timesfm: Ranked #1 on fev-bench, supports multivariate time series forecasting

This is a time series foundation model released by Google Research. It uses pre-trained weights to forecast time series data across various domains without requiring separate tuning.

Starting with version 3.0, TimesFM natively supports multivariate forecasting and covariate handling. It accepts both past-only covariates and future-known covariates, handling both univariate and multivariate scenarios.

It has achieved the top rank on major benchmarks including fev-bench, TIME Benchmark, and GIFT-Eval. It provides zero-shot generalization performance validated on over 100 real-world tasks and 50 domain datasets.

It is provided as a PyTorch-based library, with weights available for download on Hugging Face Hub. However, the weights for version 3.0 are subject to a license permitting use for non-profit and non-commercial purposes only.