MAPE vs MASE for Demand Forecast Accuracy Measurement

MAPE (Mean Absolute Percentage Error) excels at communicating forecast quality to business stakeholders because it expresses error in intuitive percentage terms. For high-volume, fast-moving consumer goods where demand is consistently above zero, MAPE provides a clear, weighted view of accuracy that aligns with revenue impact. However, MAPE catastrophically fails on intermittent demand patterns—when actual sales hit zero, the calculation divides by zero and becomes undefined. Even near-zero values produce infinite or massively skewed error percentages, making it impossible to fairly compare a slow-moving spare part against a top-selling SKU in the same portfolio.

MASE (Mean Absolute Scaled Error) takes a fundamentally different approach by scaling the forecast error against a naive baseline (typically the in-sample MAE of a seasonal random walk). This scale-independence means MASE handles zero values gracefully and produces comparable accuracy metrics across SKUs with wildly different demand volumes. For supply chains managing long-tail spare parts, aftermarket components, or highly seasonal products, MASE allows you to benchmark a forecasting model's added value over a trivial method. The trade-off is interpretability: a MASE of 0.8 tells a planner the model is 20% better than a naive forecast, but it does not directly translate to a percentage error they can intuitively grasp.

The key trade-off: If your priority is executive communication and your portfolio consists primarily of high-volume, continuous-demand items, MAPE remains the pragmatic choice despite its mathematical flaws. If you prioritize fair model comparison across a heterogeneous SKU portfolio containing intermittent or lumpy demand, choose MASE to avoid systematically over-forecasting slow-moving items. For the most rigorous evaluation, consider supplementing MASE with RMSSE (Root Mean Squared Scaled Error) to penalize large outliers more heavily, which directly translates to reduced safety stock requirements and lower inventory carrying costs.

MAPE (Mean Absolute Percentage Error) excels at intuitive communication because it expresses error as a simple percentage, making it easy for business stakeholders to grasp. For example, a MAPE of 15% is universally understood across planning, finance, and executive teams. However, this interpretability comes with a fatal mathematical flaw: MAPE is undefined when actual demand is zero and becomes infinite or wildly skewed when values approach zero. In supply chains where intermittent demand is the norm—such as spare parts, aftermarket components, and long-tail SKUs—this renders MAPE not just inaccurate, but structurally unusable for model comparison.

MASE (Mean Absolute Scaled Error) takes a fundamentally different approach by scaling the forecast error against a naive benchmark's in-sample error. This results in a scale-independent metric that works identically whether a SKU sells 10,000 units per week or one unit every six months. A MASE of 0.80 means the forecast is 20% better than a naive seasonal model, while a MASE above 1.0 signals the model is worse than a trivial baseline. Critically, MASE handles zero values without division-by-zero errors, making it the only reliable choice for portfolios with intermittent demand patterns.

The key trade-off: If your priority is communicating forecast accuracy to non-technical stakeholders and your portfolio consists entirely of high-volume, continuous-demand products, MAPE remains a viable, if imperfect, choice. However, if you need to fairly compare model performance across a heterogeneous SKU portfolio—especially one containing slow-moving items, new product introductions, or seasonal goods with zero-demand periods—MASE is the only statistically valid option. For demand planning directors evaluating AI forecasting models, MASE should be the default metric, with MAPE reserved only for simplified executive summaries where its limitations are explicitly acknowledged.