Google Unveils 330 Million-Parameter TimesFM-3, Topping 3 Forecasting Benchmarks
Updated
Updated · Google Research · Aug 31
Google Unveils 330 Million-Parameter TimesFM-3, Topping 3 Forecasting Benchmarks
1 articles · Updated · Google Research · Aug 31
Summary
TimesFM-3 is Google Research’s new time-series foundation model for multivariate forecasting, built to predict multiple related series and known future signals in one pass rather than patch by patch.
330 million parameters and more than 1 trillion pretraining time points underpin the model, which extends the TimesFM line beyond the univariate-only limits of TimesFM-2.5 released in 2025.
Three public benchmarks—Gift-Eval, FEV-Bench and Time—ranked TimesFM-3 first on both point and probabilistic forecasting metrics among pre-trained foundation models, with multivariate mode improving further over its own univariate mode.
9 quantiles per target series give users probabilistic forecasts, and Google says the model can capture effects such as a roughly 20% sales bump on promotion days when future covariates are supplied.
TimesFM-3 is available now on GitHub and Hugging Face, while BigQuery integration is scheduled in the coming weeks.
Despite topping benchmarks, can Google's new TimesFM-3 truly outsmart traditional financial models, or is its performance exaggerated by sheer scale?
How does generating entire forecast horizons in a single pass fundamentally change the way we anticipate complex real-world market shifts?
Will the ability to predict demand using external signals like weather revolutionize retail, or will unprecedented events still shatter these forecasts?