Development of demand forecasting models for sustainable supply chain management

需求管理是供应链管理的核心，但现有研究从算法开发和数据分析的角度对需求预测的研究不足。在实际操作中，需求预测的一个小误差就会使供应链上下游的信息严重错配，造成巨大的资金和资源损失，这就是众所周知的库存管理中的牛鞭效应。此外，为这种分析收集的数据可能是大的，薄的或不完整的（如删节或缺失）。从不同的数据类型中挖掘有价值的信息自然成为需要进一步探索的研究焦点。现有的需求预测算法包括经典的统计方法，如时间序列分析，如指数平滑方法和Box-Jenkins方法，以及机器学习方法，如神经网络和决策树。这些方法在对动态大数据建模时计算复杂。此外，基于单薄和不完整的数据来预测需求的研究也很少。因此，本研究旨在开发数据驱动算法，分别解决大数据和不完整数据下的需求预测问题，从而消除牛鞭效应，促进更有效的库存控制。在需求预测模型的开发中，将采用经典的统计方法和机器学习算法。经典方法包括泊松过程、反复标记时间点过程和动态调和回归；机器学习方法包括循环神经网络，如长短期记忆网络。此外，在开发预测模型时，将考虑在线客户评论、产品属性和其他外部影响因素，以调查可能的协变量的影响。同时，对缺失或删减数据的不完整样本进行分析，这与我在硕士学习期间的研究密切相关。最后，将提出的预测方法扩展到多级供应链和更长的预测范围。

Demand management plays a core role in supply chain management, but demand forecasting is insufficiently studied through the perspective of algorithm development and data analysis in the existing research. In practice, a small error in demand forecasting will make the information in the upstream and downstream of the supply chain seriously mismatched, resulting in a huge loss in finance and resources, which is the well-known bullwhip effect in inventory management. Besides, data collected for such analyses can be big, thin or incomplete (such as censored or missing). Mining valuable information from different data types has naturally become a focal point of research that needs further exploration. Existing algorithms for demand forecasting include classical statistical methods such as time series analysis including the exponential smoothing methods and the Box-Jenkins methods, and machine learning methods such as neural networks and decision trees. These methods are computationally complicated when modelling on dynamic big data. Additionally, there is little research exploring methods to forecast demands based on thin and incomplete data. This research therefore aims to develop data-driven algorithms to address the demand forecasting issue on big data and incomplete data, respectively, so as to eliminate the bullwhip effect and then facilitate more effective inventory control. In the development of demand forecasting models, both classical statistical methods and machine learning algorithms will be employed. The classical methods include the Poisson process, recurrent marked temporal point process and dynamic harmonic regression; and machine learning methods include recurrent neural networks such as the long short-term memory network. Further, online customer reviews, product attributes and other external impacting factors will all be considered in the development of forecasting models to investigate the impacts of possible covariates. Meanwhile, incomplete samples with missing or the censored data will be analyzed, which is closely related to my research during the master study. Finally, the proposed forecasting methods will be extended to the cases of multi-echelon supply chains and a longer forecast horizon.