An evaluation of forecasting methods for anticipating spare parts demand

In the planning process of a supply chain, demand forecast have an important role in planning process of a company. The forecasts have to be as accurate as possible in order to allow the optimization of production, avoiding extra stocking costs or lost sales. In the case of spare parts, the challenge arises as the demand presents intermittent behavior. Nowadays, many forecast techniques, namely ARIMA models and Croston´s method, are used to forecast spare parts demand. Alternatively to statistical methods, artificial neural networks (ANNs) are also powerful tools for solving non linear complex problems. Recurrent neural networks, such as the reservoir computing (RC), present an architecture that allows modeling of dynamic behavior, making it a strong candidate for solving the spare parts demand forecast problem. This work proposes a performance evaluation of artificial neural networks (feed-forward and recurrent) for spare parts demand forecasting, including also a comprehensive comparison of best achieving techniques, such as Croston´s and the ARIMA method. The evaluation performed here may help an organization to decide which technique suits better its needs for forecasting the spare parts demand. The experiments show that neural networks out perform the statistical methods in three out of four data sets tested.