Lag-Llama is a foundation model for probabilistic time series forecasting. It was introduced in the paper Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting by Rasul, Ashok, Williams, Ghonia, Bhagwatkar, Khorasani, Darvishi Bayazi, Adamopoulos, Riachi, Hassen, Biloš, Garg, Schneider, Chapados, Drouin, Zantedeschi, Nevmyvaka, and Rish (2024). The code, model weights, and demos are available in the official GitHub repository. ([GitHub][1])
Lag-Llama is a general-purpose forecasting foundation model for univariate probabilistic time series forecasting. It is based on a decoder-only transformer architecture that uses lags as covariates, allowing it to model temporal dependencies while outputting a probability distribution for each predicted timestep rather than only point forecasts. ([arxiv.org][2])
A key feature of Lag-Llama is its zero-shot forecasting capability: the model can be applied to unseen datasets of different frequencies and prediction lengths without dataset-specific retraining. At the same time, the released codebase also supports finetuning, and the project documentation emphasizes that the model typically performs even better when adapted to a specific dataset. The repository also provides pretraining and finetuning scripts to reproduce the experiments from the paper. ([GitHub][1])
Key traits of Lag-Llama:
Lag-Llama is intended for:
Limitations:
@misc{rasul2024lagllama,
title={Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting},
author={Kashif Rasul and Arjun Ashok and Andrew Robert Williams and Hena Ghonia and Rishika Bhagwatkar and Arian Khorasani and Mohammad Javad Darvishi Bayazi and George Adamopoulos and Roland Riachi and Nadhir Hassen and Marin Bilo{\v{s}} and Sahil Garg and Anderson Schneider and Nicolas Chapados and Alexandre Drouin and Valentina Zantedeschi and Yuriy Nevmyvaka and Irina Rish},
year={2024},
eprint={2310.08278},
archivePrefix={arXiv},
primaryClass={cs.LG}
}