Advances in Intermittent Demand Forecasting

Forecasting intermittent and lumpy demand remains one of the most challenging problems in supply chain management due to its sporadic occurrence and high variability. Demand that occurs in bursts, characterized by frequent periods of zero demand interrupted by irregular, inconsistent nonzero values, introduces significant risks. Overestimating such demand can lead to excessive production and inventory costs while underestimating it can result in missed opportunities and longer lead times (Ivanov, 2017; Moon, 2018; Kourentzes, 2013). Traditional approaches, such as Croston’s method (Croston, 1972) and its variants (Willemain et al., 2004; Shale et al., 2006), have been widely used; however, their performance is limited by the unpredictable nature of these series. Lumpy demand, marked by considerable fluctuations in demand size (Mukhopadhyay et al., 2011a; Martin et al., 2020; Xu et al., 2012), adds another layer of complexity, further exacerbated by the historical absence of evaluation metrics that adequately account for temporal shifts and associated costs (Martin et al., 2020). This thesis addresses these gaps by first reviewing the fundamental characteristics of intermittent demand and the forecasting methods traditionally applied to such series, drawing on seminal studies (Croston, 1972; Willemain et al., 2004; Li and Lim, 2018). It then investigates the use of integer-valued autoregressive models to capture the dynamics of intermittent demand. Based on these insights, the research improves the accuracy of the forecast by integrating machine learning techniques with Croston’s method and further introduces a novel global Croston approach to better capture the complexities inherent in these demand patterns. Forecast performance is evaluated using advanced error metrics, such as scaled Cumulative Error (sCE) (Kourentzes, 2013), scaled Absolute Periods in Stock (sAPIS) (Wallström and Segerstedt, 2010), Mean Absolute Scaled Error (MASE) and Root Mean Squared Scaled Error (RMSSE) (Hyndman and Koehler, 2006), which collectively assess prediction accuracy, temporal dynamics, and the costs related to inventory management. Empirical analysis on three real-world datasets validates the strengths and limitations of each approach, offering actionable benchmarks for practitioners confronting the challenges of intermittent and lumpy demand.