A0213
Title: Measures of uncertainty for model selection
Authors: Yuanyuan Li - Munich Re (United States) [presenting]
Jiming Jiang - University of California-Davis (United States)
Xiaohui Liu - Jiangxi University of Finance and Economics (China)
Abstract: Model selection is a key step in statistical analysis, but its outcome can be sensitive to the data and may not always reflect the true data-generating process. Standard model selection methods often provide a single "best" model, without accounting for the uncertainty in that choice. This can lead to overconfident conclusions and unreliable inference. The aim is to propose two measures to quantify uncertainty in model selection. The first is analogous to a confidence set, identifying a collection of plausible models rather than a single choice. The second focuses on the probability of model selection error. Both methods are applicable to classical settings with a fixed number of candidate models and modern high-dimensional problems. They are conceptually simple, computationally efficient, and supported by both theoretical analysis and empirical results. Their applications are demonstrated by applying them to real-life problems.
