Forecast accuracy measurement is one of those topics where the theoretical answer and the practical answer diverge significantly. In a graduate forecasting course, MAPE is introduced as the standard metric. In a mid-market distribution operation with thousands of SKUs, MAPE alone gives you a number that is difficult to act on and can actively mislead you about model performance. The choice of metric shapes what the forecasting team optimizes for, which shapes what happens to inventory. It is worth getting right.
This piece covers the three metrics most commonly used in demand planning contexts - MAPE, WMAPE, and sMAPE - describes where each one fails, and provides a practical framework for choosing based on what you actually need to manage.
MAPE: The Standard That Fails on Intermittent Demand
Mean Absolute Percentage Error (MAPE) is computed by averaging the absolute percentage error across all forecast periods. It is interpretable, comparable across SKUs with different demand scales, and widely understood.
The critical failure: MAPE is undefined when actual demand is zero, and it produces extreme values when actual demand is very small. A forecast of 10 units against an actual of 1 unit produces a 900% MAPE for that period. If your catalog has any items with intermittent demand - and almost all mid-market distributor catalogs do - MAPE will be dominated by these extreme error periods. The aggregate MAPE number will tell you something about your worst-performing intermittent items, not about the overall quality of your forecasting process.
The standard workaround is to exclude periods with zero actual demand from the MAPE calculation, or to exclude items with too many zero-demand periods entirely. But this exclusion is itself a data quality decision that needs to be made explicitly, not buried in the metric definition. A MAPE reported on only 60% of your SKUs because the other 40% had too many zero-demand periods is not comparable to a MAPE reported on the full catalog.
Where MAPE works: high-velocity items with stable, non-zero demand and relatively consistent weekly volumes. If your top 200 SKUs all sell at least 20 units per week, MAPE gives you a clean, comparable accuracy picture across those items. This is the context it was designed for.
WMAPE: Solves the Zero Problem, Creates a Bias Problem
Weighted Mean Absolute Percentage Error (WMAPE) - sometimes written as wMAPE - addresses the intermittent demand problem by weighting each error by actual demand. The formula is the sum of absolute errors divided by the sum of actuals, expressed as a percentage.
This eliminates the division-by-zero problem. A period with zero actual demand contributes zero to both the numerator and denominator. The metric is no longer undefined or extreme for low-demand periods. It is mathematically stable across the full demand range.
The tradeoff is a systematic bias toward high-volume items. Because the weighting is by actual demand, forecast errors on high-volume SKUs dominate the aggregate WMAPE. A 10% error on an item that sells 500 units per week moves the metric far more than a 10% error on an item that sells 5 units per week, even though both represent identical proportional forecast quality.
For a distribution company where the top 5% of SKUs account for 60% of volume, WMAPE is essentially reporting the accuracy of the top-volume forecast. The long tail of the catalog - which may represent significant inventory value and service level exposure even if it contributes little to total volume - is nearly invisible in the metric.
Where WMAPE works: when the business objective is optimizing accuracy on high-revenue items and the long tail is genuinely less important. For a specialty distributor where the top 100 items drive 80% of revenue and the remaining 900 items are fill-in business, WMAPE is a reasonable primary metric. For a MRO distributor where a long catalog of low-velocity parts is critical to customer relationships, it is not.
sMAPE: Symmetric but Asymmetric in Practice
Symmetric Mean Absolute Percentage Error (sMAPE) was designed to address a specific MAPE problem: the fact that percentage errors are not symmetric. A 100% overforecast and a 100% underforecast produce different MAPE values because the denominator differs. sMAPE uses the average of actuals and forecasts as the denominator, making over- and under-forecasting errors mathematically symmetric.
In theory, this is an improvement. In practice, sMAPE has its own edge cases that make it problematic in demand planning contexts. First, when both actuals and forecasts are close to zero, sMAPE is highly sensitive to small absolute differences - a forecast of 1 versus an actual of 0 produces very different sMAPE than a forecast of 100 versus an actual of 99. Second, sMAPE does not behave intuitively when forecasts and actuals are of very different magnitudes.
More practically: sMAPE does not tell planners or management anything actionable that MAPE or WMAPE does not. It is harder to explain and interpret, and the symmetry it provides is most useful in academic model comparison contexts where bias detection matters, not in operational planning contexts where the primary question is "how far off is the forecast and in which direction."
Where sMAPE works: research contexts where symmetric treatment of errors is necessary for fair model comparison. Operational demand planning is not that context.
What Actually Matters for Planners
The most important thing to understand about forecast accuracy metrics is that no single metric tells the full story. Planners need two pieces of information that aggregate metrics obscure: bias direction and accuracy by SKU tier.
Bias direction is whether the forecast systematically over- or under-predicts. MAPE, WMAPE, and sMAPE all produce absolute error values - they tell you the magnitude of the error but not the direction. A forecasting system that consistently underforecasts by 12% (which leads to stockouts) and one that consistently overforecasts by 12% (which leads to overstock) have identical MAPE values. These are not equivalent problems. Mean forecast error (MFE), also called mean bias, should be reported alongside accuracy metrics to detect systematic directional errors.
Accuracy by SKU tier is the practice of computing forecast accuracy separately for high-velocity, medium-velocity, and low-velocity items rather than reporting a single aggregate metric. A distribution company with a 14% overall WMAPE might have 8% WMAPE on its top-tier items and 38% WMAPE on its medium-tier items. The aggregate number hides a medium-tier problem that is generating disproportionate stockout and overstock costs.
A Practical Metric Framework for Mid-Market Distributors
Based on how demand planning teams actually use forecast accuracy data, the following combination covers most use cases without adding unnecessary complexity.
For high-velocity items (top 20% by annual demand, non-intermittent): MAPE by SKU, reported weekly. This tier drives most inventory cost and service level impact. Individual SKU visibility matters. MAPE is interpretable at this level.
For medium-velocity items: WMAPE by category or sub-category, reported monthly. Individual SKU MAPE is too noisy at this volume level. Category-level WMAPE provides a usable signal without the false precision of per-SKU metrics on moderate-volume items.
For all tiers: MFE (mean bias) reported alongside the accuracy metric, flagging any tier or category with systematic directional error above 5%. Systematic bias is a model problem that deserves investigation regardless of overall accuracy level.
For intermittent items: track fill rate and stockout frequency rather than forecast error metrics. These items are too sparse to produce reliable MAPE or WMAPE numbers, but the operational outcomes - how often they stocked out, how often there was excess - are directly measurable and actionable.
The goal is not metric optimization. It is operational outcome measurement. The right metric framework is the one that surfaces actionable information about where forecast quality is creating inventory problems, not the one that produces the most impressive-looking aggregate number.
