Plan retail replenishment with clean history, honest forecast tests and separate continuous/weekly review calculations, including a worked kettle example.
Penulis
Anichur Rahaman
Imagine a three-shop homeware retailer on a busy Friday. The popular kettle has sold out, but 140 candle holders still occupy the stockroom. The supplier needs two weeks for the next kettle delivery. This is an illustrative scene, not an observed customer result.
The kettle’s old reorder setting reflected 30 weekly sales; the planning team now sees a much higher run rate. Buying more of everything would tie up more cash without fixing the decision. They need to understand demand, supplier timing and how often they actually review stock.
This guide follows a fictional eight-week SKU history through a simple forecast test and two different replenishment policies. All quantities are teaching assumptions. The formulas are useful only when their assumptions match the operation; they do not promise a stockout-free Friday. Reviewed 4 November 2026.
Define SKU, location, unit, time bucket and planning horizon first. Sales record fulfilled purchases, not every unit customers wanted. Stockouts can censor demand; a zero-sale day does not by itself prove either zero demand or lost sales. Check availability, trading hours and reliable enquiry or lost-sale evidence.
Keep observed sales, estimated lost demand and management overrides separate. A stockout estimate is uncertain, not a transaction to insert into the sales ledger. Do not automatically delete every large order: identify whether it is a repeatable account, a one-off event or a commitment already known for the next horizon.
Forecast outbound demand and expected usable returns separately. A returned kettle may be late, damaged or unsellable. Net sales are useful financially but are not automatically the quantity to replenish. Include only inspected or otherwise reliably available return supply in the relevant period.
The table keeps the original teaching data. Week 2 has a known one-off 10-unit component excluded from this baseline, while remaining real consumption in the stock records. Week 4 includes a provisional seven-unit lost-demand estimate. Assume both annotations were available by that week’s close; deriving them from future weeks would contaminate the backtest.
| Week | Recorded units | Baseline adjustment | Planning series |
|---|---|---|---|
| 1 | 44 | 0 | 44 |
| 2 | 71 | −10: identified one-off | 61 |
| 3 | 52 | 0 | 52 |
| 4 | 31 | +7: provisional lost demand | 38 |
| 5 | 55 | 0 | 55 |
| 6 | 49 | 0 | 49 |
| 7 | 63 | 0 | 63 |
| 8 | 38 | 0 | 38 |
Recorded sales total 403; the adjusted series totals 400, an eight-week mean of 50. Similar totals do not mean the weekly timing is accurate. Retain the reason, evidence, date and owner of each adjustment and test alternative lost-demand assumptions.
Flag promotion and price changes rather than silently removing them. Trend and calendar effects need an appropriate baseline; a monthly seasonal ratio can confuse growth with seasonality or double-count a seasonal model. More repeated cycles help, but two years do not guarantee a reliable seasonal estimate. Moving holidays and new products need particular care.
A four-week moving average is an explainable candidate, not an automatic winner. Predict each test week using only the preceding four planning values. Compare against relevant naïve or seasonal baselines on the same dates and information. For a two- or three-week stock decision, also test the cumulative forecast over that horizon.
The following one-week rolling example demonstrates the arithmetic. Four test observations are far too few to establish a production method’s reliability. Save each forecast as it was made; do not rewrite its inputs with knowledge from later weeks.
| Week | Forecast | Actual planning value | Actual − forecast | Absolute error |
|---|---|---|---|---|
| 5 | 48.75 | 55 | +6.25 | 6.25 |
| 6 | 51.50 | 49 | -2.50 | 2.50 |
| 7 | 48.50 | 63 | +14.50 | 14.50 |
| 8 | 51.25 | 38 | -13.25 | 13.25 |
| Total | 200.00 | 205 | +5.00 | 36.50 |
After week 8, the four-week average would predict week 9 at 51.25, not 50. The stock-policy calculation below deliberately assumes a stationary planning mean of 50 from the full eight-week history. It is a separate simplified assumption, not a claim that the latest moving-average forecast equals 50.
For non-negative demand in these four test weeks, WAPE = sum of absolute errors ÷ sum of actuals = 36.5 ÷ 205 = 17.8%. If the denominator is zero, report N/A. Changes in volume, trend and case mix can make cross-period WAPE comparisons misleading; an aggregate can hide individual SKUs.
Define the sign: bias = sum(forecast − actual) ÷ sum(actual) = (200 − 205) ÷ 205 = −2.4%, so the example underforecasts overall. Errors can cancel; MAE here is 36.5 ÷ 4 = 9.125 units. Use these with service, waste and purchasing cost rather than a universal WAPE target.
For deeper evaluation, Forecasting: Principles and Practice describes rolling-origin tests; Hyndman’s WAPE note explains denominator and non-stationarity limitations. Consider scaled errors such as MASE/RMSSE where their scaling baseline is meaningful.
Continuous review: monitor inventory position as it changes and trigger the configured replenishment decision at or below a reorder point. The point says when, not how many units to buy. A separate quantity policy, open orders and approval rules still apply.
Periodic weekly review: at each scheduled review, calculate the order needed to reach an order-up-to target. Protect the time until the next review plus supplier lead time. Do not wait for a continuous-review threshold before performing this weekly calculation.
For the simple backorder model, inventory position is usable on-hand + confirmed on-order − backorders. It is not the same as physical shelf stock or available-to-promise. Reconcile reservations and other commitments to the demand model so they are not counted twice. Exclude damaged stock; assess late or uncertain receipts by due date rather than assuming every open purchase order will arrive on time. The worked case has no other commitments or shortages.
For this teaching calculation only: weekly mean d = 50, fixed lead time L = 2 weeks, independent stable weekly demand with sample standard deviation σ = √(644/7) = √92 ≈ 9.5917, and a normal approximation. Choose a 95% cycle service target, z ≈ 1.64485. Use consistent time units and round whole-unit requirements upward.
Continuous review: SS = z × σ × √L ≈ 22.31, rounded up to 23 units. ROP = d × L + SS = 100 + 23 = 123 units. That is a model-based threshold, not a guarantee or a 95% unit fill rate.
Weekly review has R = 1, so the protection interval is L + R = 3 weeks. Its buffer is z × σ × √3 ≈ 27.33, rounded up to 28 units. The target is S = 50 × 3 + 28 = 178 units. Reusing the two-week buffer would understate this model’s protection interval.
At the weekly review, suppose inventory position is 124: q = max(0, 178 − 124) = 54. With MOQ 60 and cartons of 12, order 60 units, subject to cash, capacity, expiry and approval. Position 124 is above the continuous ROP of 123; that does not cancel this periodic-review order.
| Assumed policy | Protection interval | Rounded buffer / target |
|---|---|---|
| Continuous review | 2 weeks | SS 23 · ROP 123 |
| Weekly review | 3 weeks | SS 28 · S 178 |
In the same continuous-review toy model, 90%, 95% and 99% cycle targets give rounded buffers of 18, 23 and 32, and reorder points 118, 123 and 132. Higher targets require more buffer here; their business value depends on shortage consequences and holding cost. They are theoretical inputs, not observed service performance.
Eight adjusted observations do not validate a normal, independent distribution. With seasonal, correlated, intermittent or varying-lead-time demand, estimate errors over the actual protection interval or simulate suitable scenarios. Track supplier order-to-usable-receipt times and uncertainty; a two-week average is not a fixed two-week promise. Review actual stockouts, fill rate and excess stock after implementation.
For the inventory concepts and policy assumptions, see the NC State safety-stock tutorial and Rossetti’s continuous/periodic review chapter. The arithmetic here is an independently calculated teaching example.
Use ABC to prioritise a declared value measure, such as annual usage cost, contribution or revenue. Thresholds are management choices, not a compulsory 80/15/5 split. XYZ can describe variability using CV = σ/mean, but it needs a defined window, sensible treatment of zero mean and checks for seasonality.
In this toy series CV ≈ 9.5917/50 = 0.192. A locally chosen X threshold may classify it as relatively steady; that alone does not authorise automated purchasing. Low-revenue spares can be business-critical, and a volatile item can still deserve stock. Never drop a CZ item merely because of its label.
Compare a more complex model with a credible simple baseline on information available at decision time. Promotions, price, substitutes and product similarities may help, including for new items; they do not guarantee predictive value. SKU count alone is not a rule for choosing or rejecting machine learning.
Choose for the decision: out-of-sample horizon errors, service/waste, operating cost, explainability and exception handling. Reconcile location and aggregate plans where needed. StoreConsole’s inventory module can be evaluated for the records and movement controls you need; confirm the enabled forecasting and purchase-approval workflow separately. A low-stock alert or product tour is not proof of an automatic forecasting engine.
In the fictional shop, the buyer now knows which figure is an observation, which is an estimate and which purchasing policy applies. They can challenge the kettle order and make a separate decision on the candle holders. The benefit is a repeatable, auditable decision—not a promise that one formula prevents every empty shelf.
Anichur Rahaman is the creator of StoreConsole. This guide explains operational planning; all shop stories, demand values and service targets are illustrative.
Tentang Penulis
Anichur Rahaman
Lanjut Membaca
