Predictive analytics in retail: inputs and forecast checks

Predictive analytics in retail estimates future demand and tests those estimates against what subsequently happened. For a retailer reviewing supplied forecasts, the practical question is whether predictions beat a simple baseline at the store, product and time horizon where someone must act. A matching company sales total is not enough.

This is a forecast review for merchandise retailers. The broader definition and forecasting methods belong in predictive analytics; here the starting point is an existing forecast export, not a new forecasting system.

What the prediction is meant to support

A weekly unit forecast can inform replenishment. A dollar sales forecast can inform a store budget. Those are different targets: a price change can raise revenue without increasing units, while a product substitution can change which shelf needs stock without changing total category sales.

Write down the target, forecast horizon and decision owner before scoring anything. The forecast horizon is the gap between the forecast's issue date and the period being predicted. A forecast issued before purchasing commitments should not be evaluated against a revision made after those commitments.

NIST's introduction to time-series analysis explains that observations over time can contain trend, seasonality and dependence between periods. In retail, that means a holiday week and an ordinary week may need different comparison periods. More history does not make unlike weeks interchangeable.

Keep this review separate from the full sales and stock reporting covered in retail analytics. The question here is what the forecast knew in advance and how far its predictions missed.

Assemble a fair comparison from the exports

Keep one comparison row per store, SKU and target period. A supplied forecast file should identify those fields, forecast units, the issue date and the forecast version. The actual-sales file should contain the same identifiers and actual units on the same calendar.

Join the files using the complete combination. A store total joined to SKU rows would repeat the total. Two forecast versions joined without selecting one would repeat the actual sales. Keep unmatched rows visible rather than quietly removing difficult products.

Add the baseline forecast as another column. This could be the previous comparable week's sales or the same seasonal period from the prior year, provided that information was available at the forecast date. Choose the baseline before seeing which one makes the supplied forecast look strongest.

For a genuine holdout check, earlier periods inform the prediction and later periods supply the actual results. Hyndman and Athanasopoulos explain in their forecasting evaluation guidance that accuracy must be assessed using observations that were not used to fit the forecasting method. A good historical fit is a different result.

Worked example: a correct total hides incorrect allocations

An illustrative US apparel retailer evaluates forecasts issued on September 7, 2026, for September 14 through September 20, 2026. Each row is one store and SKU. The baseline column was frozen at the same issue date and repeats the previous comparable week's actual units.

Use forecast minus actual as the signed error throughout. Positive means overforecasting; negative means underforecasting.

The forecasts and actuals both total 500 units. Yet the retailer overforecast jackets and underforecast T-shirts at both stores. A buying team that looked only at the total could still send the wrong assortment.

Weighted absolute percentage error, or WAPE = Sum of absolute row errors / Sum of actual units

Signed bias rate = Sum of signed row errors / Sum of actual units

Mean absolute error = Sum of absolute row errors / Number of evaluated rows

Percentages are rounded to one decimal place; unit totals are exact. The reconciliation is 500 forecast units minus 500 actual units equals the zero signed-error total. Absolute errors must be calculated before summing: taking the absolute value of the total error would incorrectly give zero.

The supplied forecast beats this baseline on these rows, but the test covers only one week. Repeat it across distinct selling periods before making a broader claim. The details of denominator choice belong in WAPE forecast accuracy.

Decide what the errors mean and who follows up

The demand planner owns the frozen forecast and its assumptions. Store operations owns explanations for closures, shortened hours or stockouts. Merchandising owns assortment and promotion context. Finance checks whether unit and revenue comparisons use consistent prices and return policies.

For the example, the immediate review is the jacket and T-shirt allocation. It is not a reason to raise or lower the entire company forecast. Check whether availability constrained the T-shirt sales before interpreting the observed underforecast as its full size.

A forecast also does not select the right action by itself. Ordering more stock depends on lead times, purchase commitments and the cost of leftovers. Prescriptive analytics addresses that next decision, with explicit constraints and an accountable owner.

Where it goes wrong

• Later revisions replace original forecasts. A forecast updated after sales begin benefits from information unavailable when the decision was made. Preserve versions and evaluate the relevant issue date.
• Stockouts are read as weak demand. Recorded sales can fall below customer demand when the item is unavailable. Keep availability exceptions beside the error results.
• Aggregate errors cancel. The example's zero bias hides 60 units of absolute error. Check both measures at the decision's actual level.
• Different rows favor different forecasts. Removing missing forecasts from one candidate but retaining them for another changes the test population. Report coverage and use a common comparison set.
• Zero actuals disappear. A row with no sales can still have a meaningful unit error. Retain it; flag a percentage measure as undefined when its entire actual-sales denominator is zero.

Check your retail forecasts in Covirage

Bring a sample of your own frozen forecast and actual-sales exports, with store, SKU, period and issue-date columns. Covirage maps the columns, and its tools compute and check forecast-versus-actual figures; the external AI model explains the results with every figure traced to its rows. Create a free Covirage account and set up your business before bringing your file.

See forecast analysis for the supporting business use case.