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

Google Research · 1 month ago · 22 ● 3 sources

TimesFM introduced TimesFM-3, a zero-shot time-series foundation model designed for multivariate forecasting in a single forward pass. TimesFM-3 has 330 million parameters and is pre-trained on more than 1 trillion time points. The model moves from univariate-only inputs to native multivariate targets with past and known future covariates, and it outputs 9 quantiles per horizon step for probabilistic forecasts.