Google Releases TimesFM-3 for Zero-Shot Multivariate Time-Series Forecasting
Google has open-sourced TimesFM-3, a 330-million-parameter foundation model pre-trained on over 1 trillion time points for zero-shot multivariate forecasting. It natively processes multiple coevolving targets and dynamic covariates in a single forward pass, topping public benchmarks like Gift-Eval and FEV-Bench.
Impact: Medium
Why it matters
You can now forecast complex multi-series datasets with future event covariates directly via open weights without training task-specific models.
TL;DR
- 01TimesFM-3 introduces native zero-shot multivariate forecasting for multiple targets and past-future covariates.
- 02Non-autoregressive decoding via Contiguous Patch Masking generates full horizons in a single pass without error compounding.
- 03The 330M-parameter model outputs point forecasts and 9 uncertainty quantiles directly out of the box.
Key facts
- Parameters
- 330 million
- Pre-training Dataset
- 1+ trillion time points
- Patch Size
- 32 time steps
- Probabilistic Output
- 9 quantiles (10th to 90th percentile)
- Benchmark Rank
- Rank 1 on Gift-Eval, FEV-Bench, and Time
Native Multivariate Architecture
Previous foundation models like TimesFM-2.5 were constrained to univariate series, ignoring cross-series correlations and external signals. TimesFM-3 scales to 330M parameters, pre-trained on a corpus of more than 1 trillion real-world and synthetic time points. It natively handles multiple target series, historical past covariates (such as foot traffic), and past-future dynamic covariates (such as scheduled marketing campaigns or weather forecasts) using a lookahead token concatenation strategy.
2D Attention and Non-Autoregressive Decoding
The model organizes tokenized 32-step patches into a 2D transformer grid:
- Causal Temporal Attention: Operates horizontally across time steps within each series to prevent future data leakage.
- Full Variate Attention: Operates vertically across distinct time series at each step to capture cross-series dependencies.
Instead of iterative autoregressive generation, TimesFM-3 applies Contiguous Patch Masking. Target and past-covariate series are masked over the forecasting horizon, while known future covariates remain visible. The model resolves the full horizon in a single forward pass, providing point predictions and 9 probability quantiles from the 10th to the 90th percentile.
Benchmark Performance and Availability
Evaluated on the Gift-Eval, FEV-Bench, and Time benchmarks, TimesFM-3 ranked first among pre-trained time-series foundation models in both point accuracy and probabilistic forecasting quality, outperforming Chronos-2, Toto 2.0, and TimesFM-2.5. The model is available immediately on GitHub and Hugging Face, with BigQuery integration rolling out via AI.FORECAST.
Try it in 2 minutes
git clone https://huggingface.co/google/timesfm-3
pip install timesfmbash
✓ When to use
- Forecasting multiple interconnected metrics with known future calendar or promotional covariates.
- Zero-shot forecasting pipelines where training dedicated models per time series is cost-prohibitive.
- Applications requiring probabilistic uncertainty bounds alongside point forecasts.
✕ When NOT to use
- Simple univariate tasks where basic statistical heuristics or lightweight ARIMA models deliver sufficient accuracy.
- Ultra-low-latency real-time stream processing with sub-millisecond execution constraints.
What to do today
- Clone the TimesFM-3 repository or load the model checkpoint from Hugging Face for multivariate testing.
- Format external future signals (promotions, holidays) into past-future dynamic covariates for non-autoregressive inference.
- Evaluate single-pass inference latency against existing autoregressive time-series pipelines.
Sources