Online retail
Replacing rule-based ordering with an economic ordering model
An ordering model that priced overstock and stockout risk explicitly, replacing rule-based replenishment across a large catalog.
$13M+
Measured financial impact
Client
Large online retailer
Services
Basis
Partner experience, prior role
The insight
“The forecast was not the problem. The order was. Every item was being treated as if running out and overbuying cost the same.”
Exhibit
Ordering the forecast is not ordering well
Expected cost of a single ordering decision, by units ordered
Illustrative, computed from stated assumptions: lognormal demand with mean 100 and coefficient of variation 0.45; a unit short costs 4× a unit left over. The forecast is unbiased, yet ordering it costs 21% more than the optimum.
Rules that treated every item the same
The retailer managed a large and fast-changing catalog with replenishment driven largely by rules: target weeks of cover, service-level settings, and planner judgment. The rules were easy to explain, but they treated every item as if the cost of running out and the cost of holding too much were the same.
Better forecasts could not close the gap
Demand for much of the catalog was volatile and intermittent, and historical sales understated true demand wherever items had been out of stock. Forecast accuracy metrics looked reasonable, yet the business carried excess inventory on some items while losing sales on others. Improving the forecast alone was not going to close that gap.
Making the order itself the decision
We reframed the problem around the order rather than the forecast. Demand was modeled as a probability distribution, corrected for stockout censoring. Holding cost, shortage cost, and ordering constraints were written into an explicit economic objective, and the ordering policy was optimized against it.
Before launch, the model was replayed against historical periods to compare its decisions with those actually made, which established the expected value before any production order was placed.
More than $13M in measured impact
The ordering model launched into production and delivered more than $13 million in measured financial impact. Just as important, the economics behind each order became explicit, giving planners and finance a shared basis for discussing tradeoffs.
What made the difference
01
Correct the history
Demand was modeled as a probability distribution and corrected for stockouts, so the model learned true demand rather than constrained sales.
02
Price both mistakes
Holding cost, shortage cost, and ordering constraints became one explicit economic objective that planners and finance could read.
03
Replay before launch
The policy was replayed against historical periods to measure its value before a single production order was placed.