Google's new AI model predicts the future from sales data, weather, and discount schedules

Details

Google Research has released TimesFM-3, the third version of its time series forecasting model. The 330-million-parameter Transformer was trained on over one trillion data points — real and synthetic time series — and operates zero-shot with no task-specific fine-tuning required.\n\nTimesFM-3 addresses a structural flaw in previous versions: predicting the future one block at a time lets errors compound. The new model marks all future time steps as blanks and fills them in a single forward pass, reducing compute overhead and eliminating error cascades. The architecture groups 32 consecutive data points into a single patch and normalizes each series so measurements of very different magnitudes can be compared directly.\n\nThe model handles three types of supplementary inputs: multiple related time series predicted simultaneously, historical covariates known only for the past (such as foot traffic data), and known future events like scheduled promotions or weather forecasts. It outputs nine values per time step rather than a single estimate, capturing prediction uncertainty across the range.\n\nA retail ice cream example demonstrates the impact: a standard model continues weekly sales patterns blind to planned promotions, while TimesFM-3 correctly predicts roughly 20% higher sales on each promotion day by learning from historical discount effects. Enterprise use cases include demand forecasting, logistics planning, and financial modeling where structured event calendars and co-variates are available.