A decoder-only foundation model for time-series forecasting
Not a pilot. Google just open-sourced TimesFM—a 200M-parameter foundation model that forecasts unseen time-series data with zero training, launching on Vertex AI this year.
Why it matters
Google Research introduced TimesFM, a production-ready foundation model for time-series forecasting that eliminates the need for domain-specific model training. This shifts how enterprises approach demand planning, financial forecasting, and operational analytics—moving from lengthy training cycles to immediate zero-shot deployment.
The key facts
8 to knowTimesFM trained on 100 billion real-world time-points
Model size: 200M parameters (orders of magnitude smaller than GPT-3.5)
Zero-shot performance matches or exceeds supervised models like PatchTST explicitly trained on target datasets
Tested on Monash Forecasting Archive: tens of thousands of time-series across traffic, weather, demand forecasting
Outperforms llmtime(ZS) prompt-based forecasting approach from GPT-3.5
Decoder-only transformer architecture with variable context and horizon lengths
Launching on Google Cloud Vertex AI later in 2024
Supports multiple temporal granularities from minutes to yearly data
Go to the source
Google Research Blogblog.research.google
Publisher excerpt: Posted by Rajat Sen and Yichen Zhou, Google Research Time-series forecasting is ubiquitous in various domains, such as retail, finance, manufacturing, healthcare and natural sciences. In retail use cases, for example, it has been observed that improving demand forecasting accuracy can meaningfully…