MAPE and WAPE both express forecast error as a percentage, but they weight errors differently. MAPE averages the percentage error of each item or period, so small-volume lines count as much as large ones. WAPE divides total absolute error by total actual demand, so volume decides the weight. For mixed portfolios, WAPE is usually the safer headline metric.
Neither measure shows direction. For that you need bias, which is why the three are best read together.
The formulas in plain terms
For each item and period, the absolute error is the gap between actual demand and forecast, ignoring sign.
- MAPE (mean absolute percentage error): divide each absolute error by that period’s actual, then average those percentages.
- WAPE (weighted absolute percentage error): add up all absolute errors, then divide by the sum of all actuals. Some teams call a closely related version MAD/Mean ratio.
- Bias: add up the signed errors (forecast minus actual) and divide by total actual. A positive result means over-forecasting, a negative one under-forecasting.
A small illustrative example shows the difference. Suppose two products: product A has actual demand of 10 and forecast of 15; product B has actual demand of 1,000 and forecast of 1,050. Product A is 50 percent off and product B is 5 percent off. MAPE averages these to 27.5 percent. WAPE adds the errors (5 plus 50 equals 55) and divides by total actual (1,010), giving about 5.4 percent. The numbers are invented for illustration, but they show why the two metrics can tell very different stories about the same forecast.
Where each metric misleads
MAPE has well-known weaknesses. It is undefined when actual demand is zero, and it explodes when actuals are very small, so intermittent or low-volume items can dominate the average. It also treats over-forecasts and under-forecasts asymmetrically: an under-forecast cannot exceed 100 percent error, but an over-forecast has no upper limit. Optimising for MAPE can therefore push a model towards forecasts that are too low.
WAPE avoids the zero-division problem at portfolio level and reflects business impact better, because large-volume items matter more. Its weakness is the flip side: a few big items can hide poor accuracy on many small ones. If a slow-moving range has real cost, such as expensive spare parts, a WAPE headline can look healthy while that range performs badly.
Bias answers a different question entirely. A forecast can have a modest WAPE and still be consistently high, which quietly inflates inventory. Or it can be consistently low and cause stockouts. Error metrics based on absolute values cannot tell you which.
Comparison table
| Criterion | MAPE | WAPE | Bias |
|---|---|---|---|
| What it measures | Average percentage miss per item-period | Total miss relative to total demand | Direction of the miss |
| Weighting | Equal across items | By volume | By volume |
| Handles zero demand | Poorly; undefined for zeros | Yes, if the total is not zero | Yes, if the total is not zero |
| Sensitive to low-volume noise | Very | Little | Little |
| Shows over or under-forecasting | No | No | Yes |
| Best use | Comparing similar high-volume items | Reporting accuracy across a mixed portfolio | Spotting systematic optimism or caution |
| Main risk | Distortion from small actuals; favours under-forecasting | Big items hide small-item problems | Positive and negative errors can cancel across items |
Choosing by situation
High-volume, stable items
Here MAPE and WAPE usually agree closely, because actuals are large and similar. MAPE is easy to explain and fine for comparing products of similar scale. WAPE is still a good choice when you need one number for a product family.
Low-volume or intermittent items
Avoid MAPE as the main measure. Zeros and tiny actuals make it unstable. Use WAPE aggregated over a sensible group or period, and consider measures designed for intermittent demand, such as scaled errors that compare against a simple baseline forecast.
Mixed portfolios
Report WAPE for the overall picture, and split it by segment (for example fast, medium and slow movers) so large items do not mask small-item problems. Add bias alongside each segment.
Aggregation level and horizon
Errors shrink as you aggregate across products, locations or time, so a forecast can look accurate at a national monthly level while being poor for a single store-week. Always state the level and the lag at which accuracy is measured, such as the forecast made one month ahead of the period.
Practical steps for a sound measurement set-up
- Decide the decision the forecast supports (buying, production, staffing) and choose the level where that decision is made.
- Report WAPE as the main accuracy measure and bias next to it, never alone.
- Segment items by volume and variability, and track each segment separately.
- Compare against a naive baseline, such as last period or last year, so you know whether the model adds value.
- Keep the measurement lag consistent, and record the forecast as it stood at that lag, not as revised later.
- Review which items drive most of the error, and fix those first.
When machine learning models are used, they are often trained to minimise one error measure, and the metric chosen shapes the behaviour of the forecast. A model trained on a percentage error may drift low; one trained on a squared error may chase outliers. Pick the training measure deliberately, then evaluate with the business-facing metrics above.
Sources and further reading
- Association for Supply Chain Management (ASCM)
- MIT Center for Transportation and Logistics (MIT CTL)
- Harvard Business Review: Supply Chain Management topic (Harvard Business Review)