TimesFM

TimesFM (Time Series Foundation Model) is a pretrained decoder-only foundation model developed by Google Research for time-series forecasting. It supports zero-shot forecasting across a wide range of time-series tasks and was presented at ICML 2024. The latest version, TimesFM 2.5, uses 200M parameters, supports up to 16k context length, and includes an optional 30M quantile head for continuous quantile forecasting up to 1k horizon.

• Pretrained Foundation Model: Download and use pre-trained checkpoints from Hugging Face without training from scratch.
• Large Context Window: TimesFM 2.5 supports up to 16k context length, a significant upgrade from the 2048 of prior versions.
• Continuous Quantile Forecasting: An optional 30M quantile head enables probabilistic forecasts from the 10th to 90th quantile.
• Multiple Backends: Supports both PyTorch and Flax (JAX) backends, with optional GPU, TPU, and Apple Silicon acceleration.
• Covariate Support via XReg: External regressors (covariates) can be incorporated through the XReg module for richer forecasting.
• ForecastConfig API: A flexible configuration object controls context length, horizon, normalization, quantile settings, and more.
• BigQuery Integration: Available as an official Google product inside BigQuery for enterprise-scale forecasting workflows.
• Open Source (Apache 2.0): Clone the repository, install via pip or uv, and run forecasts with just a few lines of Python.
• Versioned Releases: Older model versions (1.0, 2.0) are archived and installable via pip install timesfm==1.3.0 for backward compatibility.