TimesFM is a foundation model for time series from Google Research. It produces forecasts zero-shot, which means a retailer can see a credible forecast for hundreds of products in days instead of months.
A pilot starts with history: sales per product or category, at least a year, ideally with promotions and stock-outs marked. We hold out recent periods the model never sees.
We forecast with TimesFM 2.5, show an uncertainty range, and always plot a naive baseline next to it. If the foundation model does not beat the naive forecast on WAPE or MASE, we say so.
The decision layer matters more than the model. Forecasts become replenishment suggestions, with monitoring of error, stock-outs and overstock so the buyer can trust or override them.
Accuracy is verified per business. Every catalog behaves differently, and the pilot exists to find out how yours does.