Forecast Bias

Bias is measured on signed errors, not absolute ones: the common form sums (forecast − actual) over a period and divides that by the sum of actuals. A tracking signal – cumulative signed error divided by the mean absolute deviation – shows when a series has drifted out of control. Accuracy metrics carry no direction: a forecast can score well on MAPE and still be biased, and a forecast with zero bias can be badly inaccurate. The two are read together, never one instead of the other.

• Bias (%) = Σ (forecast − actual) ÷ Σ actual × 100
• Tracking signal = cumulative (forecast − actual) ÷ mean absolute deviation
• Positive value = over‑forecasting, negative value = under‑forecasting

Forecast bias in practice

What bias costs depends on whether the planning loop sees inventory. In replenishment that orders up to a target level from the current inventory position, a persistent +5 % bias lifts that level by roughly 5 % of the demand covered by lead time plus the review period: stock settles higher and stays there. Where that feedback is missing – push allocation to stores, pre‑builds ahead of a promotion, production plans derived from a commercial target – the same 5 % is added again every cycle, until the surplus is marked down or written off.

Bias is asymmetric in how visible it is. Over‑forecasting ends up on a shelf where someone counts it; under‑forecasting ends up as a stockout, and a stockout censors the sales record. Measure accuracy against sales rather than demand and the periods when the forecast was too low look like periods when it was right, so reported bias reads closer to neutral than it is. Veritico STOCK accounts for stockouts and predicts actual demand instead of fitting the model to censored sales, and it cleans the history of outliers, stockouts and promotional effects before the forecast is computed – see demand forecasting and inventory optimization and replenishment and allocation management.

Two things hold for measurement. Aggregation hides bias: items biased in opposite directions net out, the total looks clean and every item underneath is still off, so measure per item and report the share of items outside a threshold. And keep the statistical output apart from the number that enters the plan – once a forecast has passed through manual overrides and approvals, a different series is being judged. At Mondelez, unifying planning across functions moved forecast accuracy from 50 % to 70 %.

Related terms

• Forecast Accuracy
• MAPE
• Demand Planning
• Demand Forecasting
• Seasonality
• Intermittent Demand