TimesFM-3: A zero-shot foundation model for multivariate forecasting
Summary
Google Research scientists Ayush Jain and Rajat Sen introduce TimesFM-3, a 330-million-parameter time-series foundation model pretrained natively for multivariate forecasting on more than a trillion time points.
Univariate history is not enough.
Through 2.5, TimesFM forecast from a single series' history; TimesFM-3 is pretrained natively to jointly forecast related targets with past and known-future covariates.
Attention is a 2D grid.
Causal temporal attention stays within one series; full variate attention at a timestep sees every other series so promotions in one can move sales in another.
These two attention mechanisms alternate for several layers, seamlessly blending temporal patterns with cross-series relationships.
The whole horizon lands in one pass.
Contiguous Patch Masking fills masked future patches at once, avoiding autoregressive error and latency, and emits nine quantiles per target step.
Even univariate mode already leads.
On Gift-Eval, FEV-Bench, and Time it is top-ranked among pretrained foundation models for point and probabilistic metrics; multivariate mode is a further leap.
TimesFM-3 is the top-ranked model in terms of both point and probabilistic forecasting metrics among all pre-trained foundation models.
Weights now, BigQuery later.
TimesFM-3 is on GitHub and Hugging Face; BigQuery AI.FORECAST still points at TimesFM-2.5 until the new integration lands.
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.