Stock-Out Distortion and Under-Forecasting: An Inventory-Aware Diagnosis of Demand-Signal Bias in Multi-SKU Sales Forecasting

Abstract

Observed sales are only an imperfect proxy for demand. When an item is out of stock the recorded quantity falls even though customers still want it, and the signal used to train forecasting models is distorted as a result. This study diagnoses how that distortion relates to systematic under-forecasting and inventory risk in a heterogeneous lubricant and grease portfolio (PT XYZ, 33 SKUs, 2021-2025, 396 test observations under a fair-comparison protocol). Stock-availability patterns expose periods of sharp depletion where the demand signal is likely distorted, and an inventory-aware analysis ties these periods to forecast behaviour. Across SARIMA, XGBoost with a Tweedie objective, and CNN-LSTM, every model carries a negative bias (-2.57, −3.08, −4.80) and under-forecasts more often than it over-forecasts, most strongly for XGBoost (61.4%) and CNN-LSTM (57.6%). A group-wise analysis shows that this under-forecasting concentrates in regularly selling SKUs-where stock-outs destroy real demand-and weakens on intermittent SKUs whose zeros are largely structural; the trained models also amplify the negative bias relative to naive baselines. A computational-cost profile completes the comparison. The contribution is fourfold: a reading of stock-availability patterns to locate likely-distorted periods, a diagnosis of systematic under-forecast bias linking stock-out distortion to stock-out risk, a robustness analysis across demand regimes, and a computational-cost characterisation to inform model choice in demand planning.