北航经管学院“融实论坛”系列讲座
（2026年第3期）
讲座题目：Policy Averaging for Stochastic Decision Problems: Theory and an Application to the Newsvendor Problem
讲座时间：2026/5/15，10:00-12:00
讲座地点：新主楼A949
讲座嘉宾：崔翔宇，上海财经大学教授
讲座嘉宾简介：
崔翔宇教授，现为上海财经大学 统计与数据科学学院教授，管理科学与工程学会第四届理事。研究方向包括行为金融、投资组合优化、金融计量、风险管理。主持国家自然科学基金青年项目、面上项目和重点项目。成果发表于《Operations Research》，《INFORMS Journal on Computing》，《IEEE Transactions on Automatic Control》等期刊。
讲座概要：
We propose a Policy Averaging Approach (PAA) for stochastic decision problems that combines multiple candidate policies into a single data-driven decision rule. The main idea is to exploit diversification across policies, in the spirit of model averaging and risk diversification, so as to improve robustness, stability, and decision quality under uncertainty. Rather than relying on a single model specification or a single estimated policy, PAA aggregates information from competing policies and uses cross-validation to select and optimize averaging weights. Our main contribution is methodological and general. We formulate PAA for a broad class of stochastic decision problems and provide theoretical analysis to shed light on its effectiveness. The results suggest that policy averaging can improve the stability of data-driven decisions, reduce sensitivity to model misspecification, and enhance out-of-sample performance relative to individual candidate policies. We use the newsvendor problem as an illustrative application. In that setting, existing approaches often depend on restrictive distributional assumptions, specific feature-based demand models, or highly adaptive data-driven rules that may be unstable or prone to overfitting. PAA provides a unified framework that synthesizes such competing approaches while retaining interpretability and empirical flexibility. Through theoretical analysis, simulation experiments, and an empirical study, we show that PAA achieves lower expected cost, more stable performance, and better justified recommendations than benchmark methods.
