【主讲人简介】：张新雨，中国科学院数学与系统科学研究院研究员。长期从事统计和计量经济学理论与应用方面的研究工作，与合作者解决了模型平均研究中的多个难题，并将模型平均与迁移学习、随机森林、投资组合、库存管理等融合提出了新方法，同时将所提出的预测方法应用于实际问题为相关部门的决策提供了参考依据。先后主持国家自然科学基金委青年基金A类及其延续项目等，曾获中国青年科技奖。

【内容简介】：Data-driven decision-making has become ubiquitous, transforming data into actionable decisions in domains such as management, operations, finance, marketing, healthcare, and policy learning. Model averaging, by adopting an ensemble strategy of "not putting all eggs in one basket'', provides a more robust perspective for decision-making. However, the traditional model averaging in decision-making faces three key challenges: prediction -- decision mismatch,  averaging prediction -- optimization mismatch, and fixed weight -- complex world mismatch. To address these issues, we propose a unified decision averaging framework for contextual data-driven decision-making to mitigate decision uncertainty. The framework integrates multiple actionable decisions and targets optimal downstream decision performance. Specifically, candidate decisions are flexibly constructed from different predictive models, optimization methods, expert heuristics, LLM agents, and other decision-making mechanisms. We develop a context-adaptive and decision-focused criterion to determine the weight function across the candidate decisions. Theoretically, we establish the asymptotic optimality of the proposed framework in the sense of attaining the lowest possible out-of-sample expected decision loss, and further derive weight convergence and regret bounds. Numerically, our proposed method outperforms  candidate decisions and alternative averaging methods in both simulations and empirical applications to newsvendor, portfolio optimization and LLM routing problems.

【讲座时间】：2026年10月16日（星期五）下午16:00

【讲座地点】：人文社科科研楼1801会议室