Because the debut of TimesFM in 2024, we’ve seen the adoption of time-series basis fashions for real-world time-series forecasting duties throughout a number of domains, corresponding to retail, finance, observability, manufacturing, healthcare and pure sciences.
Up till TimesFM-2.5 (launched in September 2025), our fashions had been strictly restricted to univariate forecasting: forecasting utilizing solely the historical past of a single time sequence. But, most real-world forecasting issues are inherently multivariate: the place a number of time sequence and auxiliary exterior options collectively impression the long run forecast of a time sequence. Think about forecasting ice cream gross sales for a retail chain. Previous gross sales alone not often inform the total story. forecast also needs to draw on gross sales of associated merchandise (e.g., ice cream cones, syrups), historic foot visitors, and recognized future occasions like climate forecasts, promotions, and holidays.
Right now we introduce TimesFM-3, the following technology of our time-series basis mannequin that’s natively pre-trained for multivariate forecasting. TimesFM-3 has 330 million parameters and is pre-trained on a real-world and artificial time-series corpus comprising greater than 1 trillion time factors. Constructing on the effectivity and zero-shot generalization of its predecessors, TimesFM-3 provides sturdy help for complicated multivariate eventualities in a zero-shot method. It may collectively predict a number of coevolving time sequence, capturing dependencies that enhance general accuracy with out requiring task-specific fine-tuning. The mannequin natively helps:
- A number of targets: Forecast a number of associated time sequence concurrently (e.g., collectively forecasting completely different manufacturers of ice cream). The mannequin helps each level and quantile forecasts for all targets.
- Previous covariates: Incorporate options which can be solely recognized traditionally (e.g., previous foot visitors).
- Previous-future (dynamic) covariates: Leverage recognized future occasions to information the forecast (e.g., deliberate promotional campaigns or climate forecasts).
