WAPE is popular in forecasting because it offers an attractive property: a simple percentage view of total absolute error relative to actual volume. That simplicity is valuable. The problem begins when an aggregate metric is treated as a complete explanation of the forecasting system. A low WAPE can coexist with serious errors in critical products, particular locations or periods where operations have little room to react. It can also hide persistent bias because the numerator uses absolute errors. The metric answers how much total error exists, not necessarily where it sits or which direction it takes.
What WAPE does well
WAPE divides total absolute error by total actual demand. That makes it intuitive for large portfolios and avoids some of the instability created by percentage errors calculated observation by observation. When the question is whether total error is small or large relative to total volume, WAPE is a useful reference. It is also practical for tracking changes over time. If the same portfolio moves from a high WAPE to a materially lower one under comparable conditions, that represents a real statistical improvement. As a summary metric, WAPE is compact, interpretable and useful for governance.
The problem comes when it becomes the only criterion. One number cannot describe the error distribution, bias, business criticality or the downstream cost of being wrong.
Where information disappears
The denominator of WAPE is total actual volume. High-volume items therefore dominate the result. A product selling thousands of units can improve enough to offset deterioration across several low-volume products. That may be perfectly acceptable if the business also makes decisions at portfolio level. But if procurement, replenishment or service targets operate by SKU and warehouse, the granularity of WAPE does not match the granularity of the decision. Criticality is invisible as well. A low-volume product may be strategic, high margin or capable of stopping production when unavailable. Its error contributes very little to WAPE even though the business cost can be high. The metric weights volume, not economic importance.
Bias matters
WAPE uses absolute error, so an over-forecast and an under-forecast of equal size contribute the same amount. An inventory policy does not necessarily experience them in the same way. Persistent under-forecasting may create stockouts; persistent over-forecasting may create excess inventory. This means a stable WAPE can hide a systematically biased forecast. Two models with the same aggregate score may require very different operational responses if their errors lean in opposite directions.
At minimum, a bias measure should sit next to WAPE. Horizon-level analysis also helps because a long-range miss may still be correctable while the same miss close to an order cutoff can be expensive. Timing and direction contain information that WAPE deliberately removes.
How to use WAPE safely
WAPE works best as one layer in a small metric stack. It can provide the executive summary while diagnosis relies on bias, segment-level error, horizon-level error and an economic impact measure. Segmentation should follow the decision. There is no need to slice every dimension; the useful cuts are the ones where error consequences change. Criticality, lead time, margin, volatility and channel are often more informative than arbitrary reporting categories. Thresholds should also differ. Expecting the same WAPE from a stable high-volume item and an intermittent series can produce unrealistic targets. Metrics need to be interpreted within the behaviour of the segment.
The conclusion is not that WAPE is a poor metric. It is that a good summary metric can become a poor diagnostic tool when used alone. Forecasting should not optimize a global percentage for its own sake; it should improve the decisions where errors carry real consequences. It is also worth checking WAPE stability across periods. A score that improves sharply during high-volume weeks and deteriorates when the mix changes may be describing the denominator more than the underlying forecasting quality. Temporal consistency is part of the diagnosis. Comparing like-for-like windows helps separate a structural model improvement from an apparent improvement created by demand composition.