What to know

• TimesFM-3 is Google's latest foundation model advancing from univariate to multivariate time series forecasting, trained on over 1 trillion time points with 330 million parameters.
• The model uses decoder-only transformer architecture with causal temporal attention and full variate attention to capture dependencies across multiple related time series without task-specific fine-tuning.
• It supports multiple forecasting targets simultaneously, past-only features, and future-known covariates (e.g., weather forecasts, promotions), addressing real-world scenarios in retail, finance, healthcare, and other domains.
• TimesFM-3 demonstrates zero-shot generalization, extending the efficiency of its predecessors (TimesFM and TimesFM-2.5) to complex multivariate scenarios.

“TimesFM-3 adds robust support for complex multivariate scenarios in a zero-shot manner. It can jointly predict multiple coevolving time series, capturing dependencies that improve overall accuracy without requiring task-specific fine-tuning.”

Google Research, Model developer · Google Research Blog · Sep 1, 1:25 AM

Google Research Model developer