报告题目：Optimal Parameter-Transfer Learning by Time-Varying Model Averaging

报告人：孙玉莹  研究员 中国科学院

报告时间：2026年11月10日10:00-11:00

报告地点：伍卓群楼第二报告厅

校内联系人：程建华  [email protected]

报告摘要：

The presence of structural changes in economics and related fields poses challenges for out-of-sample forecasting, particularly when the sample size of the target data is limited. This paper proposes a novel optimal parameter-transfer learning approach through time-varying model averaging (Transfer-TVMA), which enhances target model predictions by adaptively transferring potentially shared parameter information under structural changes and model uncertainty. A local forward-validation weight choice criterion is developed to select time-varying combination weights for candidate models based on different populations. The asymptotic optimality and convergence properties of the selected time-varying weights are established under regularity conditions. Beyond point forecasts, we further develop a conformal prediction interval algorithm based on Transfer-TVMA and provide its asymptotic validity without the exchangeability assumption. Simulation studies and empirical applications to inflation forecasting and exchange rate forecasting demonstrate the superior predictive performance of the proposed approach compared to existing competing methods.

报告人简介：

孙玉莹，中国科学院数学与系统科学研究院研究员、博士生导师，主持国家优秀青年科学基金项目，入选北京市青年科学基金重点项目（原北京市杰青），入选中国科协“青年人才托举工程”。研究兴趣主要有计量经济学、经济预测理论与方法等。在Management Science, Journal of Econometrics, Journal of Business & Economic Statistics，《中国科学：数学》，《管理科学学报》等国内外重要期刊上发表论文30余篇；基于新理论与新方法，作为主笔，30余篇政策报告获国家领导人批示或被中办、国办采用。先后获成思危优秀科研成果奖、陈景润未来之星、关肇直青年研究奖、中国科学院数学与系统科学研究院“重要科研进展奖（2017，2019）”等。