Google open‑sources TimesFM
Google open‑sourced TimesFM, a time‑series forecasting model pre‑trained on 100 billion real‑world data points and positioned for zero‑shot forecasting of sales trends, market prices and volatility. The project is presented as outperforming supervised models without fine‑tuning in social reporting about the release. (x.com)
Time-series forecasting is the job of predicting the next points in a sequence — next week’s sales, tomorrow’s power demand, next month’s prices. Google Research has released TimesFM as open-source code and model weights for that task. (research.google) Google said on February 2, 2024 that TimesFM was pre-trained on 100 billion real-world time points and posted the model on GitHub and Hugging Face in a May 8, 2024 update. The paper behind it was later accepted at the International Conference on Machine Learning, or ICML, 2024. (research.google; github.com; icml.cc) The basic pitch is “zero-shot” forecasting: users feed in historical data, and the model produces forecasts for new series without training a custom model first. Google’s paper said that out-of-the-box performance on public datasets came close to, and in some cases matched, state-of-the-art supervised systems trained for each dataset. (arxiv.org; research.google) That is a shift from older forecasting workflows, where teams often trained separate models for each business problem and each dataset. Google said many deep-learning forecasters still require long training and validation cycles before a customer can test them on new time series. (research.google) TimesFM borrows the “foundation model” idea from large language models, but for numbers over time instead of words in a sentence. The paper describes a decoder-only transformer that is pre-trained on many different sequences so it can recognize patterns across different history lengths, forecast lengths, and time intervals. (arxiv.org) The first public checkpoint was a 200 million-parameter model for univariate forecasting, meaning one variable at a time, such as daily sales or hourly traffic. Its Hugging Face card said that release handled context windows up to 512 time points and focused on point forecasts rather than fully calibrated probability forecasts. (huggingface.co) Google has kept building on the open model since then. The GitHub repository says the latest open version is TimesFM 2.5, updated in 2025 and 2026, with 200 million parameters, support for context lengths up to 16,000, and an optional 30 million-parameter quantile head for forecasting uncertainty up to a 1,000-step horizon. (github.com) Google Cloud has also turned the research model into a product feature. BigQuery documentation says its built-in TimesFM model is an implementation of the open-source Google Research model and can be used through AI.FORECAST, AI.EVALUATE, and AI.DETECT_ANOMALIES functions. (docs.cloud.google.com) Google’s own documentation also draws a limit around the tool. BigQuery says forecast quality is comparable to conventional methods such as ARIMA, and points users to ARIMA_PLUS or ARIMA_PLUS_XREG when they need more tuning options than TimesFM offers. (docs.cloud.google.com) So the release is not just a research paper sitting on arXiv anymore. It is an open model, an active GitHub project with 2026 updates, and a forecasting engine Google has wired into BigQuery and AlloyDB. (arxiv.org; github.com; cloud.google.com)
