摘要：Regression extremiles provide a least squares analog of regression quantiles, offering a coherent risk measure based on weighted expectations rather than tail probabilities. In this study, we propose jackknife model averaging for extremile regressions (JMAER) to address model uncertainty in a data-rich environment. We derive the asymptotic properties of the extremile estimator under misspecification, allowing the parameter dimension to grow with the sample size. We prove that JMAER with weights obtained from an integrated leave-one-out cross-validation criterion achieves asymptotic optimality by minimizing out-of-sample prediction error, and we establish the convergence rate of the weights. For applications involving many candidate models, we introduce a preliminary screening procedure and establish its sure screening property, ensuring that all important variables are retained with probability approaching one; the screening-based JMAER is further shown to be asymptotically optimal. Monte Carlo simulations demonstrate its superior finite-sample performance over alternative model averaging and selection methods. An empirical application to predicting conditional extremiles of excess stock returns further illustrates the utility of the proposed method in financial risk assessment.

个人简介：李海奇，湖南大学金融与统计学院教授、博士生导师，副院长。目前研究方向为金融计量经济学、金融工程、数据科学等，研究成果发表于经济学国际顶尖和权威期刊以及中文重点期刊，如Journal of Econometrics, Econometric Theory, Journal of Business & Economic Statistics,Econometric Reviews, Journal of Futures Markets,《数量经济技术经济研究》《统计研究》《中国管理科学》《计量经济学报》等。曾主持国家自然科学基金项目、教育部人文社科规划基金项目、湖南省自然科学基金杰出青年项目等多项国家和省部级科研项目。