Google's TimesFM-3 tops forecasting benchmarks, adds multivariate support

Google's TimesFM family has been deployed in retail, finance, manufacturing, healthcare, and the sciences since it launched in 2024, but until now every version processed a single data series at a time. TimesFM-3 changes that by grouping 32 consecutive data points into one patch and normalizing each series so measurements of very different magnitudes can be compared directly. It then processes data in two alternating directions: along the time axis to find patterns within a series, and across series to learn how variables relate at a given moment, which is how it can pick up on things like a discount on one product affecting sales of another. That design lets it use three types of supplementary data, including related variables, historical factors such as past foot traffic, and known future events like planned discounts or weather forecasts.

The shift to one-shot forecasting is the other key change. Earlier versions predicted the future one block at a time, which Google says was slow, compute-heavy, and let errors compound. TimesFM-3 instead marks all future time steps as blanks and fills them in a single pass. Google's ice cream example shows the difference: a model that only knows past sales just continues the usual weekly pattern, while TimesFM-3, given a discount schedule, learns from history how much promotions boost demand and expects roughly 20 percent more units on each promotion day.

Whether that performance translates outside benchmarks is likely to depend on how smoothly the BigQuery integration lands in the coming weeks, since that is where many business forecasters would encounter the model. The competitive picture is also moving: Amazon's Chronos-2 comes close to TimesFM-3's single-variable mode but falls well behind the full version, and that gap is likely to shape how quickly rivals respond.