WAPE vs. MAPE For Retail Forecasts With Slow-Selling SKUs

Hand the same baking-department forecast to two buyers. One places replenishment orders for the department and is measured on unit misses across its assortment. The other has to decide which specialty molds deserve a place in a very small stockroom. Their idea of a bad miss will differ, and so can their verdict on the same forecast, even when both use a familiar percentage error.
Take one ordinary trading week at a hypothetical kitchenware retailer. The department sold 100 standard sheet pans, four brioche molds and zero star-shaped cake tins, and all three were available all week. Two forecasts had been saved before the orders were placed. Forecast A called for 98 sheet pans, eight brioche molds and three star tins. Forecast B called for 86 sheet pans, four molds and zero tins. Which set of numbers should the buyers use next time?
Under weighted absolute percentage error, or WAPE, Forecast A wins. Work out how many units each forecast missed on each line, ignoring the direction of the miss. A missed by two, four and three, for nine units of absolute error. Divide by the 104 units actually sold and multiply by 100, and A’s WAPE comes to about 8.7%. B missed the sheet pans by 14 and got both specialty lines exactly right, so its WAPE is 14 divided by 104, or about 13.5%.
WAPE charges an extra star tin the same one unit of error as a missing sheet pan. Because the sheet pan accounts for nearly all the department’s volume, getting its forecast close can make up for a very poor call on a small line. For the buyer who commits to unit quantities across the department, nine units of error may be preferable to 14. A buyer paying a tooling deposit on every specialty shape could reasonably put a different price on those three extra tins.
MAPE and The Week With No Star Tins
Mean absolute percentage error, or MAPE, gives each observation its own percentage error and averages the results. Start with the sheet pans and brioche molds. Forecast A missed the pans by 2% and the molds by 100%, an average of 51% over those two selling items. Forecast B missed them by 14% and 0%, an average of 7%. Here the four brioche molds get as much say as the 100 pans, and the specialty buyer can see exactly where the high-volume line’s good score had obscured a poor purchasing signal.

Then there is the third row. A percentage error for star tins would mean dividing by zero, a calculation that is undefined even for Forecast B, which correctly predicted zero. Ordinary MAPE for the complete three-item set is therefore undefined. The 51% and 7% above are MAPE for the two items with positive sales, and a report showing either figure should identify the excluded observation. Quietly filtering out the tins would let A’s three unwanted units disappear from this part of the review.
WAPE keeps all three misses because it divides by combined actual sales. If all three items sold zero, WAPE would also be undefined. A simple count still works in a completely quiet specialty department: show how many units were forecast for items that sold none, and how much buying those units would have committed. At a landed cost of $18 each, A’s three star tins represent $54 of potential inventory if the buyer follows the forecast unit for unit. A minimum order or existing stock would change the actual cash decision.
There is another way to make the error look smaller by accident. Forecast A predicted 109 units in total against 104 sold. Subtract those totals first and the department appears to have missed by only five units, about 4.8%, because the two missing sheet pans canceled out part of the seven excess specialty pieces. Five is a useful signed overforecast for the department, but a buyer ordering individual SKUs still has nine units of absolute error to reckon with, and the sheet-pan order alone would have come up two short.
Score The Decision The Buyer Can Still Change
A forecast saved on Sunday night can predict Monday’s sales well and still arrive weeks too late for a purchase order. Compare both forecasts from the last moment the buyer could change the order. If the vendor needs four weeks’ notice, save and score the four-week-ahead predictions for the same products and locations. Keep the selling week fixed, too. Combining a quiet February with a holiday promotion changes the mix of items carrying the department’s WAPE, even when neither forecast has improved.
Next, look behind the zeros. A star tin available online but hidden from search has a different sales opportunity from one featured on the home page. A sheet pan that sold out on Wednesday might have sold more than 100 if stock had lasted. Low error against those recorded sales can reward a forecast for predicting the inventory limit. Mark the affected days. If the team estimates lost demand, compare both forecasts against that same estimate, and keep the comparison against observed sales available for reference.
With several weeks of results, pool the absolute unit misses and actual units to calculate WAPE over the full period. Averaging the weekly percentages instead gives a quiet week the same influence as a busy one and can produce a different ranking. For MAPE on positive observations, keep the observation consistent: one SKU in one store in one week counts once. A chain-level percentage calculated after combining every store answers a different question from whether the right molds landed at the right stores.
One week’s result is too thin to justify switching every specialty item to B. Repeat the comparison over ordering cycles with ordinary and peak demand. When evaluating inventory forecasting software, give a buyer access to the saved predictions and individual misses behind the percentages. Are extra molds accumulating across quiet weeks? Does B miss pans in the same direction every time? Persistent errors can justify changing the buy even when the department’s headline score barely moves.
For the next sheet-pan order, the difference between A and B is a dozen pans before considering stock on hand. For the brioche molds it is four, a full week’s sales at the rate in this example. Put those quantities beside the purchase order. The department buyer can see the unit advantage behind A’s smaller WAPE, and the specialty buyer can challenge the additional molds and tins. Logistics has a different calculation: Forecast A’s extra dozen bulky pans can take up a very different amount of truck space from its extra four molds and three nesting tins. Once the buyers have chosen, hand the transport team the actual pack dimensions and order quantities.