MAPE

MAPE is attractive because it is scale‑free: a percentage can be compared across items with completely different volumes, and it needs no context to be read. That convenience is also its weakness. The denominator is the actual, so identical absolute misses produce very different percentages depending on how much sold, and a single low‑volume item can dominate the average. Reported alongside a bias figure and a volume‑weighted alternative, MAPE is useful; reported alone, it says less than it appears to.

• MAPE = (1/n) Σ |A − F| / A × 100, where A is actual demand and F the forecast
• Worked example: actuals 100, 20, 5 against forecasts 90, 30, 10 give APEs of 10 %, 50 % and 100 % – MAPE 53 %, even though total demand was 125 and the total forecast 130
• WAPE = Σ|A − F| / ΣA × 100 – the same errors weighted by volume come out at 21 %
• MAD/Mean ratio – mean absolute deviation divided by mean demand; works where actuals are frequently zero

MAPE in practice

Two properties turn MAPE into a trap on retail data. The first is an asymmetric ceiling. An underforecast can be wrong by at most 100 %, because the forecast cannot go below zero; an overforecast has no upper bound. A model or a planner optimising MAPE therefore forecasts below the mean on volatile and slow‑moving items. Take an item averaging two units a week with an occasional peak of ten: forecasting two produces an 80 % error in the peak week, forecasting six produces 200 % in an ordinary one. Minimising the metric builds a negative bias into the plan, and that bias is paid for later in stockouts rather than in the accuracy report.

The second property is the zero. MAPE cannot be computed when actual demand is zero, and at SKU × store × day – the grain a replenishment decision runs on – an item selling a few units a week produces more zero days than non‑zero ones. Dropping those observations removes precisely the cases where the forecast was above zero and nothing sold: the overforecasts. What remains is an optimistic number produced by data selection, not by a better forecast. Adding a constant to the denominator or switching to sMAPE hides the problem rather than solving it; the honest fix is to change the metric at that grain, not the arithmetic.

Equal weighting is the third issue. In a report covering 30,000 SKUs, the unweighted average is driven by the tail, not by the items that generate revenue – which is why a company can improve MAPE by two points while availability on its top sellers falls. For inventory work the practical setup is a volume‑weighted error next to a separately reported bias, evaluated at the grain and horizon the order is placed on. Veritico STOCK forecasts at item, store and day level, picks from around 80 statistical methods the one with the lowest expected error, and cleans stockouts, promotions and outliers out of the history before the forecast is computed, so the model is not fitted to censored sales – see demand forecasting and inventory optimization and replenishment and allocation management. Mondelez moved forecast accuracy from 50 % to 70 % once planning was unified across functions. On promotional forecasts, where percentage metrics are least forgiving because baseline volumes are small relative to the uplift, Kofola reached 76 % with Veritico PROMO.

Related terms

• Forecast accuracy
• Demand forecasting
• Demand planning
• Intermittent Demand
• Forecast Bias
• Seasonality