基于机器学习的多时间粒度需求预测研究——以某制造企业供应链管理为例

Research on Machine Learning-Based Demand Forecasting across Multiple Time Granularities—A Case Study of Supply Chain Management in a Manufacturing Enterprise

摘要: 本文基于国内某大型制造企业在2015年9月1日至2018年12月20日面向经销商的出货数据，探讨了在全球供应链不确定性加剧与数字化转型加速的背景下，企业需求预测面临的多源异构数据整合、非线性波动(如疫情引发的“牛鞭效应”)和跨粒度决策协同等挑战。本文通过对比XGBoost、LightGBM和随机森林(RF)模型在月、周、日三种时间粒度下的预测性能。研究发现，模型精度随预测频率提高呈系统性衰减，并由此提出构建多粒度联合优化框架的必要性，以实现不同决策层级的精度与资源分配平衡。

Abstract: Based on the shipment data of a large domestic manufacturing enterprise to distributors from September 1, 2015 to December 20, 2018, this article explores the context of intensified global supply chain uncertainties and accelerated digital transformation. Enterprise demand forecasting faces challenges such as the integration of multi-source heterogeneous data, nonlinear fluctuations (such as the “bullwhip effect” triggered by the epidemic), and cross-granularity decision-making collaboration. This paper compares the prediction performance of XGBoost, LightGBM and Random Forest (RF) models at three time granularities: month, week and day. The research finds that the model accuracy systematically decays with the increase of the prediction frequency. Based on this, the necessity of constructing a multi-granularity joint optimization framework is proposed to achieve the balance of accuracy and resource allocation at different decision-making levels.