TimesFM

TimesFM is a pretrained decoder-only foundation model for time-series forecasting, built by Google Research. It's aimed at developers and data scientists who need forecasts without training a model from scratch for every new series. Instead of fitting a model per dataset, you feed TimesFM raw historical values and it produces point forecasts and quantiles directly, using patterns learned from a large corpus of time series during pretraining.

The project has evolved through several major versions. Version 3.0 adds native multivariate forecasting and flexible covariate support (both past-only and past-and-future), and claims top rankings on fev-bench, TIME Benchmark, and GIFT-Eval. Earlier versions (1.0, 2.0, 2.5) remain available and are archived in the repo, with 2.5 offering a smaller 200M-parameter model, longer context windows, and quantile forecasting via an optional head.

Key features

• Zero-shot forecasting: generate forecasts on unseen time series without task-specific fine-tuning.
• Multivariate support: TimesFM 3.0 natively handles multi-channel series, not just single univariate sequences.
• Covariate handling: past-only and past-and-future dynamic covariates are supported without per-task tuning.
• Quantile outputs: returns point predictions alongside multiple quantiles (0.1 to 0.9) for uncertainty estimation.
• Variable-length context: accepts batches of series with different history lengths in a single call.
• Fine-tuning support: includes a LoRA fine-tuning example built on HuggingFace Transformers + PEFT.
• Legacy version access: older checkpoints (1.0, 2.0, 2.5) remain installable and usable via archived code paths.
• Integration into Google products: available through BigQuery ML, Google Sheets (Connected Sheets), and Vertex Model Garden for managed use.

Ideal use cases

TimesFM fits well when you need quick forecasts across many series (retail demand, capacity planning, sensor data, metrics dashboards) and don't want to train and maintain a separate model per series. It's useful for prototyping forecasting pipelines, benchmarking against classical or task-specific models, or as a baseline generalist forecaster inside a larger system.

It's also a reasonable choice when covariates matter, since 3.0's native covariate support avoids manual feature engineering steps that other zero-shot approaches require.

It's not a good fit if you need commercial or production use of the default TimesFM 3.0 pretrained weights, since those are licensed under a non-commercial restriction (weights up to 2.5 remain Apache-2.0). It's also not ideal if your series has highly domain-specific structure that a purpose-built, fine-tuned model would capture better, or if you require full interpretability of the forecasting mechanism, since it's a large pretrained decoder rather than a transparent statistical model.

Installation

Install via PyPI with PyTorch support:

pip install timesfm[torch]

For a local, editable install:

git clone https://github.com/google-research/timesfm.git
cd timesfm

# Using uv
uv venv
source .venv/bin/activate

# Install the package in editable mode with torch
uv pip install -e .[torch]

To use an older model version (1.0 or 2.0), install a matching older package release, for example:

pip install timesfm==1.3.0

Once installed, load a checkpoint and call the predictor directly, as shown in the repo's code examples for univariate and multivariate forecasting with TimesFM3Evaluator and ModelConfig. Checkpoints are hosted on Hugging Face, including google/timesfm-3.0-pytorch for the latest version and earlier releases in the TimesFM Hugging Face collection.