Weighted Mean Absolute Percentage Error (WMAPE)

WMAPE addresses the flaw in standard MAPE by weighting each item's error by its actual demand volume. This prevents low-volume, high-percentage-error items from distorting the overall accuracy metric.

• Formula: WMAPE = Σ(|Actual - Forecast|) / Σ(Actual)
• Weighting Factor: Each absolute error is implicitly weighted by the item's actual value
• Result: A single percentage representing the total portfolio's error relative to total demand
• Interpretation: A 10% WMAPE means total forecast error equals 10% of total actual demand

WMAPE aligns forecast accuracy measurement with business impact by prioritizing high-revenue or high-volume items. An error on a top-selling SKU contributes proportionally more to the metric than an error on a slow-moving item.

• Revenue Alignment: Directly correlates with financial impact of forecast errors
• Inventory Optimization: Guides safety stock calculations for items that matter most
• Executive Reporting: Provides a single, defensible number for supply chain performance
• Trade-Off Visibility: Reveals whether accuracy gains come from high-volume or low-volume items

WMAPE sits between Mean Absolute Percentage Error (MAPE) and Mean Absolute Error (MAE) in its properties. It retains the percentage interpretability of MAPE while avoiding its infinite values and small-denominator distortions.

• vs. MAPE: WMAPE is not skewed by items with near-zero actuals; MAPE can explode to infinity
• vs. MAE: WMAPE provides a relative, scale-independent percentage; MAE is in raw units
• vs. RMSE: WMAPE is less sensitive to large outliers than Root Mean Squared Error
• Limitation: WMAPE is undefined when total actual demand is zero across the entire portfolio

A critical limitation of WMAPE is its behavior with zero actual demand. While individual items with zero actuals do not cause division-by-zero errors, they can mask forecast inaccuracies.

• Mechanism: Items with zero actual demand contribute zero to the denominator Σ(Actual)
• Risk: A forecast of 100 units for an item with 0 actual demand adds 100 to the numerator but 0 to the denominator
• Mitigation: Filter out items with zero actuals and zero forecasts before calculation
• Alternative: Use complementary metrics like MAE for intermittent demand items

WMAPE is naturally coherent across aggregation levels, making it ideal for evaluating hierarchical forecasts. The weighted sum property ensures that category-level WMAPE can be derived from SKU-level errors.

• Bottom-Up Consistency: WMAPE at the aggregate level equals the volume-weighted combination of lower-level WMAPEs
• Roll-Up Integrity: Total portfolio WMAPE is mathematically consistent with individual category WMAPEs
• Drill-Down Capability: Enables root-cause analysis by decomposing aggregate error into component contributions
• Use Case: A retailer can identify which product category contributes most to total forecast error

WMAPE is straightforward to implement in production forecasting systems and serves as a primary evaluation metric during model selection and monitoring.

• Calculation: Requires only actuals and forecasts vectors; no complex distributional assumptions
• Model Selection: Use WMAPE as the objective for hyperparameter tuning when business impact matters
• Monitoring: Track WMAPE over time to detect concept drift or data quality degradation
• Thresholds: Typical industry benchmarks range from 10-20% WMAPE for retail demand forecasting at the SKU-week level

The foundational metric from which WMAPE is derived. MAPE calculates the average of the absolute percentage errors for each forecasted item, expressing accuracy as a percentage.

• Formula: (1/n) * Σ |(Actual - Forecast) / Actual| * 100
• Key Limitation: Treats all items equally, regardless of their business volume or revenue impact.
• Infinite Error: Produces undefined or infinite values when actual demand is zero, making it unsuitable for intermittent demand patterns.
• Use Case: Simple, intuitive communication of average error when all items have similar business importance.

A scale-dependent accuracy metric that measures the average magnitude of forecast errors in the original units of the data, without considering direction.

• Formula: (1/n) * Σ |Actual - Forecast|
• Unit Dependency: Error is expressed in the same units as the data (e.g., units sold, dollars), making cross-series comparison impossible without normalization.
• Robustness: Unlike MAPE, MAE handles zero actual values without producing infinite errors.
• WMAPE Relationship: WMAPE can be viewed as a weighted version of MAE, normalized by the total actual volume to produce a percentage.

The systematic tendency of a model to consistently over-forecast or under-forecast, measured as the mean of the raw forecast errors (not absolute values).

• Formula: (1/n) * Σ (Forecast - Actual)
• Positive Bias: Indicates persistent over-forecasting, leading to excess inventory and holding costs.
• Negative Bias: Indicates persistent under-forecasting, causing stockouts and lost sales.
• Complementary Metric: WMAPE measures dispersion (magnitude of error), while bias measures direction. A model can have a low WMAPE but a dangerous bias if errors consistently lean one way.

A modified percentage error metric designed to address MAPE's asymmetry and infinite error problems by using the average of actual and forecast values as the denominator.

• Formula: (1/n) * Σ |Actual - Forecast| / ((|Actual| + |Forecast|) / 2) * 100
• Bounded Range: Produces values between 0% and 200%, avoiding the unbounded upper limit of MAPE.
• Criticism: Despite its name, SMAPE is not truly symmetric; it penalizes under-forecasts more heavily than over-forecasts for the same absolute error.
• WMAPE Advantage: WMAPE avoids SMAPE's interpretability issues by using only actual values as the denominator, providing a clearer business interpretation of error relative to realized demand.

A quadratic scoring rule that measures the square root of the average squared differences between forecasts and actuals, heavily penalizing large errors.

• Formula: √[(1/n) * Σ (Actual - Forecast)²]
• Large Error Sensitivity: Squaring the errors means RMSE is disproportionately influenced by outliers, making it ideal when large forecast misses are disproportionately costly.
• Unit Dependency: Like MAE, RMSE is expressed in the original data units.
• WMAPE Comparison: WMAPE uses absolute errors (linear penalty), making it more robust to outliers than RMSE. Choose RMSE when the business cost of error grows quadratically with magnitude.

A monitoring statistic used to detect when a forecasting model's bias has drifted beyond acceptable limits, triggering an alert for model review or retraining.

• Formula: Σ (Actual - Forecast) / MAD, where MAD is the Mean Absolute Deviation.
• Threshold Triggers: Typically, a tracking signal exceeding ±4 indicates a statistically significant bias has emerged.
• Operational Role: While WMAPE is a periodic aggregate accuracy measure, the tracking signal provides a continuous, real-time check on model health.
• Supply Chain Application: Used in demand planning systems to automatically flag SKUs whose forecasts have become systematically unreliable, preventing silent degradation of inventory decisions.