Solving data-driven newsvendor problem with textual reviews through deep learning

The production decision of a large commodity or equipment manufacturing enterprise can be modeled as a newsvendor problem. Managers must determine the optimal production volume in advance to minimize the underage cost and the overage cost. However, the traditional newsvendor problem assumes the known demand distribution, which is not the case in practice. Data-driven approaches have become the hot research topic and opened up new avenues for such issues. Recent studies have considered demand-related features but have failed to address how to optimize production and inventory using informative textual reviews, not just numerical feature data. To address this issue, we propose a data-driven newsvendor model that leverages sentiment analysis on textual reviews using a deep learning model to solve the data-driven newsvendor problem by integrating estimation and optimization. Experiments on real data show that our proposed method reduces the average cost by approximately 14.18% compared to the most advanced deep neural network method, making it the best-performing method. Furthermore, our method is more suitable for situations where unit shortage costs are greater than unit overage costs. Finally, our method is robust in terms of sample size and can still obtain good results even with insufficient historical data.