AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-20
Optimal Feature-Based Newsvendor Policies under Censored Demand
The paper considers a newsvendor problem in which a decision-maker must learn how much stock to order from censored demand data, where demand is censored when insufficient stock was available to meet it. The authors propose a model-agnostic loss function approach that estimates optimal order quantities from observational data, characterizing the optimal policy as the minimizer of this loss. The method requires only a single estimation step, can be implemented in linear empirical risk minimization, kernel weights optimization, and deep neural networks, converges quickly to the optimal policy on synthetic data, and improves performance on real-world retail data.