In this paper, new neural network (NN) methods are proposed to forecast intermittent demand and we empirically study their performance as compared to parametric and non-parametric forecasting methods proposed in the literature. The empirical investigation uses demand data for 5,135 spare parts for the fleet of aircrafts of an airline company. Three parametric benchmark methods are examined: single exponential smoothing (SES), Croston’s method and Syntetos–Boylan approximation, along with two bootstrapping methods: Willemain’s method and Zhou and Viswanathan’s method. The benchmark NN method considered in this paper is that proposed by Gutierrez et al. (2008) The paper shows the outperformance of SES and the NN methods for (a) their forecast accuracy and (b) their inventory efficiency (trade-off between holding volumes and backordering volumes) when compared to the other methods. Moreover, among the NN methods, a new proposed method is shown to be better than that proposed by Gutierrez et al. in terms of forecast accuracy and inventory efficiency.

中文翻译：

---

关于一些新的预测间歇性需求的神经网络方法的经验性能

本文提出了一种新的神经网络（NN）方法来预测间歇性需求，并且与文献中提出的参数和非参数预测方法相比，我们进行了实证研究。实证调查使用航空公司航空公司飞机机队的5135个备件的需求数据。研究了三种参数基准方法：单指数平滑（SES），Croston方法和Syntetos-Boylan逼近，以及两种自举方法：Willemain方法以及Zhou和Viswanathan方法。本文考虑的基准神经网络方法是由古铁雷斯等。（2008）本文显示了与其他方法相比，SES和NN方法在（a）预测准确度和（b）库存效率（库存量和缺货量之间的权衡）方面表现出色。此外，在NN方法中，一种新提出的方法被证明比Gutierrez等人提出的更好。在预测准确性和库存效率方面。