Google publishes TimesFM

Time-series forecasting is the job of predicting the next points in a sequence — next week’s sales, tomorrow’s traffic, next month’s demand. Google has now published TimesFM as open-source code and model checkpoints for that task. (github.com) Google Research says TimesFM is a pre-trained forecasting model that can make “zero-shot” predictions, meaning it can be applied to new datasets without training a custom model first. Google’s February 2, 2024 research post said the model was trained on 100 billion real-world time points and was later released on GitHub and Hugging Face after the paper was accepted to the International Conference on Machine Learning 2024. (research.google) The basic idea mirrors large language models, but for sequences of numbers instead of words. In the paper, the authors said they trained a decoder-only transformer on large collections of time-series data so it could learn recurring patterns across retail, finance, manufacturing, healthcare, traffic, energy and weather data. (arxiv.org) Traditional forecasting usually means fitting a separate model to each dataset, then tuning it for that business or sensor feed. Google said a foundation model can skip much of that setup and produce out-of-the-box forecasts on unseen series, which is useful when teams do not have enough labeled history or machine-learning staff to build custom pipelines. (research.google) Google’s public materials make a narrower claim than some social posts: the company said TimesFM’s zero-shot results came close to state-of-the-art supervised models on previously unseen public benchmarks, not that it beat every supervised approach across every dataset. The paper and blog both frame the result as competitive performance across different domains, forecast horizons and time granularities. (research.google) (arxiv.org) The open repository shows the project has continued to evolve after the initial 2024 release. Google Research’s GitHub page lists TimesFM 2.5 as the latest model version, with a March 19, 2026 update note and earlier notes saying the September 15, 2025 release cut the model to 200 million parameters, expanded context length to 16,000 points from 2,048, and added optional quantile forecasts up to a 1,000-step horizon. (github.com) Google has also moved the model into products, which helps explain why the open release matters beyond research. BigQuery documentation says its built-in TimesFM model is an implementation of Google Research’s open-source model and is available in all BigQuery-supported regions through the AI.FORECAST function. (cloud.google.com) Google Workspace pushed that one step closer to spreadsheet users on February 17, 2026. In a product update, Google said Connected Sheets users can generate forecasts from BigQuery data with TimesFM inside Google Sheets, without writing Structured Query Language or Python. (workspaceupdates.googleblog.com) There are still limits. BigQuery’s documentation says the built-in TimesFM model is univariate, meaning it forecasts one series at a time, and Google directs users who need more tuning options to other models such as ARIMA_PLUS or ARIMA_PLUS_XREG. (cloud.google.com) So the release is less about replacing every forecasting system than about lowering the entry cost for a common business task. Google’s pitch is that a model trained once on very large time-series corpora can give teams a usable first forecast before they invest in building one from scratch. (research.google)