Abstract:
We study the data-driven newsvendor problem in situations where historical demand data and associated feature information are available. To address this problem, we leverage information from similar data sources and employ transfer learning to improve decision-making and facilitate statistical inference on the parameter of primary interest in the target domain. Within the framework of a semiparametric regression model, we assume the existence of a shared feature representation for confounding effects across different tasks. By learning this shared representation from various source domains, we can effectively transfer it to the target domain. This approach enables both accurate decision-making and interpretability in the target domain. We establish sufficient conditions for model identifiability, derive a finite sample performance bound for the cost function, and prove that the estimator of the parameter of primary interest in the target model is consistent and asymptotically normal. Through simulation studies, we demonstrate the superiority of our method, and we further illustrate its practical applications in inventory decision-making forbike-sharing systems.
Speaker Profile:
Zhang Xinyu, Research Fellow at the Academy of Mathematics and Systems Science, Chinese Academy of Sciences. His research focuses on econometrics and statistical theory and applications, with particular interests in model averaging methods and their interdisciplinary applications in economic forecasting, managerial statistics, machine learning, and biomedical studies. He currently serves as Area Editor for the SCI-indexed journal Journal of Statistical Computation and Simulation (JSSC) and is on the editorial boards of several journals including Systems Science and Mathematics. Dr. Zhang is a recipient of the China Youth Science and Technology Award and has led multiple research projects funded by the National Natural Science Foundation of China at Class C, B, A, and A-extension levels.
