Google's TimesFM 3 forecasts multiple series at once

Google Research released TimesFM-3 on August 31, the third version of its open time-series forecasting model and the first to natively handle multiple related series at once rather than forecasting each one alone. The 330-million-parameter model was pretrained on more than a trillion time points of real and synthetic data, and can take past covariates, like historical foot traffic, and events already known ahead of time, like a scheduled promotion or a weather forecast, into account when producing a forecast, generating the entire output in one forward pass instead of predicting step by step.

Google reports TimesFM-3 ranks first on all three public forecasting benchmarks it was tested against: GIFT-Eval, FEV-Bench, and the Time leaderboard, for both point forecasts and the nine-quantile probabilistic ranges it also outputs. It beats specialized univariate models even when run in univariate mode. The model’s code is open on GitHub and its weights are hosted on Hugging Face under a non-commercial license, a change from earlier TimesFM releases; a managed BigQuery integration is expected within weeks.

For operators, multivariate forecasting has typically meant either building bespoke models per business or accepting the accuracy loss of treating correlated series independently. A zero-shot model that handles both cases from one checkpoint, and that ranks first on the field’s standard benchmarks, is worth a direct comparison against whatever demand, inventory, or capacity forecasting you’re running today.