Forecast Accuracy

Accuracy is the complement of forecast error, so it inherits every property of the metric it is derived from. Two definitions dominate. One is based on MAPE, which averages percentage errors item by item and treats a small item and a large one as equally important. The other is based on WAPE, which divides total absolute error by total demand and therefore weights items by volume. For inventory and replenishment work the WAPE‑based figure is the safer default, because a single slow mover with a large percentage miss cannot distort it.

• Accuracy (WAPE‑based) = 1 − Σ|A − F| / ΣA, summed across items and periods
• Accuracy (MAPE‑based) = 1 − (1/n) Σ (|A − F| / A) – undefined whenever actual demand is zero
• Bias = Σ(F − A) / ΣA – the direction of the error, and a separate number that belongs next to accuracy
• State the grain and the horizon: “82 % at SKU × store × week, one lead time ahead” is a claim; “82 % accuracy” is not

Forecast accuracy in practice

Aggregation is the first thing to check on any reported number. A monthly, network‑wide, volume‑weighted figure lets errors offset each other: overshooting one store and undershooting another cancels out, and nothing in the total reveals that both orders were wrong. The same forecast measured at SKU × store × day, which is the grain a replenishment decision actually runs on, lands far lower. Neither number is false; they answer different questions, and only the second one predicts whether the shelf will be full.

Horizon is the second parameter. Accuracy measured at lag zero – this week's forecast for this week – describes a forecast nobody can order from. The number that matters is the one produced at the horizon of the lead time plus the review period, because that is when the order has to be committed.

The third issue is quieter and more expensive. When accuracy is measured against sales, the forecast is being compared with something other than demand. During a stockout sales are censored: the forecast said 40, the shelf ran out at 25, and the error is booked as an overforecast. Accuracy then looks best exactly where the company lost revenue. Days with zero availability have to be excluded from the calculation or the lost demand reconstructed. For slow movers with intermittent demand there is a limit to what any model can do, and percentage‑based metrics break down when actuals are frequently zero – there the work is done by safety stock and by shortening lead time, not by chasing accuracy. Accuracy is an input metric; availability, stock cover and write‑offs are the outcomes worth reporting to a board.

Veritico STOCK picks the forecasting method with the lowest expected error per item from roughly eighty candidates, cleans stockouts and promotional effects out of the history before it fits, and forecasts per item, warehouse and day – the grain the replenishment run consumes – see demand forecasting and inventory optimization and replenishment and allocation management. Mondelez moved forecast accuracy from 50 % to 70 % after unifying planning across functions; Kofola reached 76 % on promotional forecasts, where accuracy is structurally harder than on baseline demand.

Related terms

• Demand forecasting
• Demand planning
• Stockout
• Time Series Forecasting
• MAPE
• Forecast Bias