Stop ordering inventory to the average

The ask was deliberately small. An apparel brand doing eight figures online had brought us in to look at their data stack, and their head of ops mentioned January's clearance had been brutal: 380,000 dollars of winter stock moved at 60 percent off. We asked for one thing: the demand forecast for those SKUs, next to what the buying team actually ordered.

The forecast was fine. That was the uncomfortable part. Actual demand came in within 12 percent of the prediction. The money was lost in a different place: the plan ordered to the forecast number itself, the average expected demand, for every SKU, as if being right on average were the goal.

It is not. Because the two ways to be wrong do not cost the same.

A forecast is a bell curve, and ordering to its center means treating a lost sale and a leftover unit as equally painful. They never are.

Take a parka that costs 40 dollars landed and sells for 120. Run out early, and each unit of unmet demand forgoes 80 dollars of margin, plus whatever the shopper concludes about coming back. Overbuy, and each leftover unit ties up 40 dollars until clearance recovers maybe half. Under-ordering that SKU hurts roughly four times more per unit than over-ordering it. Now take a trend-driven knit at thin margin with no carryover value: the asymmetry flips entirely, and leftovers are the expensive direction.

Ordering the average unit count treats those two SKUs identically. The buying team is not making an error of forecasting; they are making an error of translation, collapsing a distribution into a single number and losing the one property, the shape of its uncertainty, that the ordering decision actually needs.

The method has a name older than computers, the newsvendor problem, and a modern form that is entirely practical: probabilistic forecasting. Instead of predicting "this SKU sells 800 units," the model predicts the full range with probabilities attached: a 10 percent chance demand stays under 620, a 50 percent chance under 810, a 90 percent chance under 1,060. Gradient-boosted quantile regression or a probabilistic time-series model produces exactly this, per SKU, per location, per season, trained on your own sales history with promotions and stockout periods handled honestly.

Then the economics pick the order point. The margin-to-markdown ratio of each SKU implies a critical service level: the parka's four-to-one asymmetry says order at roughly the 80th percentile of its demand curve; the thin-margin knit's says order near the 45th. The forecast supplies the shape; the unit economics supply the cut point; the order quantity falls out as arithmetic. Buying stops being a negotiation between a spreadsheet average and a merchant's gut, and becomes a policy you can audit SKU by SKU.

(Handling stockout periods honestly is the quiet skill in that sentence. Sales data records what you sold, not what shoppers wanted; every stockout censors the truth downward. Train on raw sales and the model learns your past shortages as low demand and recommends them again.)

We built this decomposition for our own demand tooling, and the pattern from the apparel brand's history was typical: high-margin evergreen SKUs chronically under-ordered, trend SKUs chronically over-ordered, each by the same well-meaning average.

Rerun against two years of history, quantile-based ordering priced the gap: about 510,000 dollars a year between margin left unclaimed on stockouts and capital burned in clearance, against the roughly 40,000 the forecast's accuracy error explained. The team had spent eighteen months tuning the forecast. The expensive mistake was never in the forecast. It was in the single number they asked it for.

TensorLabs builds probabilistic demand systems for retailers whose planning still runs on point estimates. If your last clearance season felt bigger than your forecast error, that gap has a distribution behind it. Send us your worst overbuy or stockout story and we will tell you which quantile your economics were actually asking for.