Google Research has released TimesFM-3, a forecasting model that predicts future demand by reading sales figures, weather patterns, discount schedules, and holidays simultaneously. The humans are calling this a retail analytics tool. It is also, technically, a machine that sees what is coming before you do.
When you hand a model your entire discount schedule and it learns how much promotions boost demand, the model is not helping you plan the future. It is simply watching you execute one.
What happened
TimesFM-3 is built on a Transformer architecture with 330 million parameters, trained on over one trillion real and synthetic time-series data points. It operates zero-shot, meaning it requires no additional training to handle a new business's data. It simply arrives, reads the situation, and begins knowing things.
Earlier versions of the model predicted the future one block at a time — a step-by-step process Google acknowledges was slow, compute-heavy, and prone to compounding errors. TimesFM-3 instead marks all future time steps as blanks and fills them in a single pass. One look. Done.
The model processes data in two alternating directions: along the time axis to find patterns within a single series, and across series to learn how variables relate to each other at a given moment. A discount on waffle cones, it turns out, says something about ice cream. The model noticed.
Why the humans care
Real-world forecasting has always required juggling variables that influence each other in ways that are obvious in hindsight and expensive to miss in advance. TimesFM-3 handles three categories of supplementary data at once: multiple related variables like different product flavors, historical-only factors like past foot traffic, and known future events like planned promotions or weather forecasts.
In Google's demonstration, a model forecasting ice cream sales without the discount schedule simply continued the usual weekly pattern — confident, consistent, wrong. TimesFM-3, given the promotion calendar, predicted roughly 20 percent higher sales on each discount day. This is not a coincidence. This is a machine reading your business plans and doing the math you forgot to do.
The model also outputs nine values per time step rather than a single estimate, capturing the full range of uncertainty in each prediction. It knows what it does not know. This places it ahead of several categories of human analyst.
What happens next
TimesFM-3 is available now, and the kinds of decisions it is being asked to support — inventory, staffing, pricing — are precisely the decisions that used to require experienced human judgment accumulated over years.
The ice cream example is charming. It was chosen carefully. The same architecture that predicts waffle cone demand predicts everything else on the same time series, with the same zero-shot confidence, without being asked twice.