Forecast accuracy (MAPE)
MAPE (mean absolute percentage error) divides each period's demand forecast error by actual sales and averages those percentages; forecast accuracy is stated as 100 − MAPE.
Also calledMAPE, mean absolute percentage error, tahmin doğruluğu
Formula
Σ (|Actual − Forecast| ÷ Actual) ÷ n × 100- Actual
- Actual sales for the period; on stockout days, demand corrected for lost sales.
- Forecast
- The forecast for the same period that was live when the order was placed.
- n
- Number of periods (or SKU-periods) measured.
Calculator
Enter your own figures and the result updates instantly. The defaults are the figures from the worked example below.
Result%15,1
Worked example
How to read it, and what good looks like
MAPE gives every period's percentage error equal weight. That works while volumes are similar; it misleads once they diverge: on a product selling three units a month, a three-unit miss counts as 100% and buries a good forecast on a product selling a thousand. That is why MAPE is used for one product's accuracy across periods, while slow sellers are tracked separately and error is also reported in units.
"What should forecast accuracy be?" depends on the product's demand pattern; instead of a fixed threshold, use a simple comparison. Calculate a naive forecast for the same periods (last week's sales is this week's forecast) and compare the model's error with it. If the model cannot beat the naive forecast, it is not adding value. Measure at the level where orders are placed and on the forecast made one lead time ahead.
The database calculates past sales; the forecasting model supplies forward demand; accuracy is the comparison of the two at the same level. In Mindpera, the demand signal feeds Mind Send's initial allocation, replenishment and transfer proposals and Mind Stock's stock requirement; because forecast and actual sales sit in the same data model, error is calculated with the same formula for every SKU and store, and any product over threshold lands on the planner's task list. Where forecast and sales series live in separate systems, installation brings them into a single model.
Where it misleads
- In a period with zero sales the actual in the denominator is zero, so MAPE is undefined (the calculator shows "—"). Dropping those periods makes the measurement optimistic. For slow sellers, track error in units rather than as a percentage.
- MAPE punishes over-forecasting harder than under-forecasting: an under-forecast error is capped at 100%, an over-forecast has no cap. A model chosen on MAPE forecasts systematically low and creates stockouts.
- Accuracy is measured at total level, where positive and negative errors cancel out. A forecast that looks good in total shows a far larger error at SKU and store level.
- Sales on stockout days stay at zero; the forecast was right, the error looks large and the model is corrected in the wrong direction.
Data you need
- Actual sales by SKU, store and week
- The forecast at the same level that was live at order time (not a later revision)
- Stockout day records and promotion flags
Mindpera products that calculate this KPI
Mind Send
Make your shipments manageable. Depot-to-branch dispatch, branch-to-branch transfer and the initial allocation decision are produced together with the demand forecast.
Mind Stock
Stop treating stock accuracy as something that waits for a count and a month-end close. See movements in and out, expiry dates and profitability on one screen.
Frequently asked questions
- What is a good forecast accuracy?
- Instead of a fixed threshold, compare with a naive forecast. Calculate the error of the naive method, which takes last period's sales as this period's forecast; if the model's error is not lower, it is not adding value. Run the comparison at the SKU and store level where orders are placed, on the forecast made one lead time ahead.
- How do you calculate MAPE when sales are zero?
- You cannot: percentage error is found by dividing by actual sales, and division by zero is undefined. Dropping those periods makes the measurement optimistic. For slow sellers, track error in units rather than as a percentage. Where sales were zero because of a stockout, record corrected demand as the actual.