Blog · Procurement and supply chain · Supply chain
Retail replenishment predictions need testing at the store and SKU level. Compare a supplied forecast with observed sales and stock availability, reconcile the totals, and separate forecast misses from constrained sales.
Retail supply chain predictive analytics tests whether replenishment expectations hold up at the store and SKU level. Start with the forecast your team actually used, compare it with sales, and inspect stock availability alongside the error. A small network bias can hide opposing store misses and sales suppressed by stockouts.
The decision is usually whether to change a replenishment quantity, a planner override, or the investigation assigned to an item. Those are different actions. A forecast miss on an item available all week deserves a different review from a miss on an item absent from the shelf.
Use a prediction saved before the period began. The forecasting text Forecasting: Principles and Practice explains why accuracy must be assessed on observations that were not used to fit the forecasting method. For replenishment validation, that means preserving the version the buyer could actually use, rather than substituting a later revision.
Keep the scope specific: one selling period, one forecast lead time, one unit of measure, and one store-SKU combination per comparison row. The broader predictive supply chain analytics guide covers other operational prediction questions; this review concerns retail replenishment outcomes.
Use a forecast export with store ID, SKU, selling-period start and end, forecast creation date, and forecast units. Match it with a sales export containing the same identifiers, dates, and unit basis. Keep returns separate until the team has agreed whether the forecast predicts gross sales or sales net of returns.
Availability needs its own evidence. Bring daily saleable-stock flags or a consistent store-item availability report, plus receipts and adjustments when investigating exceptions. Define an available day before calculating the rate: for this example, it means the item met the retailer's documented availability rule for that trading day.
Do not replace availability with a month-end inventory figure. The Census Bureau's retail definitions distinguish merchandise held at the end of the reporting period from sales during the month. A closing balance cannot establish whether an item was available throughout the preceding selling days.
For deeper work on persistent directional misses, use demand forecast bias by SKU and site. Keep this comparison focused on whether the supplied prediction is usable in the presence of the observed availability record.
A fictional US retailer reviews the week ending October 11, 2026. Every row represents an active item scheduled for seven trading days. Forecasts were saved at the same replenishment lead time; sales and forecasts are in individual units.
Define signed error as forecast minus recorded sales. Positive therefore means the forecast exceeds observed sales. Some organizations use the opposite sign, so write the convention on the report.
Signed bias = Sum of forecast minus sales / Sum of sales
Weighted absolute percentage error, or WAPE = Sum of absolute forecast-minus-sales errors / Sum of sales
Availability = Available store-SKU days / Eligible store-SKU days
| Store | SKU | Forecast units | Sales units | Signed error | Absolute error | Available days | Eligible days |
|---|---|---|---|---|---|---|---|
| Austin | A | 120 | 100 | 20 | 20 | 7 | 7 |
| Austin | B | 80 | 100 | -20 | 20 | 7 | 7 |
| Dallas | A | 100 | 80 | 20 | 20 | 5 | 7 |
| Dallas | B | 100 | 100 | 0 | 0 | 7 | 7 |
| Total | 400 | 380 | 20 | 60 | 26 | 28 |
| Check | Calculation | Result |
|---|---|---|
| Forecast reconciliation | 120 + 80 + 100 + 100 | 400 units |
| Sales reconciliation | 100 + 100 + 80 + 100 | 380 units |
| Error identity | 400 - 380 = 20 - 20 + 20 + 0 | 20 units |
| Signed bias | 20 / 380 | 5.3% |
| WAPE | 60 / 380 | 15.8% |
| Availability | 26 / 28 | 92.9% |
Percentages are rounded to one decimal place after calculation. The absolute error total is calculated from the individual rows; taking the absolute value of the network's signed error would discard the opposing Austin misses.
The forecast exceeds recorded sales overall, but that does not prove every location is overstocked. Dallas SKU A was available for only five of seven eligible days. Its positive error is measured against constrained sales, so investigate the availability gap before reducing its replenishment forecast.
The demand planner owns the frozen forecast and explains overrides. Store operations owns shelf-availability exceptions and checks whether recorded inventory was saleable. Replenishment or purchasing owns receipt timing and the response to supplier shortfalls. Finance can reconcile unit valuations when the review extends to inventory exposure.
For Austin SKU A, review whether an overly optimistic assumption persisted while the item remained available. For Austin SKU B, investigate a possible underprediction using the same lead time and comparable periods. For Dallas SKU A, first trace receipts, stock adjustments, and availability records. Dallas SKU B provides no forecast exception in this week, although one week alone cannot establish consistent performance.
Write the next action against the store-SKU record, with an owner and review date. Repeat the comparison on later periods before treating a single miss as a reliable pattern. Predictive analytics in retail covers other retail decisions; avoid importing customer or promotion conclusions into a replenishment result without the necessary records.
Bring the forecast version your buyers used, the matching sales export, and availability records to Covirage; it maps the columns, its tools compute and check the comparisons, and the external AI model explains them with every figure traced to its rows. Create a free Covirage account, which takes about a minute, then set up your business and bring a sample of your own data.
See the supply chain industry page for the supporting operational context.
Sales are a useful observed outcome when an item was available throughout the selling period. During stockouts, sales are constrained by supply. Keep availability flags beside forecast errors and identify any separately estimated lost demand rather than silently adding it to recorded sales.
Bias preserves the direction of the miss, so overpredictions and underpredictions can cancel. Absolute error counts both directions as misses. A retailer should inspect both by store and SKU before deciding whether to adjust a forecast or investigate replenishment execution.
Use the version available when the replenishment decision was made, with a consistent lead time. Save its creation date and intended selling period. Comparing a recently revised forecast with an older actual period would give the planning process information it did not originally have.
No. Availability also depends on receipts, supplier performance, shelf replenishment, inventory accuracy, and allocation decisions. A forecast can be close to observed demand while a late delivery causes empty shelves. Review forecast measures beside availability and receipt exceptions.