Predictive planning across supply chains and financial operations has long relied on isolated time-series models that struggle to capture real-world complexity. While basic trend forecasting tracks how a single metric moved in the past, operational decisions depend on multiple interacting variables such as pricing shifts, promotional calendars, and foot traffic. On August 31, 2026, Google Research scientists Ayush Jain and Rajat Sen announced TimesFM-3, a 330-million-parameter foundation model engineered to deliver multivariate zero-shot time-series forecasting in a single forward pass.
Moving Past Univariate Limits
Since the original TimesFM debuted in 2024, foundation models for time series have seen steady adoption across retail, finance, observability, manufacturing, healthcare, and natural sciences. However, earlier iterations through TimesFM-2.5, released in September 2025, were strictly limited to univariate analysis—evaluating only the isolated trajectory of an individual metric. As Jain and Sen noted in their release announcement, enterprise forecasting is fundamentally multivariate, requiring systems to map correlations across auxiliary metrics and core targets simultaneously.
"Past sales alone rarely tell the full story. A good forecast should also draw on sales of related products (e.g., ice cream cones, syrups), historical foot traffic, and known future events like weather forecasts, promotions, and holidays."
To bridge this gap, Google Research trained TimesFM-3 on a synthetic and real-world corpus exceeding 1 trillion time points. The architecture natively ingests multiple targets, forecasting coevolving series simultaneously while outputting point and quantile forecasts. Crucially, the model categorizes auxiliary inputs into historical covariates—such as backward-looking foot traffic—and dynamic past-future covariates, like scheduled promotions or upcoming weather patterns.
Single-Pass Architecture and 2D Attention
Technically, TimesFM-3 deploys a decoder-only transformer setup, grouping continuous data points into patches of 32 time steps and normalizing each series to balance disparate scales. For past-future covariates, the architecture concatenates current and future patches using a lookahead strategy to incorporate scheduled events directly into inference. By eliminating the compute overhead and maintenance debt of training custom ML pipelines for individual SKUs or sensors, TimesFM-3 shifts enterprise forecasting from brittle bespoke models to zero-shot shared infrastructure.