Lag-Llama: Open-Source Foundation Model for Time Series Forecasting
Explore the architecture of Lag-Llama and learn to apply it in a forecasting project using Python
In October 2023, I published an article on TimeGPT, one of the first foundation model for time series forecasting, capable of zero-shot inference, anomaly detection and conformal prediction capabilities.
However, TimeGPT is a proprietary model that is only accessed via an API token. Still, it sparked more research in foundation models for time series, as this area has been lagging compared to natural language processing (NLP) and computer vision.
Fast-forward to February 2024, and we now have an open-source foundation model for time series forecasting: Lag-Llama.
In the original paper: Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting, the model is presented as a general-purpose foundation model for univariate probabilistic forecasting. It was developed by a large team from different institutions like Morgan Stanley, ServiceNow, Université de Montréal, Mila-Quebec, and McGill University.
In this article, we explore the architecture of Lag-Llama, its capabilities and how it was trained. Then we actually use Lag-Llama in a forecasting project…
