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 time series forecasting model that predicts future values by analysing not only historical data but also related contextual signals such as weather forecasts or planned discount schedules. Unlike step-by-step forecasting approaches, the 330-million-parameter model generates all future time points in a single pass.

According to Google Research, this single-pass approach reduces compute time and limits the compounding errors that can build up when a model forecasts one step at a time and feeds each prediction back in as input for the next.

Why it matters

Forecasting underpins a wide range of operational decisions — inventory levels, staffing, pricing, energy demand and supply chain planning among them. A model that can factor in known future events, such as an upcoming promotion or a forecasted weather pattern, alongside historical patterns has the potential to produce more contextually grounded predictions than models relying on historical data alone.

The efficiency gains from single-pass prediction are also significant for organisations weighing the practicality of deploying forecasting at scale. Faster, less error-prone forecasting could make it more feasible for enterprises to run frequent, granular predictions across many product lines, locations or variables rather than relying on periodic, aggregated forecasts.

By the numbers

• 330 million parameters make up the TimesFM-3 model.

The Renascence take

Forecasting tools tend to be judged purely on statistical accuracy, but the more interesting shift here is architectural: a model that can ingest planned future events — a discount, a weather forecast — alongside historical data changes what forecasting is actually for.

Most conversations about demand forecasting focus on model accuracy in isolation, missing that the real value lies in connecting forecasts to decisions that are already known but siloed — a promotion calendar sitting in marketing, a weather feed sitting in operations. The behavioral lesson is that better predictions don't improve outcomes unless the people running the business — merchandisers, planners, service teams — actually trust and act on them quickly. Any operator evaluating a model like this should test it not just for accuracy, but for how fast and confidently it changes a real operational decision, such as a restocking call or a staffing adjustment.