TL;DR
TimesFM is a pretrained foundation model for time-series forecasting, offering generalized forecasting performance across domains.
Key features
Pretrained model for time-series forecasting (decoder-only architecture)
Supports context length up to 16k
Continuous quantile forecasting (up to 1k horizon)
Supports PyTorch and Flax backends
Covariate support via XReg
Fine-tuning examples with HuggingFace Transformers + PEFT (LoRA)
When to use it
General time-series forecasting tasks across various domains
When leveraging pretrained models for good forecasting performance with limited data
Integration with Google products like BigQuery ML, Google Sheets, and Vertex Model Garden