# Lag-Llama: Towards Foundation Models for Time Series Forecasting

**Authors:** Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Arian Khorasani, George Adamopoulos, Rishika Bhagwatkar, Marin Biloš, Hena Ghonia, Nadhir Vincent Hassen, Anderson Schneider, Sahil Garg, Alexandre Drouin, Nicolas Chapados, Yuriy Nevmyvaka, Irina Rish  
**arXiv:** [2310.08278](https://arxiv.org/abs/2310.08278)  
**Date:** 2023-10-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.08278v1/

## Key points

- The paper presents Lag-Llama, a general-purpose univariate probabilistic time-series forecasting model.
- Lag-Llama is trained on a large collection of time-series data and demonstrates impressive zero-shot prediction capabilities on unseen "out-of-distribution" time-series datasets.
- It outperforms supervised baselines in terms of predictive abilities.
- The authors analyze the scaling behavior of Lag-Llama using smoothly broken power-laws.
- The open-source code for Lag-Llama is available at https://github.com/kashif/pytorch-transformer-ts.
- Lag-Llama is introduced as a promising foundation model for time-series forecasting with excellent predictive abilities.

## Summary

The paper titled "Lag-Llama: Towards Foundation Models for Time Series Forecasting" presents the authors' work on Lag-Llama, a general-purpose univariate probabilistic time-series forecasting model. The goal of their research is to build foundation models for time-series forecasting and study their scaling behavior. Lag-Llama is trained on a large collection of time-series data and demonstrates impressive zero-shot prediction capabilities on unseen "out-of-distribution" time-series datasets, outperforming supervised baselines. To analyze the model's scaling behavior, the authors utilize smoothly broken power-laws to fit and predict its performance. The open-source code for Lag-Llama is available at https://github.com/kashif/pytorch-transformer-ts, allowing researchers and practitioners to access and utilize this powerful forecasting tool. Overall, this paper contributes to the field of time-series forecasting by introducing Lag-Llama as a promising foundation model with excellent predictive abilities. The authors' use of power-law analysis provides insights into the scaling behavior of the model, enhancing its applicability in real-world scenarios.