Google Research has announced the release of TimesFM-3, the third generation of its zero-shot time series foundation model. The new model, which features 330 million parameters, introduces native multivariate forecasting capabilities, marking a significant technical departure from its predecessor, TimesFM-2.5.
Technical Architecture and Capabilities
According to documentation from Google Research, the primary advancement in TimesFM-3 is its ability to process multivariate inputs. While previous versions were limited to the history of a single series, TimesFM-3 can handle multiple simultaneous targets. The model supports past-only covariates as well as past-future covariates—such as scheduled promotions or holidays where future values are known—without requiring fine-tuning.
The architecture is a decoder-only transformer consisting of 20 layers with a model dimension of 1280 and 16 heads. It utilizes a patching mechanism that groups 32 contiguous time steps per token. The model employs an alternating attention mechanism: causal attention is applied across time within a single series, while full attention is applied across different series at a given time step.
Unlike many traditional models that generate forecasts autoregressively, TimesFM-3 produces forecasts in a single forward pass. It achieves this by appending masked placeholder tokens for the entire forecast horizon and filling them simultaneously, while leaving known past-future covariates unmasked. The model provides nine quantiles, ranging from the 10th to the 90th percentile, per target for each step of the horizon.
Training Data and Benchmarks
The model was trained on a dataset totaling over 1 trillion time points. This training set includes Wikipedia pageviews through November 2023, Google Trends queries through the end of 2022, synthetic data, and GiftEvalPretrain (excluding overlaps with FEV-bench).
Google reports that TimesFM-3 achieved the best average rank on the Gift-Eval, FEV-Bench, and Time benchmarks when compared against other models, including Chronos-2, Toto 2.0, and TimesFM-2.5. The company claims that the model’s univariate-only mode matches or exceeds these baselines even before covariates are integrated.
Availability and Licensing
The PyTorch weights for TimesFM-3 are currently available on Hugging Face and GitHub. However, users should note that the weights are released under the TimesFM Non-Commercial License v1.0, meaning the model is not intended for immediate production use in commercial environments. Google has indicated that integration with BigQuery is planned for a future release.
While the technical release has generated interest in developer communities, some users on platforms like Reddit have raised questions regarding the specific composition of the synthetic data used in training and the breadth of the datasets, noting a lack of explicit mention of weather, finance, or macroeconomic data.