Google AI Releases TimesFM-3: A 330M Parameter Zero-Shot Foundation Model For Multivariate Time Series Forecasting

Google Research released TimesFM-3, a foundation model with 330 million parameters designed to forecast multiple related time series simultaneously, addressing a limitation of earlier versions that could only forecast one series at a time. Previous TimesFM versions were univariate, meaning they forecasted single series from their own history alone, but most real-world forecasting problems involve multiple interconnected factors, such as ice cream sales being influenced by related product sales, foot traffic, weather, promotions, and holidays. TimesFM-3 was pretrained on more than 1 trillion time points and can handle multiple targets, past covariates (historical data), and past-future covariates (known future events) without requiring task-specific fine-tuning, producing the entire forecast in a single forward pass. The model ranks first among pretrained foundation models on three major benchmarks, though its weights are restricted to non-commercial and non-production use only.