Google's TimesFM-3 cuts forecast errors by predicting 20% sales spikes from promo data

Google Research has dropped TimesFM-3, a 330-million-parameter model that finally treats forecasting as a multivariate problem rather than a solitary math exercise. Previous iterations forced businesses to analyze single data streams in isolation. This update changes the game by ingesting related variables simultaneously. Think ice cream sales. Old models looked at past scoops sold. TimesFM-3 looks at scoops, waffle cone inventory, local weather forecasts, and scheduled discount campaigns all at once. The result is a prediction engine that understands context, not just history.

The technical shift is subtle but powerful. Instead of predicting one time step after another—a slow process where early mistakes compound into later disasters—TimesFM-3 fills in all future blanks in a single pass. It groups 32 consecutive data points into patches and normalizes them, allowing it to compare apples to oranges, or rather, foot traffic to syrup sales. Google claims this one-shot approach eliminates the error accumulation that plagued earlier autoregressive methods. In their retail example, the model didn’t just see a flat line; it recognized a planned promotion and adjusted its output to expect roughly 20 percent more units on those specific days.

Performance metrics back up the architectural tweaks. On benchmarks like Gift-Eval and FEV-Bench, TimesFM-3 outperformed Amazon’s Chronos-2 and Google’s own TimesFM-2.5 in both point accuracy and uncertainty calibration. It doesn’t just give you a number; it outputs nine values per time step to map the range of possible outcomes. This probabilistic approach is critical for risk management in supply chains. You don’t just want to know how much stock to order; you need to know the likelihood of running out if a heatwave hits.

Availability is immediate for developers, with code live on GitHub and Hugging Face. BigQuery integration follows in weeks, bringing these capabilities to enterprise data warehouses without requiring custom ML pipelines. While competitors focus on scaling parameter counts, Google is betting on better data utilization. The real test won’t be benchmark scores, but whether retail planners trust an AI that suddenly sees connections between weather patterns and waffle cone demand. If it works, the era of siloed forecasting is over.