● Covered by 3 sources · 3 reportsMedium impact2 neutral1 positive
Google BigQuery Introduces TimesFM 3.0 for Multivariate Time-Series Forecasting
Key points
- TimesFM 3.0 is a 330-million parameter foundation model.
- It enables multivariate time-series forecasting in BigQuery SQL.
- TimesFM 3.0 ranks #1 on GIFT-Eval, FEV-Bench, and TIME leaderboards.
- It also upgrades univariate forecasting and anomaly detection in BigQuery.
- Mistral released Timestral, its first family of time-series foundation models.
- Timestral models range from 5M to 209M parameters.
- The largest Timestral model has 209M parameters.
- Timestral supports past-only and known-future covariates.
- Timestral combines synthetic pre-training, covariate-aware attention, and long-context learning.
- Timestral offers fine-tuning options for domain adaptation via Mistral Forge.
- IBM introduced the Granite Timeseries Ensemble.
- The Granite Timeseries Ensemble combines four Granite time-series models.
- The ensemble provides median and quantile forecasts without retraining or fitted ensemble weights.
- The ensemble includes Probabilistic Ensemble and adaptive Inter Quartile Range (IQR) Ensemble.
Multivariate Forecasting Capability
Google has released TimesFM 3.0 in preview for BigQuery, introducing native multivariate time-series forecasting. Previous versions, TimesFM 2.0 and 2.5, were limited to univariate forecasting, predicting from a single time series. TimesFM 3.0 allows for predictions that consider multiple related targets and external drivers like weather or promotions.
Model Specifications and Performance
TimesFM 3.0 is a 330-million parameter foundation model developed by Google Research. It has achieved the top ranking across three major public time-series leaderboards: GIFT-Eval, FEV-Bench, and TIME. The model demonstrates high zero-shot accuracy on both covariate and multivariate forecasting tasks.
Integration with BigQuery AI Functions
The new model integrates directly into BigQuery SQL via AI.FORECAST for multivariate forecasting. It also serves as a drop-in upgrade for existing univariate forecasting workflows and anomaly detection through AI.DETECT_ANOMALIES. Users can specify 'TimesFM 3.0' to immediately benefit from improved accuracy.
Example Workflow
An example provided forecasts daily commuter and leisure bike rides using the NYC Citi Bike public dataset. BigQuery uses a single-table input format for multivariate forecasting, combining historical observation rows with fully populated target metrics and covariates, and future horizon rows populated with known future covariates while target columns are set to NULL.
Updates
- IBM introduced the Granite Timeseries Ensemble.
- The Granite Timeseries Ensemble combines four Granite time-series models.
- The ensemble provides median and quantile forecasts without retraining or fitted ensemble weights.
- The ensemble includes Probabilistic Ensemble and adaptive Inter Quartile Range (IQR) Ensemble.
- Mistral released Timestral, its first family of time-series foundation models.
- Timestral models range from 5M to 209M parameters.
- The largest Timestral model has 209M parameters.
- Timestral supports past-only and known-future covariates.
- Timestral combines synthetic pre-training, covariate-aware attention, and long-context learning.
- Timestral offers fine-tuning options for domain adaptation via Mistral Forge.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Primary sources
GitHub mistralai/timestral GitHub ibm-granite/granite-tsfm arXiv 2602.06909 arXiv 2508.05287 arXiv 2401.03955 arXiv 2205.04216How outlets covered it
IBM introduced the Granite Timeseries Ensemble, a new method combining four Granite time-series models to produce more robust forecasts. This ensemble approach reduces the need for users to select a single best model, offering median and quantile forecasts without retraining or fitted ensemble weights.
Mistral has released Timestral, its first family of time-series foundation models, which includes models up to 209M parameters. The largest Timestral model outperforms previous state-of-the-art models on key benchmarks while using fewer parameters, offering probabilistic, multivariate forecasts and supporting past and future covariates.
Google has launched TimesFM 3.0 in preview for BigQuery, enabling native multivariate time-series forecasting directly within SQL. This update allows users to predict future values based on multiple related time series and external factors, addressing a limitation of previous univariate-only models.