Google Releases TimesFM-3 for Multivariate Time Series Forecasting

Google Research has released its latest foundation model for multivariate time series forecasting, TimesFM-3. Unlike existing models that were limited to univariate time series forecasting, TimesFM-3 is designed to solve complex real-world forecasting problems where multiple time series and external variables interact.

Key Features and Architecture

TimesFM-3 possesses 330 billion parameters and was pre-trained on over 1 trillion real and synthetic time series data points. The model supports the following capabilities in a zero-shot manner without separate fine-tuning:

• Multi-target forecasting: Simultaneously forecasts multiple related time series, supporting both point and quantile forecasts.
• Past covariates: Reflects information known only in the past (e.g., past foot traffic).
• Past-future covariates: Improves forecast accuracy by utilizing events known in the future (e.g., promotion plans, weather forecasts).

In terms of architecture, it adopts the same decoder-only transformer structure as previous versions. It processes time series data by grouping them into patches of 32 steps and undergoes a normalization process that accounts for scale differences between each time series. Specifically for past-future covariates, it applies a strategy of constructing tokens through 'lookahead'.