Almost every demand planning team reports a 'forecast accuracy' figure, and it often comes from MAPE (mean absolute percentage error). The trouble is that MAPE has structural weaknesses that, if unrecognized, make bad decisions look justified: it blows up near zero, it weights low-volume items the same as high-volume ones, and it is asymmetric between over- and under-forecasting.
MAPE per period is |error| / actual sales, then averaged. If actual sales in a month are zero the value is undefined; if actual is 1 unit and the forecast was 5, the error is 400 percent and can swing the whole portfolio average, although the business impact is 4 units. As a result an aggregated MAPE reflects small, unimportant items more than the items that actually drive capital or revenue.
WAPE (weighted absolute percentage error) fixes this: the sum of absolute errors divided by the sum of actual sales. Every unit of error counts equally, so high-volume items naturally matter more. If monetary value matters, you can multiply error and sales by price or margin so the metric aligns with financial importance.
But accuracy alone is not enough; bias must be measured separately. Bias is the sum of signed errors divided by the sum of actual sales. A forecast can be 80 percent accurate yet consistently 10 percent too high; that pattern signals systematic behavior such as sales-team optimism, and its direct result is persistent excess inventory. Random error can be covered with safety stock, but bias can only be removed by fixing the process.
Another practical point is the measurement horizon. Measure accuracy at the lead-time horizon — the forecast that existed when the ordering decision was made — not the current-month forecast that was updated when sales were nearly known. Many organizations report accuracy from the latest forecast version and so see an optimistic number that has nothing to do with the quality of the real decision.
A fourth, less-used metric is forecast value added: did each step in the forecasting process, such as manual adjustment by the sales team, improve accuracy over a simple baseline like a moving average or a naive forecast? If a manual override makes error worse, that step costs effort and creates no value. This analysis tends to move the discussion from 'our forecast is inaccurate' to 'which step adds no value'.
A common mistake is setting one accuracy target for every item. Stable (X) items can be forecast with high accuracy, but intermittent (Z) items are inherently unpredictable, and an aggressive target only encourages gaming the number. Set targets by ABC-XYZ class and decision horizon, and for volatile items shift focus from 'improve the forecast' to 'shorten lead time and add flexibility'.
A practical starting point: this month, calculate WAPE and bias at the lead-time horizon for your top 20 items by value and put them next to your current MAPE. If the gap is large or you see persistent positive bias, examine the process and incentives behind the number before investing in a new tool or model.