讲座题目：Improving time series estimation and prediction efficiency via transfer learning techniques
主讲人：李国栋 教授
主持人：周勇 教授
开始时间：2025-07-14 11:30
讲座地址：普陀校区理科大楼A512
主办单位：统计学院
报告人简介：
李国栋，本科和硕士毕业于北大数学学院，2007年于香港大学统计精算系获得统计学博士，随后在南洋理工大学任助理教授。现任香港大学统计精算系教授，数据科学研究院副院长。主要研究方向包括时间序列分析，分位数回归，高维统计数据分析和机器学习。目前发表学术论文70余篇，其中20余篇发表在统计学4大顶级期刊、计量经济顶级期刊以及机器学习的顶级会议上。
报告内容：
A representation-based transfer learning framework is proposed for time series analysis within the vector autoregression paradigm. We develop a two-stage regularized estimation methodology and establish its nonasymptotic theoretical properties. The proposed approach enhances estimation accuracy and forecasting performance in the target domain by effectively leveraging information from related source datasets through representation learning. A distinctive feature of our method is its capability to handle datasets with varying sequence lengths and asynchronous starting/ending points, thereby offering remarkable flexibility in integrating information from diverse datasets. Extensive simulation studies and an empirical analysis conducted on a comprehensive collection of macroeconomic time series across multiple countries demonstrate the superior performance of our transfer learning framework compared to conventional single-task learning methods, particularly in scenarios involving high-dimensional target datasets with limited sample sizes.
