Google Releases TimesFM-3, a Multivariate Time-Series Model

Google Research has released TimesFM-3, a time-series foundation model that integrates multivariate prediction and covariate handling into its model architecture. With 330 million parameters, it is pre-trained on over 1 trillion time points, enabling zero-shot prediction without fine-tuning.

Key Technologies and Architecture

To address the univariate limitations of existing models and the latency issues of autoregressive approaches, the model introduces non-autoregressive decoding based on Contiguous Patch Masking (CPM). Target series, past covariates, and past-future covariates are arranged in a 2D token grid, and the prediction horizon is pre-set with mask tokens to generate all predictions simultaneously in a single forward pass. This mitigates error accumulation and latency issues that arise as the prediction horizon lengthens.

Performance and Benchmarks

Across three major benchmarks—GIFT-Eval, fev-bench, and TIME—the model achieved the average rank of 1st among pre-trained foundation models for both point and probabilistic forecasting metrics. Notably, in multivariate mode, it demonstrated superior performance compared to competing models such as Chronos-2 and Toto 2.0, while even in univariate mode it outperforms existing state-of-the-art models. For probabilistic forecasting, it directly outputs 9 quantiles to quantify prediction uncertainty.

License and Practical Application Considerations

Model weights are available for download on Hugging Face; however, the TimesFM-3.0 pre-trained weights are licensed under a non-commercial license, restricting use in commercial or production environments. Additionally, due to the nature of multivariate attention, there is an upper limit (32) on the number of series that can be processed in a single forward pass. If this limit is exceeded, automatic splitting occurs, so critical covariates must be selected to fit within the input budget.