Google open‑sources TimesFM
Google published TimesFM, a time‑series forecasting model pre‑trained on about 100 billion data points and positioned for zero‑shot forecasting tasks like sales and markets. The open‑source release claims the model outperforms some supervised baselines on forecasting benchmarks. (x.com)
Time-series forecasting is the practice of predicting the next few points in a sequence — next week’s sales, tomorrow’s power demand, or a stock index at the next close. Google has released TimesFM as open-source software, giving outside developers access to a forecasting model the company says was pre-trained on 100 billion real-world time points. (research.google) Google Research described TimesFM as a decoder-only transformer, a model design borrowed from language models, and published code and checkpoints through its GitHub repository and Hugging Face collection. The repository says the current open release is TimesFM 2.5, while older 1.0 and 2.0 versions are archived in a separate directory. (github.com) The core pitch is “zero-shot” forecasting, which means a model can make predictions on a new dataset without being retrained for that specific series. Google’s research paper says TimesFM’s out-of-the-box performance came close to fully supervised models built and tuned for individual datasets across public benchmarks. (arxiv.org) Google’s 2023 research post said the first TimesFM model was trained on 100 billion time points. Google Cloud later said the BigQuery preview used a TimesFM model trained on 400 billion real-world time points, showing the line has continued to expand inside Google’s products after the original research release. (research.google, cloud.google.com) That matters because forecasting systems are usually built one dataset at a time, with separate training pipelines for retail demand, traffic, weather, or industrial sensors. A pre-trained model aims to work more like a general-purpose assistant: learn broad patterns first, then handle a new series with little or no extra setup. (arxiv.org, docs.cloud.google.com) Google has also moved TimesFM beyond research code. BigQuery Machine Learning now offers a built-in TimesFM model in all BigQuery-supported regions, and Google Cloud documentation says that product implementation is based on the open-source Google Research model. (docs.cloud.google.com) The open repository adds some caveats. Google says the public code is “not an officially supported Google product,” and the repository recommends at least 32 gigabytes of memory to load dependencies for some versions. (github.com) The model itself is relatively small by modern foundation-model standards. Third-party documentation built from the repository says TimesFM 2.5 is a 200 million-parameter PyTorch model with support for context windows up to 16,000 points and optional quantile forecasts, which estimate a range of likely outcomes instead of a single number. (deepwiki.com) Google’s paper frames the release as part of a broader effort to apply language-model techniques to numerical sequences rather than words. With code, checkpoints, and a cloud product now in place, TimesFM is no longer just a research paper — it is infrastructure other teams can test against their own data. (arxiv.org, github.com)
