OpenLTM provides a comprehensive codebase for developing and evaluating Large Time-Series Models (LTMs). It targets researchers and practitioners in time-series analysis, offering pre-training code, datasets, and scripts for various state-of-the-art models like Timer-XL, Moirai, and GPT4TS, enabling efficient zero-shot and few-shot forecasting.

How It Works

The project leverages foundation backbones, primarily Transformers, and large-scale pre-training to create scalable deep models for time-series data. This approach allows models to generalize across diverse datasets and downstream tasks, offering strong zero-shot forecasting capabilities by learning a "language" of time series.

Quick Start & Requirements

• Install via pip install -r requirements.txt.
• Requires Python 3.11.
• Datasets should be placed in the ./dataset folder. Pre-training datasets include UTSD (1 billion time points) and ERA5-Family. Supervised training datasets are available from TSLib.
• Scripts for supervised training, pre-training, and adaptation are provided.
• Official Hugging Face model links are available for Chronos, Moirai, TimesFM, Timer-XL, Time-MoE, and TTMs.

Highlighted Details

• Implements and provides pre-training code for multiple LTMs including Timer-XL, Moirai, GPT4TS, and TTMs.
• Includes scripts for supervised training, large-scale pre-training (UTSD, ERA5), and model adaptation (full-shot, few-shot).
• Supports zero-shot forecasting evaluation with links to pre-trained models on Hugging Face.
• Offers a pipeline to develop and integrate custom large time-series models.

Maintenance & Community

The project acknowledges contributions from various GitHub repositories and lists key contributors for inquiries. Specific community channels (Discord/Slack) or roadmaps are not explicitly mentioned in the README.

Licensing & Compatibility

The README does not explicitly state a license. However, given the project's nature and common practices in academic research, it is likely intended for research purposes. Commercial use compatibility would require explicit clarification.

Limitations & Caveats

While LTMs are presented as scalable, the README notes they are still smaller than foundation models in other modalities. Pre-training may require significant computational resources (e.g., A100 GPUs), though adaptation can be done on hardware like RTX 4090s. The project is actively updated with new models and implementations.