lag-llama

Why we included this project

Lag-Llama is one of the first open-source foundation models for time series forecasting, and the practical appeal is that you can throw a series you have never seen at it and get a usable forecast. Because it is probabilistic, each predicted step comes back as a full distribution rather than a single number, so you get confidence bands without building a bespoke model for every dataset. It works zero-shot across different data frequencies and prediction horizons, and the README is upfront about the edges: it explains context-length tuning and RoPE scaling for when you need to go beyond what the model was trained on. When zero-shot accuracy is not enough, the repo provides pretraining and finetuning scripts along with concrete guidance on context length and learning rate. Weights are on Hugging Face and the code is Apache-2.0, which keeps it easy to fold into research or production work.