Google open‑sources TimesFM

Most forecasting inside companies still works like a custom suit: a team picks one dataset, trains one model, tunes it by hand, and starts over when the pattern changes. Google’s TimesFM tries to work more like autocomplete for numbers: feed it a history of values, and it predicts what likely comes next without training a bespoke model first. (research.google) A time series is just a number measured over time: daily sales, hourly electricity use, weekly hotel bookings. Forecasting means learning the shape of that sequence well enough to guess the next few points, the way you can often continue a rhythm after hearing the first bars. (research.google) Google trained TimesFM on a corpus of 100 billion real-world time points, spread across different domains and time scales. The idea is the same one behind language models: instead of learning one company’s history from scratch, the model learns broad patterns first and then applies them to a new series. (research.google) The architecture is decoder-only, which means it predicts the next values from the values it has already seen, one step after another. In plain English, it reads the left side of a chart and keeps extending the line to the right. (research.google) Google Research released TimesFM as open source on GitHub, where the repository now lists TimesFM 2.5 as the latest model version and links to public checkpoints. The repository also says the model was presented in the paper “A decoder-only foundation model for time-series forecasting,” at the International Conference on Machine Learning in 2024. (github.com) The second part of the news is where this stops being just a research release. BigQuery, Google Cloud’s data warehouse, now includes a built-in TimesFM model, so analysts can run forecasts with the `AI.FORECAST` function directly in Structured Query Language instead of exporting data into a separate machine-learning stack. (cloud.google.com 1) (cloud.google.com 2) Google’s documentation says `AI.FORECAST` uses BigQuery Machine Learning’s built-in TimesFM model and avoids the need to create and train your own forecasting model. That changes the workflow from “build infrastructure, train, deploy, monitor” to “write a query against the table you already use for reporting.” (cloud.google.com) BigQuery’s docs also describe the built-in model as univariate, which means it forecasts one value column from its own history rather than from a bundle of outside drivers. If you want a forecast for daily units sold, you give it the units-sold series and timestamps, not a giant feature-engineering project with weather, promotions, and holidays on day one. (cloud.google.com) Google’s tutorial shows the same model forecasting San Francisco bike-share trips from a public BigQuery table, including multiple time series in one workflow. That is the practical pitch: a retailer, airline, or manufacturer can keep the data in BigQuery and plug forecasts into the same pipeline that already feeds dashboards, pricing, staffing, or inventory decisions. (cloud.google.com) Google has also wired the model into `AI.DETECT_ANOMALIES`, which flags unusual spikes or drops in time-series data. So the same foundation model can be used for two adjacent jobs inside operations teams: predicting next week’s baseline and spotting when today’s number suddenly breaks it. (cloud.google.com) There are still limits. The open-source repository says the public package is “not an officially supported Google product,” while the BigQuery version is the supported product path, and BigQuery’s docs frame TimesFM as a preview feature inside BigQuery Machine Learning. (github.com) (cloud.google.com) So the real shift is not that forecasting became solved in April 2026. It is that Google moved a research model for time-series prediction into the place where many companies already store sales, demand, and supply data, which makes “run a forecast” look a lot more like “run a query.” (cloud.google.com 1) (cloud.google.com 2)