Authors: Ivan Svetunkov and Anna Sroginis

Abstract: Effective demand forecasting is critical for inventory management, production planning, and decision making across industries. One of the main challenges lies in understanding the origins of zeros in demand, which may happen naturally or due to some anomalies such as stockouts and recording errors. Misinterpreting these zeroes can lead to the application of inappropriate forecasting methods, leading to poor decision making. To address this, we propose a two-stage model-based classification framework that identifies artificially occurring zeroes in the first step and then distinguishes between different types of demand in the second one. The identified anomalies and demand types are then used as features for forecasting approaches. Simulations reveal that our approach significantly increases the demand forecasting accuracy. The empirical study demonstrates how the features can be used in forecasting with LightGBM and regression, leading to the increase in accuracy of the forecasting methods.