Google's new AI model predicts the future from sales data, weather, and discount schedules
decoder ·
Google Research has released TimesFM-3, an AI forecasting model that predicts future values from time series such as daily sales while using related data and known upcoming events. It is built on a Transformer, like its predecessors, but groups 32 consecutive data points into one patch and normalizes each series to a common scale. The model processes data in two alternating directions: along the time axis it uses only past values to find patterns within a series, and across series it compares variables at a given time to learn relationships, such as how a discount on one product affects another. TimesFM-3 has 330 million parameters and was trained on over one trillion real and synthetic time-series data points, and it works zero-shot without extra training. It handles supplementary data in three forms: predictions for multiple related variables, past-only factors such as historical foot traffic, and known future events like planned discounts or weather forecasts, and it outputs nine values per time step to capture uncertainty. Earlier versions forecast one block at a time, which Google says was slow, compute-heavy, and prone to compounding errors; TimesFM-3 instead marks all future time steps as blanks and fills them in a single pass. In an ice cream example, giving the model a discount schedule leads it to expect about 20 percent more units on promotion days, while a model without that information just continues the usual weekly pattern. Google says TimesFM-3 ranks first among pretrained forecasting models on Gift-Eval, FEV-Bench, and Time for both point accuracy and uncertainty calibration, ahead of Amazon's Chronos-2, the Toto-2.0 family, and TimesFM-2.5; it is available on GitHub and Hugging Face, with BigQuery support planned, and the TimesFM family has been deployed in retail, finance, manufacturing, healthcare, and the sciences since 2024, with TimesFM-3 adding multivariate support that earlier versions lacked.
decoder