Demand Forecast Accuracy, WAPE, MAPE & Bias Calculator

Finin2min two-minute explanation

Measure actual versus forecast demand using WAPE, MAPE, signed forecast bias and RMSE with correct zero-demand handling. All results are management scenarios, not automatic accounting, contractual, audit or tax determinations. Validate each data source and period before relying on the numbers.

What the metrics answer

Comparing a forecast against past actual demand gives management a way to check planning quality. The input data must be an earlier frozen forecast and actual fulfilled or demanded units for exactly the same dates and items. A forecast revised after the period ended is not a valid backtest. The calculator reports unit errors and percentage errors separately to avoid turning a statistical metric into a spurious financial loss.

WAPE definition and interpretation

WAPE equals the sum of absolute forecast error divided by the sum of actual demand. Every row contributes its absolute unit miss to the numerator. Unlike signed bias, excess forecasts in one period cannot cancel shortfalls elsewhere. The metric is often useful for a group of different demand periods because high-volume dates contribute more to the denominator. However, the result must not be sold as a guarantee of future forecast accuracy.

Zero actuals and MAPE

Individual absolute percentage error divides by observed demand, which is impossible where actual demand is zero. This implementation explicitly excludes zero-actual observations from the MAPE average while still counting their unit errors in WAPE when total demand is positive. The number of excluded observations is displayed so a high-intermittency product is not misrepresented as consistently predictable. If every actual value is zero, WAPE is undefined rather than 0 percent.

Signed bias and demand risk

Here, signed error is forecast minus actual. A positive signed value suggests systematic overforecasting, which can increase slow-moving stock and working-capital needs. A negative value indicates underforecasting, which can encourage stockouts, expediting, or missed selling opportunities. Signed errors can cancel; therefore a small bias cannot replace WAPE or row-level exception analysis. Examine the expensive directions of error separately before changing reorder policy.

RMSE and outlying errors

RMSE squares each observed error, averages the squared values and then takes the square root. As a result, a small number of large forecasting misses can dominate the score. This is appropriate when those misses are disproportionately harmful but it may also emphasise isolated anomalies. Compare RMSE only for forecasts with compatible units and time steps, and never mix unit errors with revenue errors in the same report.

Six-week worked example

The sample has weekly actual demand of 120, 180, 0, 220, 170 and 210 units. Its corresponding forecasts are 140, 160, 25, 215, 190 and 200. Total observed demand is 900 units and total absolute error is 100 units, giving WAPE of 11.11%. Week 3 contributes 25 units to the error total but is excluded from MAPE because its actual demand is zero.

Management review checklist

Validate forecast snapshot dates; ensure returns, cancellations and promotions are treated consistently; remove duplicate lines; and group unlike products separately. Investigate the rows with the largest absolute misses before averaging metrics. Keep a baseline forecast such as previous-period demand for comparison. Classify whether the cost of underforecasting is materially different from the cost of overforecasting: an apparently accurate average may be economically poor if one error direction is much more expensive.

Limits and appropriate use

The calculator is a descriptive scorecard, not a forecasting algorithm, demand-probability distribution or stock-optimisation engine. It does not infer lead time, safety stock, supplier fill rate, service-level probability or the financial cost of errors. A professional analyst should reconcile source demand, time buckets and SKU mappings to the planning system before presenting recommendations. All data remain in the browser and are not transmitted to the provider.

Official sources and further reading

• AWS – Evaluating Predictor Accuracy

Sources explain the calculation conventions and rules referred to above. Notifications and standards change; confirm on the official site before you file, book or advise.