Why Teams Measure Forecast Accuracy Wrong

A matching total is not proof of a useful forecast

A retailer can predict the correct total while placing the expected demand in the wrong stores or weeks. Opposing local errors cancel at the group level, but the stock consequences do not. The measure should match the decision the forecast is intended to support.

Core supports weekly and monthly demand baselines and planning context. Atlas supports reporting and analysis on the accepted data scope. The article’s measures are explicit worked calculations, not a claim that a named accuracy dashboard or automated model-selection process is preconfigured.

Retain what the forecast knew at the time

A forecast edited after actual trading is no longer the same prediction that supported the original buy. Keep the issue date, version, horizon and available inputs identifiable. Compare a one-week forecast with another one-week forecast made at an equivalent cutoff.

Forecast evaluation research distinguishes held-out test evidence from data already used to fit a method. A good explanation of past trading can still perform poorly on an unseen period. A planner needs the operational forecast version, not a retrospectively improved number with no decision history.

Expose errors that disappear in aggregation

The worked forecast predicts a hundred units for A and a hundred for B. Actual demand is eighty and 120. Both forecast and actual totals are 200, but A is twenty too high and B twenty too low. The local absolute errors total forty.

Weighted absolute percentage error, defined here as total absolute unit error divided by total actual units, is forty divided by 200: 20%. Mean absolute error across the two store observations is twenty units. A matching total therefore coexists with meaningful local allocation risk.

Compare an alternative on the same evidence

A separate forecast of ninety for A and 110 for B has errors of ten in each location against the same eighty/120 actuals. Its total still matches, but absolute error totals twenty and the stated WAPE is 10%. That is a useful comparison for this particular test window.

It does not prove the alternative remains superior next season or at a different hierarchy level. Use enough relevant periods and a consistent decision horizon. The example keeps two simple observations to make cancellation visible, not to establish statistical confidence or a production forecast winner.

Separate directional bias from absolute error

Bias asks whether the method tends to over- or under-forecast under a stated sign convention. Absolute error asks how far it misses without cancelling opposing signs. Both matter: a directionally neutral total can still distribute stock badly, as this example shows.

Percentage measures can also become unstable when actual demand is very small or absent. State the denominator and treatment of those cases rather than silently dropping inconvenient products. Do not rename WAPE as an accuracy percentage or subtract it from a hundred without explaining what that would mean.

Adding a low-volume observation can change percentage measures sharply. A forecast of twelve against two actual units has a ten-unit absolute error and a 500% absolute percentage error for that observation. State how such items enter the evaluation rather than hiding them; the existing two-store WAPE would become fifty absolute units divided by 202 actual units, or 24.8% rounded.

Check whether actual sales were constrained

A forecast of 120 can look too high if only eighty units were available, even when customers wanted more. Separate realised sales from the demand concept the forecast predicts, and record any estimated availability adjustment with its uncertainty.

Atlas supplies the scoped evidence; Core retains the planning assumption that may need revision. The calculation alone does not diagnose whether the miss came from demand, stock availability, price or a calendar mismatch. Use the measure to locate the question before changing a buying rule.

Evaluate the horizon that supports the commitment. A next-day replenishment forecast and a multi-week buying forecast can face different uncertainty. Comparing their error percentages without their issue time and horizon can reward the easier task rather than identify the better planning method.

Measure the forecast that can improve the next decision

For this test, the original local forecast has 20% WAPE and the alternative 10%, despite both matching the total. Retain the same cutoff, scope, horizon and demand definition when comparing another period. No automated model change or stock action is represented.

Merchmix connects the forecast version and trading evidence so the planner can review a specific decision. A credible accuracy statement describes what was predicted, when and at which grain. The goal is better supported stock decisions, not a flattering percentage that hides opposing errors.

Common questions