Large Universe Models in retail

What it ingests

Point-of-sale data, inventory positions, competitor pricing and promotions, weather forecasts, local event calendars, social and search trends, supplier lead times, and returns data.

The belief that matters

Demand per item per location for the coming period, revised as the drivers move rather than fitted to a seasonal curve.

Why history is not enough

Seasonal forecasts encode what usually happens. The costly errors happen when something unusual is happening and the evidence for it exists — a heatwave forecast, a competitor's promotion, a local event, a supply disruption on a substitute.

Each of those arrives on a stream. None of them appear in the sales history until it is too late to reorder.

Substitution and cannibalisation

Demand beliefs are not independent. A stockout moves demand to a substitute; a promotion cannibalises a neighbour. Holding these as a linked graph makes the second-order effects explicit instead of surprising.

Limits

Retail is high-volume and low-margin per decision, so the economics require that the model run cheaply across a very large number of item-location pairs. This is an engineering constraint on deployment, not an objection to the approach.