报告题目：Robust Matrix Factor Models with Qn-based Autocovariance for Outlier Detection

报告人：郝程程，上海对外经贸大学

报告时间：2026年10月20日(星期二) 10:15-11:00

报告地点：伍卓群楼第一报告厅

校内联系方式：邹婷婷[email protected]

报告摘要：

Matrix factor models provide two-way dimension reduction for matrix-valued time series, but loading-space estimators based on sample lagged autocovariances can be highly sensitive to extreme observations. We propose two robust spectral estimators, rTOPUP and rTIPUP, by replacing the entrywise lagged autocovariances underlying TOPUP and TIPUP with scale-difference estimators constructed form the robust Qn statistic. Under Gaussian short-range dependence, a uniform root-T condition for the relevant Qn scales, and mode-specific identification conditions, we establish root-T convergence of the robust lagged-autocovariance operators and the resulting loading-space estimators. We also develop a residual-based procedure that first detects outlying time points through robustly calibrated Frobenius norms and then localizes outlying cells using Qn-standardized residuals. Simulations under cellwise additive contamination show that rTIPUP substantially improves loading-space recovery relative to TIPUP while incurring only a modest loss under clean sampling; rTOPUP is less affected by contamination but remains inaccurate in the simulation design. An application to China’s exports across 97 commodity categories and 30 trading partners over 2014-2024 identifies 2024 as the only globally anomalous year and localizes 146 commodity-partner cells contributing to that departure.

报告人简介：

郝程程，上海对外经贸大学副教授，瑞典斯德哥尔摩大学统计学博士。主要研究方向为张量数据分析、异常点诊断，在《Scandinavian Journal of Statistics》等期刊上发表一作论文10余篇；主持完成国自科、上海市科委、瑞典皇家科学院等课题，指导学生获全国研究生统计建模大赛一等奖、“三下乡”全国优秀调研报告、“挑战杯”上海市一等奖等，获上海高校青年东方学者、上海市白玉兰人才、上海市教学成果特等奖。