Which demand-forecasting approach would you fund for a retail inventory rollout?

A fictional retailer must choose its current seasonal forecast or a machine-learning forecast for replenishment. Both were evaluated on the same eight unseen weeks, using only information available at each forecast date and the same two-week lead time and ordering policy. The holdout covers the stores and product groups in the proposed rollout, including ordinary promotions. Daily requests include recorded unfilled orders, so demand is not limited to completed sales.

The exhibit reports weekly averages. Every week has 10,000 requested units. Each unfilled unit costs an expected $12 in lost contribution; each excess unit costs $3 in holding or disposal for the period. These charges are not counted elsewhere. Incremental forecast operating cost is shown separately. All figures and costs are invented for this exercise.

The minimum weekly unit fill rate is 97%; storage can accommodate up to 800 excess units. The machine-learning option saved $900–$1,500 in each holdout week after its operating cost. Either approach can use the current replenishment integration. There is no separate implementation charge for this decision. Recommend which approach to fund for the represented stores.