AI & RoboticsMore in AI & Robotics→
Google's TimesFM-3 Forecasts the Future From Sales, Weather and Promotions
Key Takeaways
- Google Research released TimesFM-3, a Transformer-based forecasting model that combines time series with related data and known future events.
- The model groups 32 consecutive data points into patches and normalizes each series to a common scale so values of different magnitudes can be compared.
- It processes data along two alternating directions: within a single series using only past values, and across series to learn relationships between variables.
- Three types of supplementary input are supported: jointly predicted related variables, historically known factors, and known future events such as discounts or weather forecasts.
DE
DT Editorial Team··via the-decoder.com