How Much of Each Product Should I Order So I Neither Run Out Nor Pile Up?

Editor’s note

This problem is commonly called “how much should I order”, “stock control” or “merchandise planning”. In the academic literature it is studied as Demand Forecasting, Time Series Forecasting and Inventory Policy Optimization. Without these concepts you cannot tell a ‘sales report’ (history view) apart from a ‘demand forecast’ (a point estimate with a confidence interval) when buying an ERP or planning tool. The first only shows the past; the second produces a decision.

The most-skipped point in the field: forecast accuracy is not the goal — minimising inventory cost is. Many SMBs chase low MAPE, but bias (systematic over- or under-forecast) hurts inventory far more than variance. A model with fifteen percent bias costs much more than a twenty-five-percent-MAPE model that is unbiased. Worse: weeks with stockouts in history mask true demand — a model trained on uncorrected data systematically under-forecasts. Censored-demand correction is the hidden critical step.

Step-by-step path — for the SMB

Stage 1 — Measure before you plan. At least eighteen to twenty-four months of SKU-level sales: daily or weekly granularity, channel split (store, e-commerce, wholesale), price history, promotion calendar, stockout days. ABC analysis: the top twenty percent of SKUs usually account for eighty percent of sales — invest model effort here. The long tail can stay on simple methods.

Stage 2 — Extract the knowledge capital. Supplier lead times (mean plus standard deviation), order multiples (pallet, case), warehouse capacity constraints, target service level (decide with the leadership: ninety-five or ninety-eight percent), seasonal calendar, promotion plan, new-product launch process. This knowledge capital shapes the inventory policy more than the forecast itself.

Stage 3 — Pilot. Twelve to sixteen weeks. Start with fifty to one hundred fifty fast-moving SKUs in a single warehouse or store group. Success criteria fixed up front: WAPE down thirty percent, absolute bias under five percent, fill rate up four points, inventory turnover up fifteen percent. Model output appears as a suggestion to the purchasing manager — the human still signs off.

Stage 4 — Scale-up. Four to eight months across all SKUs and locations. Weekly automatic retrain. ERP or WMS handles order creation. Monthly KPI dashboard: WAPE, bias, fill rate, turnover, dead-stock ratio. High-error SKU list reviewed every week.

Risks — what can go wrong

1. Stockout masking. SKUs with extended stockout history train into systematic under-forecasts and persistent stockouts. Censored-demand correction is mandatory.
2. Over-parameterisation on thin data. New products or intermittent SKUs do not support deep models. SKU classification plus method matching is required.
3. Promotions and launches. A model trained on the past treats promotion periods as normal demand. Promotion calendars must be separated and manual override must exist.
4. Service-level fantasy. Targeting ninety-nine percent fill rate roughly doubles safety stock versus ninety-five percent. The target must be set jointly with finance and operations.

Technical view of the solution method

Objective function choice:

• Objective 1 — Minimise forecast error: MAPE, WAPE, RMSE, MASE. Used at model-selection stage.
• Objective 2 — Minimise stockout count: Service-level oriented; pushes safety stock up.
• Objective 3 — Minimise total inventory cost: Carrying cost plus stockout penalty plus order cost together.
• Objective 4 — Maximise fill rate / service level: Customer-focused under a budget constraint.

Multi-objective: weighted sum or goal programming. In most practical deployments Objective 3 (total inventory cost) is the operational objective while Objective 4 (fill rate) becomes a constraint.

Academic sources

Listed in the sources field of the frontmatter.