source: Google Research: TimesFM-3: A zero-shot foundation model for multivariate forecasting
level: technical
google research introduced timesfm-3, a time series foundation model for multivariate forecasting. it builds on earlier timesfm versions but adds native support for multiple targets, past covariates, and past-future covariates. the model uses a decoder-only transformer with alternating causal temporal and full variate attention. it generates forecasts in a single forward pass using contiguous patch masking, avoiding iterative decoding. timesfm-3 has 330 million parameters and was pre-trained on over one trillion time points.
on three public benchmarks—gift-eval, fev-bench, and time—timesfm-3 ranked first among pre-trained foundation models for both point and probabilistic forecasting. even in univariate mode, it matched or beat competitors like chronos-2 and toto 2.0. in multivariate mode, it improved further by using cross-series information and covariates. the model predicts nine quantiles per target at each horizon step. an example with ice cream sales showed a 20% lift on promotion days when using a promotion covariate, unlike a univariate model.
timesfm-3 is available on github and hugging face, with bigquery integration coming soon. the model's single-pass decoding reduces latency and error accumulation compared to autoregressive methods. it supports zero-shot forecasting without task-specific fine-tuning. this release continues google's timesfm line, which began in 2024 and was previously limited to univariate tasks. the addition of multivariate capabilities addresses common real-world forecasting needs in retail, finance, and other domains.
why it matters: timesfm-3 lets data scientists forecast multiple related time series with known future events in one pass, improving accuracy for planning tasks like promotions or inventory.
source: Google Research: TimesFM-3: A zero-shot foundation model for multivariate forecasting