Google TimesFM-3: AI Forecasting From Sales, Weather, and Promotions

Google Research has released TimesFM-3, an AI model built to forecast time-series data such as daily sales figures. The model can use related variables and known upcoming events, including weather forecasts, discount campaigns, and holidays, instead of relying on a single historical metric. For businesses, the key takeaway is that forecasts can become more context-aware when planned events and connected demand signals are included.

TimesFM-3 is based on a Transformer architecture and groups 32 consecutive data points into one patch. It normalizes different series to a shared scale, allowing data with very different magnitudes to be compared more directly. The model looks along the time axis for patterns in a single series and across series to learn relationships between variables at the same point in time.

Earlier TimesFM versions predicted future blocks step by step, which Google says was slower, more compute-heavy, and more exposed to compounding errors. TimesFM-3 instead marks all future time steps as blanks and fills them in a single pass. In Google’s ice cream example, a model with access to the discount schedule expects roughly 20 percent more units on promotion days, while a model using only past sales continues the usual weekly pattern.

According to Google, TimesFM-3 ranks first among pretrained forecasting models on Gift-Eval, FEV-Bench, and Time for both point accuracy and uncertainty calibration. The model has 330 million parameters, was trained on real and synthetic time series totaling more than one trillion data points, and works zero-shot without extra training for new tasks. TimesFM-3 is available on GitHub and Hugging Face, with Google planning to add it to BigQuery in the coming weeks.