Enhancing forecast accuracy for lumpy demand using hybrid machine learning model

Abstract

Demand forecasting is a critical aspect of supply chain management, underpinning decision-making processes that span from strategic operations planning to daily workload management. Given its importance, substantial research efforts have been devoted to developing and optimizing forecasting tools to enhance supply chain efficiency.

However, much of this research has focused on regular demand due to its prevalence, leaving a significant gap in addressing irregular and sporadic demand patterns. These patterns, known as intermittent or lumpy demand, are characterized by frequent zero-demand intervals interspersed with unpredictable spikes, posing substantial challenges in inventory management. This is especially problematic for the retail industry, where profit margins are narrow and slow-moving, high-value, or niche products often experience lumpy demand.

Traditional forecasting methods, such as Croston’s method, have been widely used for these scenarios but fall short in achieving high accuracy due to the difficulty in predicting both the timing and magnitude of demand spikes. To overcome these limitations, this research introduces a hybrid forecasting model that integrates Croston’s method with a Boosting framework — a machine learning technique that iteratively corrects residual errors to enhance predictive performance. By applying this hybrid approach to real-world retail data, the study demonstrates improved accuracy in forecasting lumpy demand, offering retailers a robust tool to better manage unpredictable demand patterns. This model not only reduces the risk of stockouts and overstocking but also addresses a critical gap in retail demand forecasting, providing a practical and effective solution for handling erratic, low-frequency demand.