Google open‑sources TimesFM
Google has open‑sourced TimesFM, a pre‑trained time‑series forecasting model trained on 100 billion real‑world data points that reportedly performs zero‑shot forecasting for sales, market prices and volatility without fine‑tuning. The project was announced on X and positions the model as a tool for predictive analytics in business and finance. (x.com)
Time-series forecasting is a way to predict the next points in a sequence — next week’s sales, tomorrow’s power demand, or the next move in a stock price. Google has now released the code for its TimesFM forecasting model on GitHub. (github.com) Google Research describes TimesFM as a pre-trained “foundation model” for forecasting, meaning it learns patterns from a very large training set and can then be applied to new series without building a separate model from scratch. Google’s research blog said the model was trained on 100 billion real-world time points. (research.google) The public repository says TimesFM is a pre-trained model for time-series forecasting and links it to the paper “A decoder-only foundation model for time-series forecasting,” which was presented at the International Conference on Machine Learning in 2024. The repository also says the latest open version is TimesFM 2.5, while older 1.0 and 2.0 versions are archived. (github.com) Forecasting models usually need task-specific training on each dataset before they can make useful predictions. Google’s paper says TimesFM was built to do “zero-shot” forecasting, which means a user can feed in historical values from a new dataset and ask for a forecast without fine-tuning the model first. (arxiv.org) Google said in its research post that TimesFM showed strong zero-shot results across public benchmarks from retail, finance, manufacturing, healthcare, and natural sciences. In the paper, the authors wrote that the model’s out-of-the-box accuracy came close to fully supervised systems on a range of benchmark datasets. (research.google) (arxiv.org) The release also fits into a broader product push inside Google’s data tools. Google Cloud said in December 2025 that TimesFM had been integrated into BigQuery and AlloyDB, and BigQuery documentation now describes a built-in TimesFM model available across BigQuery-supported regions. (cloud.google.com 1) (cloud.google.com 2) Google has also started putting TimesFM in spreadsheet workflows aimed at business users. In a February 17, 2026 update, Google said Connected Sheets for BigQuery could use BigQuery Machine Learning and TimesFM to forecast sales, demand volume, and other business metrics directly inside Google Sheets. (workspaceupdates.googleblog.com) The open repository makes the model easier for developers and researchers to inspect and run outside Google’s hosted products, but Google also says the open version “is not an officially supported Google product.” The repository links to downloadable checkpoints hosted on Hugging Face. (github.com) (huggingface.co) The immediate test is whether TimesFM’s open release turns a research result into a standard forecasting tool. Google is already using the model inside BigQuery and Sheets, and the GitHub release gives outside users a direct way to try the same approach on their own data. (cloud.google.com) (github.com)
