Research on the Newsvendor Problem Based on Average Estimation of Deep Learning Models

Inventory control largely relies on forecasting future demand. Since market demand is often unstable, determining an appropriate or even optimal inventory level can not only reduce stock overages but also lower the likelihood of stockouts. In existing literature, future demand distribution is often assumed to be known, which is impractical in real-world scenarios. When the demand distribution is unknown, most approaches focus only on estimating the unknown parameters within a specific demand distribution. However, in the field of inventory management, there exists significant uncertainty regarding demand models. We cannot precisely determine the form of the demand function, nor can we identify which variables influence demand. To address the model uncertainty in inventory management, this paper utilizes model averaging methods and proposes deep learning model averaging approaches based on network parameters and variable selection. These methods effectively avoid the high costs associated with incorrect initial choices of network parameters or variable selection. Numerical simulation results indicate that in most cases, the model averaging method outperforms traditional approaches. Furthermore, empirical studies validate the superiority of this method.