Accurate demand forecasting is critical and difficult for managers, especially for complex demand patterns. In this paper, we develop methods for demand forecasting of sparse, transient and erratic medical consumables. Firstly, combining statistical learning of historical data with basic linear regression, price discount estimates are proposed. To reduce sparse estimates, the transformation of historical demand data is added to the linear regression model. Secondly, some general methods are proposed to deal with demand patterns that we cannot clearly capture. Thirdly, we propose optimized model specifications to select optimal model and reduce redundant variables to avoid underfitting or overfitting. In the last, some numerical experiments are carried out based on the model we propose and some completive models in the actual demand data set. In this study, we develop the most realistic price discount response function based on the problem background, which can further improve demand forecasting performance. This paper also discusses many interesting findings and conclusions.

准确的需求预测对于管理者而言至关重要且困难重重，尤其是对于复杂的需求模式而言。在本文中，我们开发了稀疏，短暂和不稳定的医疗耗材需求预测的方法。首先，将历史数据的统计学习与基本线性回归相结合，提出了价格折扣估计。为了减少稀疏估计，将历史需求数据的转换添加到线性回归模型中。其次，提出了一些通用的方法来应对我们无法清晰把握的需求模式。第三，我们提出优化的模型规格，以选择最优模型并减少冗余变量，以避免拟合不足或过度拟合。最后，基于我们提出的模型和实际需求数据集中的一些完全模型进行了一些数值实验。在本研究中，我们基于问题背景开发了最切合实际的价格折扣响应函数，可以进一步提高需求预测性能。本文还讨论了许多有趣的发现和结论。