.

Limitations & Caveats

• While the research code is available, it is noted as "messier" than the released package.
• Support for Python versions other than 3.11 is expected but not guaranteed at the time of the README.