Lag-Llama

Lag-Llama is a foundation model for univariate probabilistic time-series forecasting from ServiceNow Research, Morgan Stanley, and Mila. The decoder-only transformer uses time lags as covariates, which lets a single pretrained model produce calibrated probabilistic forecasts on unseen datasets zero-shot, matching or beating dataset-specific supervised baselines in the original study.

The code and checkpoints are open under Apache 2.0, built on the GluonTS framework, and the paper includes neural scaling laws for sizing decisions. It was among the first open time-series foundation models and remains a reference point for the category. Honest limitations: it is univariate only, so multivariate problems need other tools, and the GluonTS-based setup is heavier than a pip install of a modern HF checkpoint.