Predict multivariate time series with auxiliary variables using a pre-trained model
When you need to estimate future values from daily data streams, you can calculate them alongside other variables that move with past records. TimesFM, created by Google Research, is a pre-trained time series model that handles various data formats without separate tuning. Version 3.0 specifically adds the ability to predict multiple channels of data simultaneously and naturally incorporates past and future auxiliary variables.
New prediction approach in version 3.0
TimesFM 3.0 supports multivariate time series, where multiple variables move together, in addition to single variables. It can use information that exists only in the past and information known for the future as auxiliary variables, enabling more refined predictions in complex situations. The model is ready to use in its pre-trained state and does not require adjustment for specific tasks. It also recorded first place in three major time series forecasting benchmarks, demonstrating its performance.
Environment and installation
The model is written in Python and supports PyTorch or the MLX backend for Apple silicon. You can install the PyTorch version from PyPI using the pip install timesfm[torch] command, and if you use an Apple chip, you can run it without PyTorch using pip install timesfm[mlx]. To develop locally, clone the repository, create a virtual environment, and install it in editable mode. Checkpoints for version 2.5 and below follow the Apache-2.0 license, but the pre-trained weights for version 3.0 are restricted to non-commercial and non-production use.
Commercial use and license distinctions
The source code is released under the Apache-2.0 license, allowing free modification and use. However, the pre-trained weights for version 3.0 are subject to a separate non-commercial license (timesfm-non-commercial-license-v1.0). Therefore, directly hosting downloaded weights or using them for commercial purposes is not permitted. Use in commercial and production environments is only possible through official Google Cloud services such as BigQuery ML. Weights for version 2.5 and below retain the Apache-2.0 license, allowing commercial use.
Items to check before application
You must verify the license conditions when using this model. If you plan to directly download version 3.0 weights and deploy them to commercial services or production environments, you must use official channels such as BigQuery ML or Vertex AI on Google Cloud. If you are using it in a non-commercial environment for personal projects or research, you can install it locally and use it immediately. Additionally, when using the MLX backend, it is advisable to check for numerical consistency with PyTorch and to inspect errors at patch connection points during long time series predictions.
