讲座题目：Recurrent Neural Networks for Nonlinear Time Series
内容摘要：Bridging classical time-series econometrics with modern machine-learning tools, we establish theoretical guarantees for recurrent neural networks trained on time series generated by nonlinear vector autoregressive moving-average models with exogenous variables. We derive upper bounds on predictive risk that decompose into approximation and estimation errors. Approximation error depends on smoothness and effective dimension, while estimation error depends on architecture; both vanish as network complexity grows with sample size. Under an invertibility condition, recurrence yields parsimonious representations of temporal dependence and faster convergence than nonparametric regressions based on high-order autoregressive truncations.
主讲人简介：沈周瑜，北京大学光华管理学院商务统计与经济计量系助理教授。 2025年于美国芝加哥大学布斯商学院获统计与计量经济学博士学位，师从修大成教授；2020年毕业于中国科学技术大学，获统计学学士学位。研究方向主要聚焦于统计学与机器学习的交叉领域，特别关注在经济学背景下应用机器学习方法所面临的理论与方法学问题。相关成果发表于Journal of Business & Economic Statistics等国际学术期刊，并在 2024 年国际金融计量学会（Society for Financial Econometrics, SoFiE）年会上荣获 Bates-White 最佳论文奖。
报告时间：2026年3月17日(周二)，16:30-18:00
线下地点：厦门大学经济楼C108
线上地点：腾讯会议 ID：230 321 309
