Google Unveils TimesFM-3 AI Model for Multivariate Forecasting

Google's latest AI model, TimesFM-3, has been released and it can predict future events based on sales data, weather, and discount schedules. Unlike its predecessors, this model processes data in two alternating directions: along the time axis and across series.

This approach allows it to capture patterns within a single series while also comparing all variables at a given point in time. TimesFM-3 has 330 million parameters and was trained on over one trillion data points, making it more accurate than its predecessors.

The model is available on GitHub and Hugging Face, and Google plans to add it to BigQuery soon. This new version can handle multiple related variables at once and incorporates factors known only for the past, such as historical foot traffic.