Google today released the TimesFM-3 model, the third generation of its time series foundation model series. With native multivariate forecasting, dual attention, continuous patch masking, and other features, it can perform zero-shot multivariate time series forecasting.
IT Home learned from Google that TimesFM-3 has 330 million parameters and was pretrained on real-world and synthetic time series corpora containing more than 1 trillion time points. TimesFM-3 inherits the efficiency and zero-shot generalization capabilities of its predecessor while expanding zero-shot support for complex multivariate scenarios.
The model can also jointly forecast multiple time series that evolve together, capturing their dependencies to improve overall accuracy without fine-tuning for specific tasks. Its natively supported features are as follows:
Multi-target forecasting: Forecast multiple related time series simultaneously, such as jointly predicting ice cream sales for different brands. The model supports both point forecasts and quantile forecasts for all targets.
Past covariates: Incorporate features known only from a historical perspective, such as past foot traffic.
Past-future (dynamic) covariates: Use known future events to guide forecasts, such as planned promotions or weather forecasts.
Reference:
TimesFM-3: A zero-shot foundation model for multivariate forecasting