Google Releases TimesFM-3 for Multivariate Forecasting

Google released TimesFM-3, the third model in its time series foundation series. It has 330 million parameters and was trained on over 1 trillion time points from real and synthetic data. The model handles multivariate forecasting natively, with dual attention and continuous patch masking.

TimesFM-3 can predict multiple related time series together, like sales of different ice cream brands, without fine-tuning. It also uses past and future features, such as planned promotions or weather forecasts, to improve accuracy.

Building on previous versions, it offers strong zero-shot generalization, making it useful for finance, retail, and weather predictions.