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

What happened

Google Research has released TimesFM-3, a new forecasting model that predicts future values across entire time series in a single pass rather than generating them step by step. The 330-million-parameter model can incorporate related data streams and known future events — such as planned discounts or weather forecasts — alongside historical figures to produce its predictions.

According to The Decoder, the model's "fill in all future points at once" approach departs from the conventional step-by-step forecasting method used by many time-series models, where each new prediction is generated based on the previous one. That sequential process is computationally heavier and prone to compounding errors as mistakes in early predictions ripple through later ones.

By processing the full forecast horizon in one pass, TimesFM-3 is designed to reduce both compute time and the accumulation of errors that typically affects longer-range forecasts.

Why it matters

Forecasting underpins a wide range of operational decisions — inventory planning, demand sensing, staffing, pricing and capacity management — and the accuracy of those forecasts directly shapes service levels and cost. A model that can factor in known future events, such as a scheduled promotion or an expected weather pattern, alongside historical data moves forecasting closer to the kind of contextual reasoning that operations teams currently do manually or through fragmented tools.

The architectural shift away from step-by-step generation is significant on its own terms: faster, less error-prone forecasting at scale makes it more practical to run frequent, granular predictions across large product catalogues, store networks or service regions, rather than relying on periodic, coarse-grained forecasts.

By the numbers

• 330 million parameters make up the TimesFM-3 model, a relatively compact size for a forecasting model of this capability.

The Renascence take

It's tempting to read this as a pure infrastructure story, but the real implication sits closer to the customer than it appears.

Most forecasting failures that customers actually feel — empty shelves during a promotion, understaffed contact centres during a heatwave, surge pricing that looks arbitrary — trace back to models that can't reason about known future context, not just historical patterns. A model built to natively ingest promotional calendars and weather data alongside past demand is really an attempt to close the gap between "what we predicted" and "what we knew was coming." The operators who benefit won't be the ones with the biggest data teams, but the ones who are disciplined about feeding structured, known-event data into their forecasting pipelines in the first place — the model is only as context-aware as the inputs it's given.