Dynamic demand forecasting and ordering for new products considering the influence of inventory shortages
针对电商新产品，提出一个预测后优化的框架，通过新方法GCWR平衡分类与回归误差，并考虑缺货对需求的影响，在京东数据集上验证了预测误差降低6%-9%、利润提升4.3%。
Motivated by the practice of a self-operated Chinese e-commerce platform, this article improves demand forecasting and ordering decisions for new products by proposing a predict-then-optimise framework that takes stockouts into account. First, we compare two clustering-based prediction methods, Cluster-While-Regress (CWR) and Cluster-Then-Regress (CTR), and find out a tradeoff between classification and intra-cluster regression errors. Then, we introduce a new method, Generialized-Cluster-While-Regress (GCWR), which balances this tradeoff through a hyperparameter and is theoretically proven to outperform both CWR and CTR when the data size is sufficiently large. Recognising the significant impact of stockouts on both demand and its forecasting accuracy, we explicitly incorporate stockouts into a multi-period ordering optimisation problem. We employ GCWR to model stochastic and nonlinear demand, capturing complex real-world scenarios. The resulting dynamic and non-convex optimisation problem is highly difficult to solve, and thus we propose an efficient data-driven one-step lookahead approximate solution. Case studies using JD.com's dataset show that GCWR can reduce prediction error by 6%–9% compared to CWR and CTR on average. Meanwhile, our data-driven method can increase average profits by 4.3% compared to the current practice of JD.com, demonstrating the superiority of our proposed framework.