A newspaper left unsold at midday is a straight loss. An empty rack is a lost sale and a disappointed reader. For Groupe Rossel, that trade-off comes up every day, for every edition, across thousands of outlets.
A two-stage design
Forecast the demand. A single global model predicts daily sales for each outlet and edition pair. Its strength lies in the explanatory variables: a finely modelled French calendar, with school holidays by region, public holidays and long weekends; sales history aggregated at several levels, under strict discipline against data leakage, with every variable lagged by at least thirteen days to match the real forecasting horizon; and a geographic transfer signal that estimates demand shifting across when a neighbouring outlet is closed.
Decide the quantity. The forecast alone is not enough. The system then applies an economic model that weighs the cost of an unsold copy against the revenue lost to an empty rack, with a target service level of 96.7%. The output can be used directly: the optimal quantity to deliver, per outlet, per edition, per day.
Validation under real conditions
The thing that makes the difference is a test harness that replays the whole chain day by day, under exactly the conditions of production, and compares the results with the quantities actually shipped. Every change to the model has to prove that it beats what is already there, not on a laboratory metric but on the real orders.
The whole system is industrialised on the cloud and steered by business indicators: unsold rate, stockout rate and forecast bias by title and by city.