TimesFM-3: A zero-shot foundation model for multivariate forecasting

Details

TimesFM-3 is Google Research's latest time series foundation model, and it marks a substantial capability jump: all prior TimesFM versions (through 2.5, released September 2025) were strictly univariate — predicting a single series from its own history. TimesFM-3, pre-trained on over 1 trillion time points, now forecasts multiple correlated series simultaneously, incorporates historically-observed covariates (e.g., past foot traffic), and leverages known future events like planned promotions or weather forecasts.\n\nThe model carries 330 million parameters and uses the same decoder-only transformer architecture as its predecessors, with input patches of 32 time steps. The key change is in token construction: multiple target series, past covariates, and past-future covariates are each handled natively, without needing task-specific fine-tuning per dataset. Both point and quantile forecasts are supported for all targets simultaneously.\n\nGoogle claims TimesFM-3 "significantly outperforms" other forecasting models across major benchmarks, though the blog post does not name specific baselines or cite numeric margins — practitioners will want to check the GitHub and HuggingFace releases for evaluation details.\n\nReal-world applicability spans retail demand planning, financial modeling, infrastructure observability, healthcare, and natural sciences. For teams already running TimesFM-2.5 in production, the upgrade eliminates expensive per-dataset fine-tuning while unlocking cross-series dependency modeling.