Spare Parts Demand Prediction by Using a Random Forest Approach

Paper in proceeding, 2023

Spare parts forecasting, in general, is a complex task, due to its intermittent and erratic demand patterns. Furthermore, the underlying reason for the demand is usually not considered since most forecasting methods are based on demand history, merely. Spare parts demand is dependent on the need for replacement of components due to repair or maintenance reasons, which varies due to, e.g., life cycle, utilization of the finished product, and the number of finished products. Although machine learning methods have been more prevalent for spare parts forecasting in recent years, the prediction of initial demand, i.e., what to be stocked before the breakdowns and maintenance needs occur is understudied. So, the purpose of this paper is to investigate if the spare part need can be predicted before the demand occurs and, by that, increase the availability to the customers and decrease the cost of unavailability, e.g., expediting costs and lost sales. By adopting a machine learning model based on decision trees, called Random Forest, we predicted the probability of initial sales based on the installed base, and categorical variables such as product group, vital code, function, and weight, for spare parts at an automotive company. The analysis was made for three different markets in the Asia-pacific region and predictions were made for three different time horizons, 1, 3, and 6 months and the model performance shows potentially good results with an accuracy of around 70%. We also analyzed the business impact concerning availability and supply chain-related costs where, for example, we obtained a substantially lower total supply chain cost.