题目：Model Averaging with Real-Time Frequency Decomposition
主讲人：洪永淼
讲座时间：2026年9月21日（周一） 10:30
地点： 综合楼644会议室
主讲人简介：洪永淼，中国科学院数学与系统科学研究院关肇直首席研究员，中国科学院大学经济与管理学院院长，发展中国家科学院院士，世界计量经济学会会士，亚太人工智能学会会士，亚洲金融经济研究局高级会士，教育部高等学校经济学类专业教学指导委员会副主任委员。曾任美国康奈尔大学经济学与国际研究讲席教授、统计学教授，中国留美经济学会会长。
研究领域为计量经济学、时间序列分析、金融计量学、统计学，在Annals of Statistics、Biometrika、Econometrica、Journal of American Statistical Association、Journalof Political Economy、Journalof Royal Statistical Society B、Management Science、Quarterly Journalof Economics、Review of Economic Studies、Review of Financial Studies、《经济研究》《管理世界》《中国工业经济》《管理科学学报》《中国科学院院刊》等经济学、金融学和统计学中英文主流期刊发表文章190余篇。出版《Python经济大数据分析》《概率论与统计学》《高级计量经济学》、Probability and Statistics for Economists、Foundations of Modern Econometrics: A Unified Approach等中英文著作。2014-2025年连续12年入选Elsevier经济学/统计学中国高被引学者榜单，获2022年高等教育（本科）国家级教学成果奖一等奖。
讲座摘要：The existing forecasting literature typically assumes a uniform relationship between predictors and the target variable across frequencies, thereby overlooking the heterogeneous predictive information contained in different frequency components. We propose a model averaging approach with real-time frequency decomposition that incorporates such frequency heterogeneity into both candidate model construction and weight selection. At each forecast origin, the proposed method decomposes predictors into frequency-specific components using only information available at that time, constructs component-specific candidate models in which each component separately forecasts the target, and selects combination weights by minimizing a forward-validation criterion. This real-time construction differs from conventional model averaging approaches in that each newly available observation leads to an updated decomposition, which may revise previously obtained component values and thereby make the historical entries of the design matrices forecast-origin-specific. We establish the asymptotic optimality of the selected combination weights in the sense of achieving the lowest possible out-of-sample prediction risk. The convergence rate of the selected weights toward the optimal weights is also derived. Monte Carlo simulations and an empirical application demonstrate the superior performance of the proposed method.
