TimesFM-3: A zero-shot foundation model for multivariate forecasting

Open original

Google released TimesFM-3, a 330-million-parameter model that forecasts multiple related time series simultaneously by processing them in a single forward pass. It outperforms competing foundation models on standard benchmarks by learning relationships between different series and incorporating known future events like promotions or weather.

  • TimesFM-3 uses a 2D attention mechanism (causal temporal + full variate) that lets the model see patterns across time within a series and correlations between different series at the same time step, enabling it to capture how one series influences another.
  • The model generates entire forecast horizons in one pass using "Contiguous Patch Masking" instead of iteratively predicting one step at a time, reducing latency and error accumulation.
  • It ranked first on three major benchmarks (Gift-Eval, FEV-Bench, Time) for both point and probabilistic forecasting accuracy against recent competitors like Chronos-2 and Toto 2.0.

Google released TimesFM-3, a 330-million-parameter time series forecasting model trained on over 1 trillion data points that handles multivariate forecasting—predicting multiple related time series simultaneously while incorporating external information. Unlike its predecessors, which only worked with single time series in isolation, TimesFM-3 natively supports multiple targets, past covariates (historical features), and dynamic covariates (known future events like promotions or weather forecasts). The model uses a decoder-only transformer that groups time steps into 32-point patches, then applies two types of attention: causal temporal attention that looks backward within a single series to prevent data leakage, and full variate attention that lets the model see all other series at each time step to learn cross-series correlations. The key technical advance is Contiguous Patch Masking, which generates the entire forecast horizon in one forward pass instead of iteratively predicting one patch at a time. This eliminates latency, reduces error accumulation, and cuts computation. For future events that are known (like scheduled promotions), the model can see those signals; for unknown future values, it masks them. The architecture outputs nine quantiles per target at every time step, giving a probabilistic view of forecast uncertainty rather than just point estimates. In a retail ice cream example, TimesFM-3 correctly anticipated a 20% sales bump on promotion days by learning the relationship between promotions and sales from historical data, while a standard univariate model couldn't account for future promotions at all. On three major benchmarks—Gift-Eval, FEV-Bench, and Time—TimesFM-3 ranked first among all pre-trained foundation models for both point forecasts and probabilistic accuracy. Even in univariate mode (when evaluated without any covariate information), it matched or beat competitors like Chronos-2 and Toto 2.0. The full multivariate mode pushed it further ahead. The model is available on GitHub and Hugging Face now, with BigQuery integration coming in the coming weeks.

Ask about this article