Comparative Analysis of Probabilistic Models for Intermittent Demand Forecasting

The distinguishing characteristic of intermittent demand compared to non-intermittent demand lies in the occurrence of zero values for certain time periods, defining intervals that transform it into a time series. Typically, a single interval value is chosen for all stock keeping units (SKUs) in software packages. When analyzing demand distribution on a SKU basis, demand is considered fast-moving if its frequency exceeds the selected interval, otherwise, it is classified as slow-moving or intermittent demand. Given the absence of a fixed interval in real-life scenarios, it is reasonable to assume that all demand distributions exhibit intermittent behaviour. This paper aims to apply relatively novel methods to forecast both expected demand and occurrences in the next forecasting period for a sample of automotive spare parts. Intermittent demand forecasting traces back to Croston’s work in 1972, which was based on single exponential smoothing, disregarding trend and seasonality and being prone to positive bias. To address these limitations, newer methods such as Holt-Winter, Syntetos and Boylan Approximation (SBA), and Teunter, Syntetos, and Babai (TSB) have been introduced. In this study, Croston’s method, SBA, and TSB methods are utilized for forecasting intermittent demand of automotive spare parts, with a comparative evaluation based on forecast error types including mean squared error, root mean squared error, and mean absolute error.