组织者：吴宇楠

Abstract：Model averaging is widely used to mitigate model uncertainty, yet two fundamental questions remain insufficiently resolved: (i) how to design a principled model screening step that improves averaging with rigorous guarantees, and (ii) when averaging is provably preferable to model selection for minimizing prediction risk. This talk revisits model averaging through these two lenses.

First, we develop a risk-minimization-oriented screening framework that goes beyond the common practice of merely trimming the candidate list. Building on the notion of full asymptotic optimality, we propose screening rules that, among all subsets of a given size, choose the subset that minimizes the corresponding model-averaging risk. The resulting post-screening averaging estimator is shown to theoretically dominate competing screening approaches of the same cardinality and, under mild conditions, retains full asymptotic optimality relative to averaging over the full model set. For implementation, we provide a consistent plug-in procedure and an efficient dynamic-programming algorithm.

Second, extending the framework of Peng and Yang (2022), we study the broader selection-versus-averaging dilemma by identifying the key determinants under which model averaging is significantly better than model selection in terms of achieving lower prediction risk. Our results work under more flexible conditions on the data generating process, which substantially broadens the scope of existing studies. A simulation study is conducted to support the theoretical findings.