The inventory management and production planning of parts with irregular demand patterns are challenging for manufacturing companies. These patterns often occur in the strategically critical spare parts sector, where the inventory and capital commitment costs are high. For this reason, an accurate forecast can improve service levels and ensure efficient stock keeping. For this problem, time-series-based forecasting methods are often used to predict future demands. Furthermore, the research of recent years in terms of stochastic forecasting also focused on Artificial Intelligence (AI) methods, mainly Artificial Neural Networks (ANN). In contrast to previous studies, this paper compares the prediction results of various ANN configurations and classical forecasting methods for all of the different demand categories according to Syntetos et al. [1], which means that erratic, lumpy, smooth, and intermittent demands are regarded separately. This study compares eleven statistical forecasting configurations with eight single hidden layer neural network configurations. Furthermore, the influence of the number of hidden neurons on the prediction performance is investigated with the learning algorithms Backpropagation (BP) and Levenberg-Marquardt (LM) by evaluating them separately, which has not been covered in the context of all irregular demand categories yet. The study is based on actual demand data from 29 spare parts of a mechanical engineering company.