Read forecast WAPE and bias without misleading accuracy claims

WAPE measures the size of errors

Weighted absolute percentage error compares the sum of absolute forecast errors with the sum of actual demand over the selected observations. Multiply by 100 to express it as a percentage. It is not an average of per-item percentage errors, and high-volume items contribute more to the result.

Worked example: WAPE and bias can disagree

Illustrative forecasts of 110 and 90 against actuals of 100 and 100 have absolute errors of 10 and 10. WAPE is 20 / 200 × 100 = 10%. Signed errors cancel, giving zero net bias under the forecast-minus-actual convention. Zero bias therefore does not mean zero error.

Compare the same horizon and population

ReOrder freezes weekly live forecast snapshots and evaluates complete 30-day outcomes when sufficient data is available. Pending forecasts must not be described as measured results. Compare the same time window and aggregation, and inspect product-level errors as well as a store total.

Zero actual demand needs separate handling

When total actual demand is zero, a percentage denominator is unavailable. Do not turn a missing percentage into zero error. Examine unit errors and the product's data instead. Backtest selection and realised outcome scoring are different evidence sets and must remain labelled separately.

Do not turn an illustration into a performance claim

The website's example values are not customer outcomes. ReOrder does not claim a universal accuracy percentage or superiority over a competitor without a comparable evaluation. Inspect forecast history, stockout assumptions and complete actuals before drawing a conclusion.

In the ReOrder workflow

Insights avoids misleading partial scores. The first live-data app load each week freezes a new 30-day forecast run. After the period ends, the normal Today refresh reads the matching net Shopify sales and scores it automatically.

This description shares the maintained content used by the in-app Help Center.

Check the plan with your own store

Start with one Shopify store, confirm the inputs and review every recommendation before acting. The current release supports up to 3,000 variants and reads up to 1,000 recent orders, with warnings for incomplete imports.