Google’s new forecasting model beats everyone. You can’t use it at work (yet).

On Monday, Google unveiled TimesFM-3, a sophisticated time-series forecasting model consisting of 330 million parameters and trained on more than a trillion data points. Currently accessible on Hugging Face under a non-commercial license, the model aims to predict data trends similarly to how large language models predict the next word.

With the increasing complexity of real-world forecasting, which often involves multiple interconnected time series and auxiliary data, Google's TimesFM-3 stands out by being pre-trained to handle these intricate connections without requiring additional data. The model's architecture resembles that of prior transformer-based models, with a key distinction in its dual attention layers.

The first layer processes time series data sequentially, ensuring causality, while the second analyzes cross-series relationships, enabling the model to account for external factors such as promotional activities or weather conditions. Unlike previous versions that generated forecasts sequentially, TimesFM-3 executes a single forward pass, appending masked tokens for the entire forecast horizon and infilling them simultaneously.

Currently, TimesFM-3's weights are restricted to non-commercial, non-production use under a separate license, but Google anticipates transitioning to a commercial model in the near future.