Google's TimesFM-3 Predicts Sales With Weather Data

Google Research put out a new forecasting model this week called TimesFM-3, built to predict what happens next in a stream of numbers like daily sales, using more than just the numbers themselves. The pitch is simple: real sales data doesn't move in isolation. A grocery chain trying to figure out how much ice cream it'll sell next week isn't just looking at last week's totals. It's weighing weather forecasts, discount schedules, holiday calendars, and even sales of related items like waffle cones.

TimesFM-3 is Google's attempt to bake all of that into one system. It's still built on a Transformer, the architecture behind its predecessors, but the way it handles data has changed. The model groups data points into chunks, scales wildly different measurements onto common ground, and scans information in two directions at once, tracking patterns over time within a single series and comparing relationships across multiple series simultaneously.

Trained on more than a trillion data points combining real and synthetic sources, it's built to work without extra training for new forecasting jobs. Here's how Google describes the problem it's trying to solve.