Teams in retail, finance, manufacturing, healthcare, and the sciences can now forecast demand using the full context around their numbers, including weather, discount schedules, and the sales of related products. Google TimesFM-3, released by Google Research on September 12, 2026, reads historical data alongside known upcoming events to sharpen each prediction. It ranks first among pretrained forecasting models on three public benchmarks in both point accuracy and uncertainty calibration.<\/p>\n
Real-world forecasts rarely depend on a single number. Google illustrates the model with a retail chain predicting ice cream sales. A useful forecast factors in related products such as waffle cones and syrup, along with past foot traffic, weather, discount campaigns, and holidays. TimesFM-3 handles three types of supplementary data at once. It predicts multiple related variables together, such as different ice cream flavors. It incorporates factors known only for the past, such as historical foot traffic. It also uses known future events, such as planned discounts and weather forecasts.<\/p>\n
Rather than a single point estimate, the model outputs nine values per time step so a forecast carries a range and a measure of its own uncertainty.<\/p>\n
TimesFM-3 is built on a Transformer, the same base architecture as earlier versions in the family. It groups 32 consecutive data points into a single patch and normalizes each series to a common scale, so measurements of very different magnitudes can be compared directly.<\/p>\n
The model processes data in two alternating directions. Along the time axis, it looks for patterns within a single series and draws only on past values, which prevents future information from leaking into the forecast. Across series, it compares all variables at a given point in time and learns how they relate, so it can pick up on how a discount on one product moves sales of another.<\/p>\n
The model has 330 million parameters and was trained on real and synthetic time series totaling more than one trillion data points, according to Google. Like its predecessors, it works zero-shot and needs no extra training for a new task.<\/p>\n
Earlier versions predicted the future one block at a time. Google says that approach was slow, compute-heavy, and allowed errors to compound as each prediction built on the last. TimesFM-3 marks all future time steps as blanks and fills them in a single pass.<\/p>\n
The ice cream example shows the gain. A model that knows only past sales continues the usual weekly pattern and stays blind to planned promotions. When TimesFM-3 receives the discount schedule, it learns from history how much promotions lift demand and expects roughly 20 percent more units on each promotion day.<\/p>\n
On Gift-Eval, FEV-Bench, and Time, TimesFM-3 ranks first among all pretrained forecasting models in both point accuracy and uncertainty calibration, according to Google. The field it is measured against includes Amazon’s Chronos-2, the Toto-2.0 family, and Google’s own TimesFM-2.5. Even when limited to a single variable, TimesFM-3 matches or beats the field, and adding more data widens the gap. On Gift-Eval it leads by a wide margin even in single-variable mode. Chronos-2 comes close to that single-variable mode on FEV-Bench but falls well behind the full multivariate version. On the Time benchmark the Toto-2.0 family follows in second place.<\/p>\n
TimesFM-3 is available on GitHub and Hugging Face, and Google plans to add it to BigQuery in the coming weeks. TimesFM-2.5 currently handles single-variable forecasting in BigQuery through the AI.FORECAST command. Since the family launched in 2024, Google says it has been deployed across retail, finance, manufacturing, healthcare, and the sciences. Every version through TimesFM-2.5, released in September 2025, could process only one data series at a time, which makes multivariate support the main advance in TimesFM-3.<\/p>\n
Google is also building forecasting models beyond time series. In early August, Google DeepMind released WeatherNext Cyclones, an open-source system for tropical cyclones that predicts storm tracks and intensity about a day further out than leading operational models.<\/p>\n
TimesFM-3 is a time series forecasting model from Google Research that predicts future values, such as daily sales, from past data. It draws on related products, factors known only for the past like historical foot traffic, and known future events like planned discounts and weather forecasts. It has 330 million parameters, was trained on more than one trillion data points, and works zero-shot.<\/p>\n
Every version through TimesFM-2.5 could process only one data series at a time. TimesFM-3 adds multivariate support, so it can forecast several related variables together and use supplementary past and future data. It also replaces block-by-block prediction with a one-shot method that marks future steps as blanks and fills them in a single pass, which Google says reduces the compounding errors of the older approach.<\/p>\n
TimesFM-3 is available on GitHub and Hugging Face, and Google plans to add it to BigQuery in the coming weeks. In BigQuery, TimesFM-2.5 currently handles single-variable forecasting through the AI.FORECAST command.<\/p>\n