What is MAPE and where it misleads the retail operator
MAPE, mean absolute percentage error, is the average percentage error of a demand forecast. It answers one question: on average, by what percentage did the forecast miss actual sales?
Don’t confuse it with bias. MAPE measures the size of the error, not its direction. A forecast that misses 10% high and 10% low has high MAPE and zero bias.
How the operation measures it
The math is simple. For each SKU in each store, take the week’s absolute error, divide by actual sales, make it a percentage. Then average every point.
The forecasting system prints this number per SKU, per store, per category, per chain. Most chains read only the aggregate. That is where the trouble starts.
What the number hides
MAPE has a known flaw: it explodes when actual sales are small. A SKU that sells two units, with a forecast of four, has a MAPE of 100%. The same two-unit miss on a SKU selling two hundred is 1%.
The average blends the two and lies about both. In practice, the chain optimizes aggregate MAPE, it drops from 30% to 22%, and shelf-out on the items that matter stays flat.
The number improved on low-volume SKUs, which are many and weigh little. The ones that pay the bills, high-velocity and perishable, were left out of the gain.
A healthy aggregate MAPE coexists with chronic shelf-out on the twenty SKUs that carry the category.
Why this changes an AI project
When the success criterion is “lower the MAPE”, the project ships the number and not the operation. The right indicator is weighted by volume or margin, not by SKU count.
WMAPE, the weighted version, fixes part of this. The real fix is choosing which SKU the forecast has to get right. It is the same reading error we describe in stock cover and in service level.
Measuring it wrong here is expensive. The operation trusts a number that climbs while the shelf stays empty. It is the kind of problem we step in to solve with criterion, not with another dashboard.
Take the number your planning team reports as “forecast accuracy”. Is it aggregate? Weighted by what? If you cannot answer in ten seconds, the number is already misleading you.
Think about the term each area of your operation uses but defines differently. “Accuracy” is usually that term. Tell me yours. In one hour I will send back a one-page card: the formula, the measurement points, what the number actually means, and where it misleads. If the card shows the system producing that number has to be rebuilt, the two-week Diagnóstico is the next step.