Google introduces TimesFM-3 foundation model for multivariate forecasting

Summary

Google Research has introduced TimesFM-3. This zero-shot foundation model, with 330 million parameters, extends time-series forecasting from single-series to multivariate forecasting. The model supports multiple targets, past covariates, and known future signals in a single non-autoregressive forward pass.

Why it matters

Practical forecasting problems, such as retail demand, energy load, and clinical monitoring, are inherently multivariate. However, previous time-series foundation models were locked into univariate inputs. TimesFM-3 lowers the barrier for teams that lack the data or expertise to build custom multivariate pipelines by handling inter-series dependencies and external covariates without task-specific fine-tuning. Additionally, its single-pass, non-autoregressive decoding reduces inference latency and error accumulation compared to the iterative decoding in previous TimesFM versions.

Highlights

• A 330-million-parameter, decoder-only transformer pre-trained on over 1 trillion real and synthetic time points
• A 2D attention mechanism to capture inter-series correlations: causal temporal (horizontal) + full variable (vertical)
• Non-autoregressive decoding via Contiguous Patch Masking: the entire horizon is produced in a single forward pass, providing 9 quantile outputs for each target

Critical take

The article relies on the claim of 'significantly superior performance compared to other forecasting models in key benchmarks.' However, no specific benchmark is named, and no numerical data is reported, making it difficult to verify the claim from the text alone. Furthermore, describing forward-looking token construction as a 'clever strategy' reads more like marketing language than technical justification. The blog post cuts off mid-sentence, resulting in the omission of details regarding quantile evaluation and ablation studies.